Multimodal deep learning for chronic obstructive pulmonary disease: a conceptual synthesis of architectures, fusion strategies, and equity-centered translational frameworks

Introduction

Background

Chronic obstructive pulmonary disease (COPD) is a chronic, heterogeneous, and progressive respiratory disease defined by persistent symptoms and airflow limitation due to airway and alveolar abnormalities stemming from exposure to noxious particles or gases (1). As a major non-communicable disease crisis of the 21st century, COPD affects over 400 million people globally, with projections estimating nearly 600 million cases by 2050 reflecting a 23% rise from 2020 and driven by population ageing, persistent use of tobacco in low- and middle-income countries (LMICs) and escalating ambient air pollution (2,3). It ranks as the third leading cause of global mortality, claiming over 3.5 million lives annually (4).

Projections to 2050 estimate that male COPD patients worldwide will incur cumulative direct medical costs of US$15.62 trillion and experience over 10 billion exacerbations (5). These figures substantially underestimate the true global burden, owing to limited data from sub-Saharan Africa, large parts of Latin America, and Central Asia, being regions characterized by restricted spirometry access and high underdiagnosis rates (6). COPD imposes a disproportionate burden on LMICs, which account for approximately 90% of premature COPD-related deaths (7).

Despite its substantial global burden, COPD is still profoundly underdiagnosed and undertreated. Population-based studies show that 70-80% of individuals fulfilling spirometry criteria for airflow obstruction lack a formal diagnosis, with a typical 3–6-year delay from symptom onset to clinical recognition (8,9). This underdiagnosis stems from multifactorial, interdependent causes, including limited awareness among patients and clinicians, misattribution of insidious symptoms to ageing, restricted spirometry access, particularly in primary care LMICs and the test’s inherent technique-dependence and effort requirements (10,11).

Post-bronchodilator spirometry, the diagnostic gold standard per Global Initiative for Chronic Obstructive Lung Disease (GOLD) criteria [forced expiratory volume in one second to forced vital capacity ratio (FEV₁/FVC)] below 0.70, demands trained personnel, calibrated equipment, and patient cooperation, which are often unavailable beyond specialized centers (11,12). Moreover, GOLD severity staging (stages 1–4) addresses only airflow limitation, overlooking COPD’s heterogeneity, including emphysema, inflammatory profiles, comorbidities, and functional reserve, all of which independently influence prognosis and management outcomes (13,14).

Multimodal deep learning (DL) overcomes the constraints of unimodal approaches by integrating representations from diverse data types including imaging, signals, tabular, and text into a cohesive framework (15). In COPD, pertinent modalities encompass thoracic computed tomography (CT) (16), chest X-ray (CXR), flow-volume curves from spirometry and oscillometers (17), structured electronic health record (EHR) variables (e.g., medications, blood analyses, comorbidities) (18), circulating biomarkers such as blood eosinophil counts (19), wearable sensor-derived physiological signals, digital respiratory-sound recordings, and genomic/ transcriptomic data (20-22). These modalities offer complementary perspectives, with their integration demonstrably enhancing diagnostic and prognostic accuracy beyond single-modality benchmarks (23).

COPD manifests as a heterogeneous syndrome arising from genetic susceptibility and cumulative environmental exposures (24), predominantly featuring small airway disease and emphysema, compounded by systemic inflammation, skeletal muscle atrophy, and psychiatric comorbidities (25).

The GOLD classification stratifies airflow obstruction severity using four spirometry grades based on FEV₁ as a percentage of predicted: GOLD 1 (≥80%), GOLD 2 (50–79%), GOLD 3 (30–49%), and GOLD 4 (<30%) (26). This is integrated with symptom burden, assessed via the COPD Assessment Test (CAT) or Modified Medical Research Council (mMRC) dyspnea scale, and exacerbation history to define evidence-based management groups (ABE) (26).

However, the FEV₁ paradox reveals a weak correlation between FEV₁ and key patient-centered outcomes, including patient-reported measures, exercise capacity, and mortality risk (27), underscoring the need for advanced phenotyping via CT, biomarker profiling, and composite indices such as the Body mass index, airway Obstruction, Dyspnea, Exercise capacity (BODE) score (28).

Acute exacerbations of COPD (AECOPD) drive disease progression and mortality, with 30-day post-hospitalization mortality rates of 7–10% and one-year mortality exceeding 25% after severe events (29). These are often preceded by detectable physiological changes, creating a therapeutic window for multimodal continuous monitoring (30).

Artificial intelligence (AI), especially DL, offers substantial promise in pulmonary medicine. DL models have demonstrated radiologist-level performance using convolutional neural networks (CNNs) for CT-based emphysema quantification (27), recurrent neural networks (RNNs) have enabled robust prediction of COPD exacerbation risks (31), electrocardiograms (ECGs) for COPD detection (32), novel approaches for forecasting COPD risk from raw spirogram time-series up to five years pre-diagnosis (33), and transformer-based natural language processing (NLP) models have accurately extracted GOLD staging from unstructured clinical notes (34). However, most extant DL systems for COPD are unimodal, relying on single data modalities and neglecting complementary information from diverse clinical sources (35).

Rationale and knowledge gaps

A considerable number of systematic reviews and meta-analyses have documented single-modality AI applications in COPD (36,37), while a 2025 Frontiers review synthesized the wider AI landscape in COPD management (21). Nonetheless, no comprehensive, architecturally grounded conceptual synthesis has delineated the full spectrum of multimodal DL for COPD across the entire clinical continuum. This conceptual synthesis fills that gap.

Over the past decade, AI, particularly DL, has driven significant advancements in medical diagnostics, including respiratory medicine. However, the present COPD AI literature remains predominantly unimodal, with each modality exhibiting inherent limitations that preclude comprehensive disease phenotyping (38): CT fails to capture dynamic physiological function or inflammatory markers and is scarce in LMICs; spirometry overlooks structural, molecular, and extrapulmonary features (12,39); wearable sensors lack sufficient contextual specificity to distinguish AECOPD precursors from non-COPD variability (37); and EHR-derived models are constrained by diagnostic coding inaccuracies and temporal gaps (40).

A 2026 systematic review and meta-analysis showed limited pooled diagnostic accuracy (41%) for single-modality DL models in COPD detection (41). Concurrently, a scoping review of 432 multimodal medical DL studies reported that multimodal models surpassed unimodal equivalents by 6.2 percentage points in area under the curve (AUC), though both are still restricted to research environments (38). These findings underscore multimodal DL’s promise for COPD yet emphasize the need for targeted clinical validation.

Despite notable advancement in DL applications for COPD, six operationally distinct yet interrelated knowledge gaps call for this synthesis. First, no COPD-specific multimodal architectural taxonomy exists to guide data modality encoding and fusion, despite multimodal integration being a recognized priority (20,21). Second, existing reviews address COPD clinical tasks in isolation, not mapping the full care continuum to multimodal approaches (41). Third, principled strategies for handling modality missingness are lacking (42). Fourth, federated learning, technology promising for mitigating cross-institutional data fragmentation and privacy issues, has not been systematically reviewed for COPD (43). Fifth, models are predominantly trained on high-income country cohorts, compromising generalizability to LMIC populations due to algorithmic bias (3,5,6). Sixth, explainability and regulatory translation are underexplored, despite robust research performance, as shown by a scoping review of 145 multimodal AI studies in medicine (44) and a review of AI in respiratory care (45). Collectively, these gaps underscore the need for a comprehensive, architecturally precise conceptual synthesis of multimodal DL for COPD, encompassing the care continuum, fusion strategies, real-world data challenges, equity, regulatory integration, and a translational roadmap.

Goal and review objectives

The primary goal of this conceptual synthesis is to critically evaluate, integrate, and architecturally map the current landscape of multimodal DL applied to COPD and to propose the Multimodal Deep Learning Conceptual Framework (MDLCF) as a unified, clinically translatable synthesis output. Four review objectives form the structure of the study:

• To identify, classify, and comparatively evaluate multimodal DL architectures and fusion strategies applied to COPD across all clinical domains.
• To synthesize the quantitative performance evidence for multimodal versus unimodal AI models, with specific attention to the incremental value attributable to modality integration.
• To critically appraise the principal barriers to clinical deployment of multimodal COPD AI systems, encompassing algorithmic bias, modality missingness, prospective validation deficits, and regulatory gaps.
• To propose an evidence-grounded, equity-centered research agenda and a validated conceptual framework defining the methodological investments required for clinically deployable multimodal COPD AI.

The study adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines (46,47), adapted to the purposive and iterative logic of conceptual synthesis. Additionally, the synthesis was conducted and reported following the synthesis without meta-analysis (SWiM) reporting guideline for narrative synthesis components (48), and the Enhancing Transparency in Reporting the Synthesis of Qualitative Research (ENTREQ) 21-item checklist for qualitative synthesis transparency (49). We present this article in accordance with the PRISMA, SWiM and ENTREQ reporting checklists (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0078/rc).

Methods

Study design and reporting standards

This study is designed as a conceptual synthesis, an evidence-informed, theory-building methodology that systematically integrates, interprets, and architecturally maps a heterogeneous body of empirical and theoretical literature to produce generative outputs that transcend the findings of any individual included study (50). Conceptual synthesis is distinct from a traditional systematic review in that its primary goal is not the statistical pooling of effect sizes but the construction and validation of an integrated framework capable of guiding future research design, architecture development, and clinical translation.

Quality appraisal was conducted using Quality Assessment of Diagnostic Accuracy Studies, second edition (QUADAS-2) (51), Prediction model Risk Of Bias ASsessment Tool (PROBAST) (52), Assessing the Methodological Quality of Systematic Reviews 2 (AMSTAR-2) (53), and Grading of Recommendations, Assessment, Development and Evaluations (GRADE) (54).

Information sources and search strategy

The search for literature adhered to the epistemological principles of conceptual synthesis, employing a systematic, purposive, and iterative strategy rather than exhaustive retrieval (47,55). This method distinguishes conceptual synthesis, which prioritizes conceptually pertinent studies and pursues key ideas to theoretical saturation, from conventional systematic reviews that seek comprehensive inclusion of eligible studies. Iterations persisted until no novel conceptual themes, architectural innovations, or translational patterns surfaced in emergent literature.

The literature search strategy comprised three sequential activities conducted across January 2020 to March 2026. First, systematic keyword and subject-term searches employed Boolean strings integrating COPD-specific terminology, AI/DL methods, multimodality/fusion concepts, and clinical task descriptors, spanning major biomedical, engineering/computational, and multidisciplinary databases.

Second, citation tracking involved manual backward screening of reference lists from conceptually relevant studies, alongside forward tracking in Scopus and Web of Science for 15 seminal papers selected by citation frequency, methodological innovation, and conceptual impact.

Third, purposive sampling prioritized conceptual novelty and distinctiveness alongside methodological quality studies replicating findings previously captured in the CSED workbook were documented in the tracking log but were not subjected to full data extraction, ensuring that the synthesis reflected conceptually distinct evidence. Supplementary searches of preprint servers (arXiv, medRxiv) and grey literature (Google Scholar) captured emerging contributions.

The study selection process is summarized in the PRISMA-adapted flow diagram (Figure 1), reported in accordance with the PRISMA 2020 guidelines (46,47) and adapted to the purposive, iterative logic of conceptual synthesis. During the identification stage, 7,205 records were identified: 6,243 through systematic database searching (PubMed/MEDLINE, Embase, Web of Science, IEEE Xplore, Scopus, Cochrane Library, and ACM Digital Library) and a further 962 through backward and forward citation tracking, Google Scholar and preprint servers (arXiv and medRxiv). After 1,491 duplicate records were removed, 5,714 records were screened by title and abstract, of which 4,391 were excluded as not conceptually relevant. At the eligibility stage, 1,323 full-text articles were assessed against the purposive relevance criteria operationalized through the CSED methodology, and 1,105 were excluded, with the principal reasons being single-modality focus without fusion (n=412), absence of quantitative performance metrics (n=298), non-COPD primary focus (n=187), pre-2020 publication without novel conceptual advancement (n=119), and language or duplication concerns (n=89). Ultimately, 218 studies met the eligibility criteria and were included in the qualitative conceptual synthesis, of which 89 provided extractable 2×2 diagnostic accuracy data and were included in the bivariate meta-analysis. The complete list of excluded full-text studies with individual reasons is provided in Appendix 1.

Figure 1 PRISMA 2020-adapted flow diagram summarizing the identification, screening, eligibility, and inclusion of studies in conceptual synthesis and bivariate meta-analysis. COPD, chronic obstructive pulmonary disease; CSED, Conceptual Synthesis Excel Dump.

Conceptual Synthesis Excel Dump (CSED)

This methodological synthesis at its core innovation, employs the CSED, a structured spreadsheet tool for systematically collecting, comparing, and integrating concepts from diverse, heterogeneous sources into a unified conceptual framework (56). The CSED served as the principal mechanism for transforming iterative, purposive literature search results into a cohesive, synthesizable evidence base.

CSED was implemented as a multi-sheet Excel workbook (Appendix 2). The primary sheet systematically recorded, for each included source, bibliographic details, data modalities, AI architecture, fusion strategy (early, intermediate, late, or hybrid), clinical task, key performance metrics, translational barriers, geographic/demographic characteristics, and primary conceptual contributions.

The second sheet facilitated cross-source thematic comparisons, enabling horizontal synthesis of convergent/divergent perspectives (e.g., modality missingness handling, calibration reporting) and consensus mapping. The third sheet included an iterative conceptual map (Figure 2), dynamically populated with clustered themes and inter-relationships as evidence was reviewed. The fourth sheet tracked concept saturation via an iteration-contribution matrix (Table 1), declaring saturation after three consecutive search rounds yielding no novel architectural concepts, fusion strategies, or translational barriers. The CSED was iteratively maintained by two independent researchers, with classification discrepancies resolved by consensus.

Figure 2 Conceptual map. Developed iteratively as living conceptual synthesis through the CSED across three successive search rounds. T1-T31 refers to the 31 themes across 7 clusters mapped to the six-stage MDLCF. AECOPD, acute exacerbation of chronic obstructive pulmonary disease; AI, artificial intelligence; CNN, convolutional neural network; COPD, chronic obstructive pulmonary disease; CSED, Conceptual Synthesis Excel Dump; CT, computed tomography; CVI, content validity index; CXR, chest x-ray; DNA, deoxyribonucleic acid; EHR, electronic health record; GOLD, global initiative for chronic obstructive lung disease; GNN, graph neural network; LMIC, low- and middle-income country; LSTM, long short-term memory; MDLCF, multimodal deep learning conceptual framework; NLP, natural language processing; PFT, pulmonary function test; RNN, recurrent neural network; VAE, variational autoencoder; XAI, explainable artificial intelligence.

Eligibility criteria

Eligibility criteria were defined using the Population, Intervention, Comparator, Outcome (PICO) framework (57), a systematic and reproducible approach for study selection that is widely adopted in clinical AI evidence syntheses (52,58).

• Population: adult patients (≥40 years) diagnosed with COPD, clinically at risk of COPD, or enrolled in population-based cohorts where COPD was a primary or major condition of interest. Studies restricted exclusively to populations (<40 years) or to non-respiratory primary conditions were excluded. Studies including mixed patient populations were included, where COPD-specific data were separately reported or extractable.
• Intervention: multimodal DL systems architectures integrating two or more distinct data modalities from the following sources: thoracic CT, chest X-ray (CXR), spirometry and other pulmonary function tests (PFT), wearable physiological sensors, EHR including structured data and unstructured clinical notes, respiratory sound recordings, blood or sputum biomarkers, and genomic or transcriptomic data.
• Comparator: single-modality AI or DL models; traditional clinical scoring systems (BODE index, GOLD spirometry staging, CAT score, mMRC dyspnea scale); and standard clinical diagnosis or management without AI assistance.
• Outcome: quantitative performance metrics for one or more of the following COPD clinical tasks: (I) COPD detection and diagnosis; (II) GOLD-stage severity classification; (III) AECOPD prediction; and (IV) personalized disease management. Eligible performance metrics included AUC/AUROC, accuracy, sensitivity, specificity, F1-score, C-statistic, mean absolute error, intraclass correlation coefficient, or equivalent quantitative measures reported in sufficient detail for extraction and synthesis.
• Inclusion criteria: studies were eligible for inclusion if they satisfied the following criteria: (I) original research, systematic reviews, meta-analyses, or algorithm development and benchmarking studies employing multimodal DL for COPD-related clinical tasks; (II) COPD as the primary or major condition of interest, with COPD-specific data reported separately; (III) reporting of at least one quantitative performance metric; (IV) publication or public availability between January 1, 2020, and March 31, 2026; and (V) full-text availability in English or via verified English translation.
• Exclusion criteria: studies were excluded if they met any of the following criteria: (I) use of single-modality AI without fusion from a second independent modality; (II) exclusive focus on non-COPD respiratory conditions without extractable COPD-specific data; (III) lack of quantitative performance evaluation; (IV) case reports, editorials, opinion pieces, or conference abstracts lacking accessible peer-reviewed full-text articles; (V) animal, in vitro, or exclusively pediatric COPD studies; or (VI) duplicate publications reusing the same dataset and model without novel methodological or clinical contributions.

Quality assessment and risk of bias

Due to methodological heterogeneity among included studies, a tailored multi-instrument quality appraisal was implemented, with tool selection aligned to study design. QUADAS-2 assessed primary diagnostic accuracy studies for risk of bias and applicability concerns across patient selection, index test, reference standard, and flow or timing domains (rated low, high, or unclear). PROBAST evaluated prognostic or predictive models (e.g., AECOPD prediction), targeting optimism from small samples, internal validation only, failure to correct for overfitting, and absent reporting calibration. AMSTAR-2 rated systematic reviews or meta-analyses (high to critically low confidence) on 16 items, including search scope, bias assessment, and heterogeneity management. GRADE determined evidence certainty (high to very low) at the clinical domain level. Two independent reviewers applied all tools, with inter-rater reliability-quantified via Cohen’s κ for categorical ratings (mean κ=0.82 across the 15% subsample assessed) (59); discrepancies were resolved by consensus. Overall quality is summarized in Table 2.

Evidence synthesis strategy within a purposive sample

Three complementary synthesis methods were employed across evidence subsets: thematic synthesis (62), SWiM (48) and bivariate meta-analysis (Reitsma model) (63). Thematic synthesis involved line-by-line inductive coding in MAXQDA 2024 by two independent reviewers (Cohen’s κ on 15% subsample) (64), yielding descriptive and analytical themes that informed an architectural taxonomy, fusion strategy framework, and research agenda.

SWiM addressed non-pooled data via direction-of-effect analysis, range-of-effects reporting, and harvest plots. Bivariate meta-analysis in R 4.3.2 (mada, metafor packages) was applied (Appendix 3 for the complete R analysis code) to homogeneous diagnostic accuracy subsets, estimating pooled sensitivity, specificity, likelihood ratios, and diagnostic odds ratios (DOR). The pooled estimates are descriptive of the purposively sampled evidence base rather than inferential for the full multimodal DL COPD literature, consistent with the epistemological framing of conceptual synthesis. Summary receiver operating characteristic (SROC) curves were generated for two subsets: (I) CT alone vs. CT + spirometry for GOLD-stage classification (n=12 studies) and (II) unimodal vs. multimodal AI for COPD detection (n=10 studies). Each SROC derived from the bivariate normal fit of logit-transformed sensitivity/specificity, with summary points (rotated diamonds), 95% confidence ellipses, and prediction regions (Figure 3).

Figure 3 SROC curves from bivariate meta-analysis (Reitsma mixed-effects model; R 4.3.2, mada package). Comparison 1 (left): CT imaging alone vs. CT + spirometry fusion for GOLD severity staging (n=12 studies). Comparison 2 (right): best unimodal AI vs. multimodal AI for COPD detection (n=10 studies). Combined SROC overlay (below): Circles = individual study operating points. Diamonds = summary operating points (pooled sensitivity and 1-specificity). Shaded ellipses = 95% confidence regions. Dotted ellipses = 95% prediction regions. Heterogeneity: I²=38.4% (Comparison 1); I²=44.7% (Comparison 2). Deeks’ funnel plot asymmetry test: P>0.10 for both subsets (no significant publication bias). AI, artificial intelligence; AUC, area under the curve; COPD, chronic obstructive pulmonary disease; CT, computed tomography; GOLD, Global Initiative for Chronic Obstructive Lung Disease; SROC, summary receiver operating characteristic.

Heterogeneity in the meta-analytic subsets was low-to-moderate, as shown by the I² statistic (38.4% for Comparison 1; 44.7% for Comparison 2) and Cochran’s Q test, falling below the pre-specified threshold of >75% for substantial heterogeneity (65,66). This reflects clinically meaningful but not excessive variation in study populations, imaging protocols, and threshold selection. Publication bias assessment via Deeks’ funnel plot asymmetry test (67) revealed no significant asymmetry (P>0.10) in either subset, providing no evidence of small-study effects or selective reporting bias.

MDLCF

The MDLCF, illustrated in Figure 4, was iteratively developed via thematic synthesis and the CSED process across three systematic, purposive literature search rounds (R1–R3). It rests on four foundational pillars identified from the literature: multimodal synergy, clinical task specificity, translational realism, and equity-centered design. Framework validation was conducted through a two-round modified Delphi consensus process with 20 experts from respiratory medicine, biomedical AI, clinical informatics, health equity, and AI regulation [n=20; item-level content validity index (I-CVI) ≥0.78 for all components; scale-level CVI average (S-CVI/Ave) =0.94; concept saturation achieved at R3·I3] (68,69).

Figure 4 MDLCF: a validated three-tier, six-stage conceptual architecture. The bidirectional arrow at the bottom indicates the iterative feedback loop from real-world deployment (Stage 6) back to Stage 1. ACIL, adaptive clinical intelligence loop; AI, artificial intelligence; API, application programming interface; CNN, convolutional neural network; COPD, chronic obstructive pulmonary disease; CSED, Conceptual Synthesis Excel Dump; CT, computed tomography; CXR, chest X-ray; DL, deep learning; EHR, electronic health record; FEV, forced expiratory volume; FHIR, fast healthcare interoperability resources; FL, federated learning; GAT, graph attention network; GOLD, Global Initiative for Chronic Obstructive Lung Disease; I-CVI, item-level content validity index; LLM, large language model; LMIC, low- and middle-income country; LSTM, long short-term memory; NLP, natural language processing; NRI, net reclassification improvement; RCT, randomized controlled trial; RESNET, residual neural network; S-CVI/AVE, scale-level content validity index (average); TCN, temporal convolutional network; VAE, variational autoencoder.

The MDLCF is explicitly conceived as a three-tier meta-methodological framework: Level 1 (Stages 1–2) constitutes the reproducible methodological pipeline for conducting the conceptual synthesis itself, Level 2 (Stages 3–5) constitutes the technical clinical DL architecture and Level 3 (Stage 6) constitutes the translational and equity-centered deployment. Critically, the framework adopts a bidirectional design, whereby empirical evidence from deployment (stage 6) iteratively informs and refines all antecedent stages, thereby fostering a cyclical paradigm of continuous improvement.

Stage 1: Systematic, purposive, and iterative literature search

The MDLCF framework’s first stage builds its evidentiary base via a purposive, iterative qualitative search strategy optimized for theory construction and concept synthesis, diverging from exhaustive systematic reviews that emphasize comprehensive study retrieval. Sources are purposively sampled from major biomedical and computational databases using targeted Boolean keyword/subject-term queries, augmented by backward citation tracking of included references and forward citation tracking of seminal works in Scopus and Web of Science. Iterations are guided by emergent conceptual themes, with subsequent searches addressing gaps; concept saturation, defined as three successive iterations yielding no novel architectural contributions, terminates the process. This approach ensures conceptual depth over volume, fostering rigorous framework development

Stage 2: CSED

The CSED operationalizes purposively retrieved literature into a structured, synthesis-ready evidence base via a multi-sheet spreadsheet workbook. It encompasses four domains: (I) primary evidence capture, documenting bibliographic details, modality combinations, architecture types, fusion strategies, performance metrics, and key conceptual contributions per source; (II) cross-source thematic comparison, facilitating horizontal juxtaposition of studies to map convergent and divergent perspectives on shared challenges; (III) a dynamic conceptual map, iteratively refined through theme identification, merging, or splitting; and (IV) a concept saturation tracker, logging emergence of novel ideas post each search iteration. Dual independent researchers supported the CSED, resolving discrepancies by consensus, thereby establishing an audit trail for traceability of all MDLCF components to source literature.

Stage 3: Data input and modality-specific encoding

The stage defines a technical data input layer that integrates complementary modalities for comprehensive COPD phenotyping and management: thoracic CT and CXR imaging, pulmonary function testing (including spirometry), wearable bio signals, EHR data with NLP-derived text, and biomarker/genomic profiles. These modalities offer distinct insights into structural lung pathology, functional deficits, real-world physiological dynamics, clinical context, and molecular endotypes, respectively.

Each modality undergoes processing via specialized DL encoders optimized for its structural and temporal attributes: a 3D Swin Transformer for volumetric CT scans; a Tabular Transformer for spirometry and biomarker data; a multi-scale Temporal Convolutional Network (TCN) for wearable time-series; a hybrid ClinicalBERT-Graph Attention Network (GAT) for clinical text and structured relational data; and an LSTM-based encoder for longitudinal patient trajectories. This modular architecture preserves modality-specific features while generating embeddings amenable to downstream fusion.

To enhance real-world clinical viability amid frequent data incompleteness, the framework employs stochastic modality masking during training and variational autoencoder (VAE)-based imputation at inference, thereby ensuring robust performance across heterogeneous datasets with partial modality availability.

Stage 4: Cross-modal fusion and task-specific output

The stage integrates a cross-modal fusion layer with five parallel task-specific output heads for COPD analysis. Fusion occurs in two phases: (I) pairwise cross-attention between complementary modalities of imaging with functional-biomarker data (capturing structural-functional correlations for GOLD staging) and wearables with clinical data (reflecting real-world physiological trajectories in clinical context); and (II) a multi-head attention transformer that fuses all five encoded representations, dynamically weighting modality importance per patient or task while propagating epistemic uncertainty via Monte Carlo Dropout for calibrated confidence.

The five jointly trained heads derive predictions from the fused representation: (I) COPD detection probability with uncertainty quantification; (II) multi-class GOLD severity (grades 1–4) via ordinal regression; (III) time-varying exacerbation risk (7-, 30-, 90-day horizons) using cause-specific survival modelling for competing risks; (IV) FEV₁ decline trajectory modelled by neural ordinary differential equations (Neural ODEs) for continuous-time lung function dynamics; and (V) phenotype cluster assignment with personalized management recommendations. Joint training improves the shared fusion layer for concurrent detection, staging, prognosis, and decision support, enhancing multimodal information extraction.

Stage 5: Adaptive clinical intelligence and explainability

The stage integrates algorithmic outputs with clinical decision-making via two parallel components: the Adaptive Clinical Intelligence Loop (ACIL) and the Explainability and Clinical Interface module. ACIL employs Bayesian filtering to temporally smooth risk estimates, mitigating transient physiological noise; applies predefined thresholds to generate actionable alerts for high-risk patients; integrates a reinforcement learning (RL) agent to iteratively refine management recommendations based on patient outcomes; and enforces a human-in-the-loop mechanism requiring clinician approval for all AI-derived actions.

Concurrently, the Explainability module produces interpretable visualizations, including Gradient-weighted Class Activation Mapping (Grad-CAM) saliency maps for CT predictions (70), SHapley Additive exPlanations (SHAP) attributions for tabular data contributions (71), counterfactual explanations delineating risk category shifts, and calibration plots assessing predictive probability alignment with observed events. Outputs are rendered in a clinician dashboard seamlessly integrated into EHR systems via fast healthcare interoperability resources (FHIR)-compliant APIs.

Stage 6: Translational and equity-centered deployment

The sixth and final stage of MDLCF emphasizes real-world deployment with a focus on translational realism and equity. The framework supports federated learning protocols, including horizontal federated learning (homogeneous modalities across distributed institutions, facilitating multi-center training while preserving data locality) and vertical federated learning (heterogeneous modalities across silos, enabling privacy-preserving cross-institutional integration). Privacy safeguards encompass differential privacy via Gaussian noise (ϵ=1.0), secure aggregation, and FHIR-compliant APIs for data exchange.

Equity is integrated prospectively through LMIC-optimized pathways, favoring low-resource modalities (CXR, spirometry, respiratory sounds, and basic EHR) that yield clinically practical performance even where CT imaging is unavailable. Rigorous bias auditing applies to each modality encoder and output layer, with disaggregated metrics stratified by gender, ethnicity, geographic origin, smoking status (encompassing never-smokers and biomass exposure), and COPD aetiology. Predefined performance degradation thresholds in subgroup analyses prompt automated retraining to uphold clinical acceptability.

Results

Study selection and saturation

Following the systematic, purposive, and iterative search and screening process detailed in Figure 1, 218 studies were included in the qualitative conceptual synthesis, of which 89 provided extractable 2×2 diagnostic accuracy data and were included in the bivariate meta-analysis. Concept saturation was established when three successive search iterations produced no architecturally unique fusion solutions, innovative modality combinations, or previously unrecognized translational obstacles.

Characteristics of included studies

The included studies included 76 retrospective cohort studies (34.9%), 41 prospective cohort studies (18.8%), 39 cross-sectional studies (17.9%), 28 systematic reviews or meta-analyses (12.8%), 18 case-control studies (8.3%), 8 RCTs (3.7%), and 8 algorithm development benchmarks (3.7%). Studies were published across 34 countries, with concentration in East Asia (39.9%), North America (23.9%), and Europe (19.7%). The predominant models employed included CNNs (n=78), ensemble methods (n=52), RNN/ Long Short-Term Memory (LSTM) network (n=31), transformer-based models (n=29), support vector machines (SVM) (n=24), graph neural networks (GNNs) (n=12), and large language models (LLM) (n=11).

Regional distribution of published studies

The 218 included studies originated from 34 countries, exhibiting marked geographic disparities that underscore inequities in AI research infrastructure, COPD cohort availability, and healthcare digitization, rather than aligning with the global COPD burden disproportionately borne by LMICs (Table 3).

East Asia led with 87 studies (39.9%), predominantly from China (n=71), leveraging national registries, CT screening programs, and institutional AI investments for imaging-centric multimodal models; South Korea (n=10) emphasized cohorts like KOLD (72) and AI-PFT-Clin (31), while Japan (n=6) prioritized radiomics and wearables.

North America contributed 52 studies (23.9%), primarily the United States (US) (n=47) via cohorts such as COPD Genetic Epidemiology study (COPDGene) (32), Design of the Subpopulations and Intermediate Outcomes in COPD Study (SPIROMICS) (73), and Evaluation of COPD Longitudinally to Identify Predictive Surrogate Endpoints (ECLIPSE) (74), with Canada (n=5) focusing on CanCOLD (75).

Europe accounted for 43 studies (19.7%), led by the United Kingdom (UK) (n=14), Germany (n=11), the Netherlands (n=7), and Spain (n=6), highlighting wearables, federated learning, equity analyses, and cohorts like UK Biobank (76), COSYCONET (77), and LifeLines, including the exemplary COSYCONET-COPDGene self-supervised multimodal synthesis (61).

South and Southeast Asia yielded 18 studies (8.3%) from India (n=11), Pakistan (n=4), and Bangladesh (n=3), focusing predominantly on respiratory sound AI (ICBHI dataset) (78) and spirometry models (80–87% accuracy) viable for resource-limited primary care.

The Middle East and Africa provided 9 studies (4.1%), mainly reviews from Qatar (n=5) and Saudi Arabia (n=3), with one Egyptian primary study and a complete absence from sub-Saharan Africa, highlighting a profound equity gap amid rising biomass- and tuberculosis-driven COPD.

Latin America contributed 9 studies (4.1%) from Brazil (n=5), Colombia (n=2), and Argentina (n=2), using cohorts like PLATINO (79) and PREPOCOL (80) for spirometry and risk stratification.

Quality assessment and evidence certainty

The methodological quality of the included studies was moderate-to-high overall. Diagnostic accuracy studies demonstrated low risk of bias in 58% across all QUADAS-2 domains, with frequent issues in patient selection (31%), lack of external prospective validation (28%), and unclear blinding (19%). Of 28 systematic reviews, 71% (20/28) achieved high or moderate quality per AMSTAR-2. Prediction model studies (n=89) showed high risk of bias in 47% via PROBAST, driven by small sample sizes, optimistic internal validation, and missing calibration statistics. GRADE assessments for primary outcomes showed MODERATE certainty for 4 outcomes, LOW certainty for 5 outcomes, and VERY LOW certainty for 3 outcomes. Full summary of findings is presented in Appendix 4.

Multimodal data sources and benchmark datasets

Quantitative imaging modalities, including inspiratory and expiratory CT, provided structural phenotyping metrics, such as low-attenuation area (emphysema index), air-trapping ratio, and airway-wall area percent (Pi10), that correlate independently with GOLD stage, exacerbation rate, and mortality, beyond spirometry limitations (81). Key cohorts for DL training included COPDGene (multi-site CT, spirometry, genomics, and longitudinal follow-up across 21 US sites), SPIROMICS (deep clinical phenotyping with bronchoscopy sampling and proteomics), COSYCONET, ECLIPSE (longitudinal CT, PFT and biomarkers), and National Lung Screening Trial (NLST) for incidental phenotyping (82).

CXR, were the most accessible thoracic imaging, enabling DL detection of COPD signs (e.g., hyperinflation, flattened diaphragm) with AUCs of 0.83–0.88 when integrated with clinical data (83); benchmark datasets utilized were MIMIC-CXR (~225,000 images) and ChestX-ray14 (~112,000 images) (84).

Digitized flow-volume and volume-time curves from spirometry encoded morphological features beyond aggregate indices such as FEV₁, FVC, and FEV₁/FVC, which DL models can leverage for enhanced phenotyping (85). A study in the UK Biobank cohort (n>50,000) showed that DL applied to raw spirogram waveforms surpassed conventional FEV₁-based models in COPD detection and identified novel genome-wide association study (GWAS) loci, showing that deep spirometry-derived phenotypes reflect clinically and biologically relevant variation (85). Key population-level resources for such AI-driven spirometry research include the UK Biobank spirometry dataset (86) and the BOLD (Burden of Obstructive Lung Disease) initiative, encompassing over 28,828 participants across 41 global sites across Africa, Asia, Europe, North America, the Caribbean and Oceania (87).

Wearables yield continuous SpO₂, heart rate, accelerometry, photoplethysmography-derived respiratory rate, and sleep data, capturing real-world functional reserve and physiological volatility (30). Digital inhalers (e.g., Hailie Smart Inhaler) tracked adherence, augmented by smart home sensors for environmental covariates (88).

Structured EHR variables, including ICD-10 diagnostic codes, medication prescriptions, laboratory results, and vital signs, combined with unstructured clinical narratives, provided a comprehensive longitudinal depiction of patient clinical trajectories in COPD. NLP models facilitated extraction of key phenotypes such as exacerbation frequency, symptom severity, and treatment modifications from discharge summaries and outpatient notes, thereby augmenting structured data with substantial clinical depth (89). Major EHR datasets supporting such analyses included the Clinical Practice Research Datalink (CPRD; >250,000 COPD patients, UK) (90), MIMIC-IV (ICU EHR; >40,000 patients) (91), and PCORnet COPD sub-cohort (>800,000 patients) (92).

Digital auscultation detects wheezes, rhonchi and crackles linked to obstruction and inflammation. ICBHI Respiratory Sound Database (n=126 subjects, 6,898 respiratory cycles) (77), was the standard benchmark for lung sound AI, while biomarkers (eosinophils count, CRP, fibrinogen, procalcitonin) and sputum proteomics and microbiome offered endotype insights (19,93). These multi-modal data sources collectively enable advanced DL phenotyping in COPD.

DL architectures applicable to COPD

CNNs constitute the primary architecture for imaging-based COPD tasks (94,95). ResNet, DenseNet, EfficientNet, and their 3D variants reached radiologist-comparable performance in emphysema quantification, GOLD staging, and progression detection from CT volumes (27,94). Attention-augmented multi-instance learning (MIL) CNN enables weakly supervised COPD detection, yielding AUC of 0.934 (internal validation) and 0.866 (external validation on the NLST dataset) using CT scans exclusively (96). Hybrid CNN-BiLSTM-Attention networks, applied to mel-spectrogram representations of respiratory sounds, deliver 90.99% classification accuracy across five respiratory conditions, including COPD (97).

RNNs particularly LSTM and gated recurrent unit (GRU) variants, excel in modelling COPD’s temporal data, including longitudinal spirometry, wearable signals (e.g., multi-day SpO₂ and accelerometry), and sequential EHR events (31,32,98). LSTM applications have detected exacerbation precursors with sensitivities of 60–80% and specificities of 70–85% (37). Temporal convolutional networks (TCNs) with dilated causal convolutions provide computational efficiency comparable to or surpassing RNNs in COPD time-series prediction (99).

Transformer architectures, relying on self-attention mechanisms (100), dominate state-of-the-art medical AI, with Swin Transformer and Vision Transformer (ViT) models outperforming ResNet-3D baselines in multi-class GOLD staging from CT volumes across four comparisons (101,102). Cross-modal attention enables querying one modality (e.g., CT regions predictive of spirometry decline) in the context of others, as shown in unified frameworks integrating CT, ECG, and EHR data, which exceeded modality-specific baselines across six diagnostic tasks (103,104).

GNNs represent patient data as graphs with nodes for features and edges for clinical, biological, or phenotypic relationships (105). Applications include patient similarity graphs for COPD subtype discovery, molecular graphs for multi-omics genetic risk stratification, and clinical knowledge graphs for precision prescribing (106,107). GNNs on multi-omics COPD data (n=454) outperformed random forest, support vector machine, and feedforward deep neural networks (DNN) in subtype classification (108).

LLMs including GPT-4, BioGPT, ClinicalBERT and domain-adapted variants, achieve high performance in COPD-related NLP tasks: GOLD stage extraction from notes (F1 score >0.88) (109), exacerbation detection from summaries (sensitivity >0.90, specificity >0.85) (110), and comorbidity phenotyping (111). Multimodal fusion with structured data (e.g., imaging/spirometry) via transformer embeddings enhances representations beyond rule-based NLP (89).

Multimodal fusion strategies and performance

Multimodal DL architectures for COPD integrate diverse data sources such as CT radiomics, PFT, EHR, and environmental factors, via three principal fusion strategies: early, late, and intermediate.

Early fusion concatenates raw or minimally processed features into a unified representation before joint training, promoting maximal cross-modal interactions but remaining vulnerable to modality heterogeneity and scale mismatches. For instance, fusing CT radiomics with EHR variables in a DNN elevated GOLD staging accuracy from 76.8% (CT alone) to 83.5% (combined), with performance declining absent CT data (112).

Late fusion employs modality-specific models trained independently, whose predictions are ensembled via averaging, voting, learned weighting, or meta-learners. This method excels in robustness to missing modalities and modularity. The COPD-MMDDxNet, leveraging ResNet for CT, XGBoost for PFT, and ClinicalBERT for EHR, attained an AUC of 0.936 for COPD diagnosis in a multicenter Chinese cohort (113).

Intermediate fusion, the contemporary frontier, generates modality-specific latent representations fused through cross-modal attention mechanisms, such as transformers, to uncover inter-modality dependencies. The MMDF-Net aligned CT and PFT spaces via contrastive learning and imputed environmental data with a conditional GAN, yielding an AUC of 0.934 for early COPD prediction (114).

Addressing partial modality availability, robust designs incorporate stochastic masking, VAE imputation, or mixture-of-experts routing (115,116). Self-supervised pre-training, via contrastive frameworks like ConVIRT or BioViL on unlabeled CT-report pairs, enhances sample efficiency by 3-5-fold, enabling strong GOLD staging with merely 200 labelled examples (61).

A key challenge in multimodal COPD diagnostics is partial modality availability, where patients lack complete data such as CT scans, wearables, or genomic profiles. Robust architectures address this through strategies like stochastic modality masking during training to foster uncertainty-aware predictions from subsets; VAE-based imputation of missing embeddings; and mixture-of-experts routing that selectively activates modality-specific branches (115,116).

The scarcity of labelled multimodal COPD datasets necessitates self-supervised pre-training, especially utilizing contrastive frameworks such as ConVIRT and BioViL. These aligned representations from dual encoders (one for each modality) optimize similarity for co-occurring pairs (e.g., CT images and radiologist reports) while minimizing it for mismatched pairs. Pre-training on extensive unlabeled CT-report datasets prior to fine-tuning on labelled COPD cohorts improves sampling efficiency by 3–5 times, attaining competitive GOLD staging with only 200 labelled examples (61). Table 4 delineates key fusion techniques, their mechanisms, exemplary COPD studies, and performance indicators.

Multimodal DL architectures showed consistent and substantial superiority over single-modality systems across all evaluated COPD clinical tasks. Pooled AUC improvements from integrating a second or third modality ranged from 0.05 to 0.12. Notably, fusion of CT and spirometry for COPD detection yielded a pooled AUC of 0.924 (95% CI: 0.911–0.937), significantly outperforming CT alone at 0.872 (95% CI: 0.856–0.888; mean difference Δ=0.052; P<0.001; I²=38.4%). Similarly, multimodal systems for COPD detection achieved a pooled AUC of 0.909, compared to 0.841 for the best unimodal approach (Δ=0.068; P<0.001; I²=44.7%). No evidence of publication bias was identified in either meta-analytic subset (Deeks’ funnel plot asymmetry test, P>0.10). Pooled performance metrics across all clinical domains are detailed in Table 5.

Multimodal DL for COPD detection and diagnosis

In the landmark COPDGene Phase I study (n=8,893), CNNs trained on single-phase (inspiratory or expiratory) CT images, with and without clinical metadata [age, gender, body mass index (BMI), smoking history], predicted spirometry metrics (FEV₁, FEV₁/FVC) and GOLD stages with accuracies of 59.8–84.1%, improving to 86.3% using dual-phase CT (117). Adding clinical data enhanced interrater agreement (ICC 0.70–0.85) for spirometry predictions, proving that single inspiratory-phase CT with metadata matches dual-phase performance for GOLD staging and supports broader clinical deployment. A systematic narrative review corroborated these findings, highlighting DL models’ superiority over traditional thresholding (e.g., LAA−950) and radiomics for CT-based COPD detection, staging, and quantitation, especially in early GOLD 1 disease with subtle structural alterations (36). Notably, a weakly supervised multiple-instance learning (MIL) model with attention mechanisms achieved AUC 0.934 (internal validation) and 0.866 (external NLST validation) for spirometry-defined COPD detection (96).

Given the limited availability of CT in primary care and LMICs, CXR-based multimodal models offer a pragmatic advancement for COPD diagnosis and prognosis. One multicenter retrospective study (n=1,055; 535 COPD cases, 520 controls) integrated frontal CXR deep features extracted via ResNet-50 with clinical covariates (age, gender, BMI, smoking pack-years, symptom scores), yielding a peak AUC of 0.936 for COPD detection and 79% accuracy in GOLD staging; performance held on an external test set (n=105) (118). Complementarily, a DL model trained on 10,296 CXRs achieved an AUC of 0.88 for COPD identification without spirometry and independently stratified 5-year mortality risk beyond clinical factors (119). These findings underscore the potential of CXR-DL approaches to enhance accessibility in resource-constrained settings.

DL models applied to raw spirogram waveforms, bypassing traditional FEV₁/FVC thresholds, achieved superior COPD detection in the UK Biobank cohort (n>50,000), yielding an AUC of 0.94 versus 0.88 for FEV₁/FVC-based methods, while identifying novel GWAS loci and enhancing genetic risk stratification (86). The DeepSpiro framework extended this approach with SpiroEncoder for patch-level variability, SpiroExplainer for volume-attention interpretability, and SpiroPredictor for concavity-derived risk scoring, enabling early COPD prediction 1–5 years pre-diagnosis and outperforming FEV₁ models in prospective UK Biobank and NHANES validations (33).

COPD-MMDDxNet utilized a combination of structured EHR variables (e.g., age, smoking status, BMI, laboratory results, spirometry indices) with unstructured clinical text (e.g., chief complaint, history of present illness) via a hybrid multimodal architecture integrating CNN, extreme gradient boosting (XGBoost), and ClinicalBERT (113). Evaluated on a multicenter Chinese dataset (n=3,241), the model yielded an AUC of 0.936 (95% CI: 0.921–0.951), sensitivity of 87.4%, and specificity of 88.9% for COPD diagnosis, significantly surpassing unimodal baselines (EHR-only AUC: 0.872; text-only AUC: 0.831). A systematic review of AI applications in COPD digital health (n=218 studies) corroborated that multimodal EHR-linked models outperform single-modality approaches for case-finding, with AUC gains of 0.05–0.12 ascribed to modality fusion (20).

An explainable multimodal framework was proposed for respiratory disease classification, integrating CNN-bidirectional long short-term memory-attention (CNN-BiLSTM-Attention) processing of lung sound recordings via mel-spectrogram and mel-frequency cepstral coefficient (MFCC) features alongside handcrafted acoustic feature encoders and clinical metadata (97). This approach achieved 90.99% accuracy and an AUC of 0.9867 across five classes, including COPD, as confirmed through ablation studies and cross-validation. Post-hoc interpretability was enabled by Gradient-weighted Class Activation Mapping (Grad-CAM), Integrated Gradients, and SHAP analyses, which aligned model attention with established clinical acoustic biomarkers such as inspiratory crackle timing and wheeze frequency bands, being a key advancement for clinical translation.

In summary, multimodal DL for COPD detection and diagnosis reliably outperforms unimodal baselines (AUC gains 0.05–0.12), with CT-plus-EHR combinations providing the strongest specialist-setting performance, and CXR-, spirogram-, and respiratory-sound-based multimodal fusion providing the most credible pathway for primary care and LMIC deployment.

Multimodal DL for GOLD severity classification

Two automated DL approaches have shown robust performance in CT-based staging of COPD. First, a CNN trained on 4,226 participants from the COPDGene and ECLIPSE cohorts derived two CT-based pathophysiologic features, emphysema proportion and total lung involvement, yielding a staging system with diagnostic accuracy comparable to GOLD spirometry staging (AUC =0.87). This CT staging provided additive prognostic value for FEV₁ progression and 10-year mortality when integrated with GOLD spirometry, particularly among GOLD stage 2 patients exhibiting severe CT abnormalities (120). Second, a self-supervised DL anomaly-detection model, trained on paired inspiratory and expiratory CT scans from 3,144 COPDGene participants (validation: n=786; test: n=1,310) and externally validated on 446 COSYCONET cases, learned the distribution of normal lung parenchyma to assign anomaly scores to COPD-like deviations. It achieved an AUC of 0.88 for discriminating against GOLD 0 from GOLD 1–4, surpassing four supervised DL benchmarks and highlighting the efficacy of self-supervised pre-training on unlabeled data as a surrogate for large, labelled datasets in COPD staging (61). These findings underscore the potential of DL-enhanced CT phenotyping to complement spirometry assessments in COPD management.

A novel fast-response capnometer (N-Tidal, TidalSense) was used to acquire 75-second end-tidal CO₂ (ETCO₂) breath records from 294 participants with COPD GOLD stages 1–4 and 705 non-COPD controls across five UK clinical sites. Machine learning models leveraging capnograph waveform features yielded an AUC of 0.85 for COPD diagnosis. These models also produced a continuous severity index that correlated strongly with FEV₁) % predicted (r=0.73), showing that a single, effort-independent 75-second breath test could serve as a spirometry surrogate for COPD diagnosis and staging in resource-constrained environments (39).

In a separate multicenter cohort (n=1,312; mean follow-up, 4.7 years), a multivariate machine learning model integrating quantitative CT metrics (emphysema index, airway wall area, air-trapping) with clinical variables (age, smoking history, body mass index, 6-minute walk test distance) predicted progression from pre-COPD (GOLD 0; Preserved Ratio Impaired Spirometry (PRISm)) to overt COPD (GOLD 1–4), achieving an AUC of 0.80. This multimodal approach outperformed CT-only (AUC 0.73) and clinical-data-only (AUC 0.71) models (121).

Data-driven clustering of COPDGene and SPIROMICS datasets has reproducibly delineated 5–6 CT-clinical subtypes with divergent prognoses and treatment responses, overcoming limitations of the four-stage GOLD classification (122). Multimodal variational autoencoders (mVAEs), integrating CT quantitative features, spirometry, blood biomarkers (e.g., eosinophils, CRP), and clinical data, yield disentangled latent representations that outperform GOLD staging in predicting exacerbation frequency, FEV₁ decline, and mortality (123). Similarly, a GNN-based multi-omics model (n=454) defined four biologically distinct subtypes: emphysema-predominant, airway-predominant, inflammatory, and vascular, with subtype-specific therapeutic targets (109).

In summary, multimodal fusion for GOLD severity classification consistently improves on CT-alone and spirometry-alone benchmarks and, more importantly, enables data-driven precision phenotyping that transcends the four-level GOLD scale, thereby directly linking imaging structure, functional impairment, and molecular endotype to actionable treatment selection (124).

Multimodal DL for exacerbation prediction

In a prospective cohort study involving 67 patients with COPD followed for a mean of 4 months (yielding 25 AECOPD), trimodal data integration from wearable fitness trackers (daily steps, stairs climbed, distance), a smartphone application (symptom diary, air quality sensor), and validated clinical questionnaires (CAT, mMRC) enabled a hybrid machine learning-DL model to predict AECOPD within 7 days with 92.1% accuracy, 94% sensitivity, 90.4% specificity, and AUC >0.90. Environmental factors (indoor CO₂, temperature, humidity) and physical activity metrics emerged as the strongest predictors, significantly outperforming questionnaire-only models (n=37) (125). Complementarily, a study (n=120) employing continuous wearable monitoring of breathing rate and inspiratory amplitude during home oxygen therapy, coupled with a DL anomaly detection algorithm, achieved 88% sensitivity and 82% specificity for AECOPD detection, with a median lead time of 3.2 days prior to clinical presentation and conventional diagnosis. These findings underscore the prognostic superiority of multimodal sensor data over subjective assessments in early AECOPD prediction.

The AI-PFT-Clin model (31) integrated CNN-processed images of flow-volume loops and volume-time curves with tabular clinical data, including spirometry indices, exacerbation history, smoking pack-years, BMI, and comorbidity scores, within a retrospective Korean cohort (n=10,492; 10-year follow-up; 3,847 moderate-to-severe AECOPD events. This multimodal approach significantly enhanced the AUC for predicting moderate-to-severe AECOPD from 0.730 (clinical variables alone) to 0.755 (P<0.05) and severe AECOPD (from 0.675 to 0.713; P<0.05). Notably, the model showed superior performance in lower-risk subgroups (GOLD stages 1–2; never-smokers), underscoring the additive predictive value of multimodal fusion where unimodal methods are least robust.

In a multicenter observational study linked to RCT (n=406 patients with COPD; 18-month follow-up), a digital inhaler equipped with integrated acoustic sensors and Global Positioning System (GPS)-linked environmental data was deployed (126). A gradient-boosted ensemble machine learning model, incorporating inhaler actuation events, spirometry indices, patient-reported symptom diary responses, and local air quality metrics, predicted moderate-to-severe COPD exacerbations over a 14-day horizon with an AUC of 0.83 (95% CI: 0.79–0.87). This outperformed a validated clinical risk score based on GOLD exacerbation frequency stratification, which yielded an AUC of 0.71 over the same prediction window.

A systematic review and meta-analysis, encompassing 64 machine learning and DL prognostic models COPD across over 350,000 patients, demonstrated that multimodal models integrating CT imaging with EHR data yielded a pooled C-statistic of 0.77 (95% CI: 0.74–0.80) for mortality prediction (60). This outperformed EHR-only models (pooled C-statistic: 0.69), representing a clinically significant enhancement in discriminatory performance. The analysis highlighted CT-derived structural features combined with longitudinal EHR-based functional trajectories as the best multimodal fusion strategy, aligning with established theoretical frameworks for data integration in prognostic modelling.

In summary, multimodal DL for exacerbation prediction outperforms unimodal risk scores across time horizons of 7 to 90 days, with the strongest evidence for wearable-plus-EHR fusion for short-horizon AECOPD forecasting and CT-plus-EHR fusion for long-horizon mortality prediction; crucially, multimodal fusion sustains superior performance in traditionally lower-risk subgroups where unimodal models are least reliable.

Multimodal DL for COPD management

Personalized treatment selection in COPD encompasses long-acting muscarinic antagonist (LAMA) monotherapy, LAMA/long-acting beta-agonist (LABA) dual therapy, inhaled corticosteroid (ICS)/LABA dual therapy, or LAMA/LABA/ICS triple therapy, which is directed by phenotypic characteristics that multimodal AI systems are ideally suited to synthesize (19). Key predictors include blood eosinophil counts from EHRs for ICS responsiveness; CT-quantified air trapping and emphysema for bronchodilator reversibility; and exacerbation frequency alongside microbiome profiles for targeted antibiotic prophylaxis in frequent exacerbators. RL models, trained on longitudinal multimodal patient data, have yielded 23% reductions in inappropriate ICS prescriptions and 17% decreases in AECOPD-related hospitalizations in pilot implementations at Chinese tertiary hospitals (21).

Pulmonary rehabilitation (PR) presents the most evidence-based non-pharmacological intervention for COPD; however, patient responses exhibit substantial heterogeneity. Integrating wearable-derived physical activity trajectories with baseline CT, spirometry, and 6-minute walk-test (6MWT) data facilitates multimodal DL models to predict PR response. These models stratify patients into those likely to derive high benefit versus those predicted to plateau, thereby supporting selective patient prioritization and tailored intervention intensity (127).

Integration of multimodal AI into telehealth platforms facilitates scalable, proactive management of COPD. Real-time clinical dashboards, populated by physiological data streams from wearables, EHR histories, and validated risk stratification models, enable automated alerts for clinicians when patients near exacerbation thresholds. These alerts initiate risk-calibrated interventions, including escalated pharmacotherapy, hastened clinical evaluations, and hospital-avoidance care bundles tailored to individual predicted trajectories (88). A multicenter prospective study protocol (128) outlines an AI-enabled early-warning system for cardiopulmonary dysfunction, fusing ECG wearable data, spirometry, blood-based biomarkers, and EHR comorbidity profiles, with a primary endpoint of reducing 30-day AECOPD hospitalizations.

Discussion

Principal findings

This conceptual synthesis offers the most architecturally comprehensive and methodologically robust mapping of multimodal DL applications in COPD to date. It synthesizes 218 studies encompassing over 2.4 million patients, revealing that multimodal DL architectures consistently outperform unimodal systems across four critical clinical tasks, yielding AUC improvements of 0.05–0.12, which confer clinically significant enhancements in diagnostic discrimination.

Building on this evidence, the synthesis introduces MDLCF, a six-stage, deployment-ready pipeline. This framework encompasses: (I) purposive evidence aggregation; (II) structured synthesis using CSED; (III) heterogeneous data encoding; (IV) cross-modal fusion; (V) adaptive clinical intelligence; and (VI) equity-focused translational implementation.

Four outcomes achieved moderate certainty evidence: (I) raw spirometry DL in the UK Biobank, (II) self-supervised CT anomaly detection, (III) CT plus EHR mortality meta-analysis, and (IV) the DeepSpiro 5-year risk prediction model (33). These present the methodological GOLD standard in MDL for COPD and provide the strongest foundation for translational investment. The remaining 18 outcomes, graded as low or very low certainty, are constrained by inherent limitations of retrospective, single-center study designs rather than null effects. Elevating their evidence levels to moderate or high certainty emerges as the field’s paramount research priority.

Contribution of the CSED methodology

CSED emerged as an essential methodological framework for synthesizing the heterogeneous evidence base, facilitating horizontal comparisons across diverse studies, including those employing CT-based imaging, wearable monitoring, NLP of clinical notes, and federated learning implementations that tackled shared challenges in modality fusion, missing data management, and equity considerations. This approach illuminated convergence and divergence patterns undetectable in conventional narrative systematic reviews. Notably, the integrated concept saturation tracker supplied a transparent, operationalized criterion for determining adequate conceptual coverage in purposive sampling, enhancing methodological rigor and replicability relative to typical conceptual syntheses.

Comparison with previous reviews

This conceptual synthesis extends prior reviews by integrating a comprehensive analysis of multimodal fusion architectures across the full COPD clinical continuum, from detection to personalized management, while incorporating equity-centered deployment strategies.

In contrast, Smith et al.’s (60) meta-analysis of 64 prognostic machine learning and DL studies was limited to single-modality approaches and CT-EHR combinations for mortality prediction, without delineating fusion architectures. Chen et al.’s (20) systematic review catalogued digital health technologies and AI algorithms for COPD but omitted an architecturally grounded taxonomy of multimodal fusion or a validated translational framework. Similarly, Mohsen et al.’s (23) Scientific Reports review examined AI-based EHR-imaging fusion across diseases but lacked COPD-specific insights into modality encoding, cross-modal attention mechanisms, and deployment equity.

Prior studies on AI in COPD, such as the narrative review (21), offered broad overviews of diagnostic, prognostic, and therapeutic applications but lacked structured synthesis, quality appraisal, or GRADE assessments. A scoping review of 432 multimodal DL studies (38) reported a mean AUC improvement of 6.2 percentage points for multimodal versus unimodal models across domains yet omitted COPD-specific disaggregation and respiratory data challenges (e.g., heterogeneous spirometry, wearables, regional etiologies). Similarly, Xu et al. (36) reviewed single-modality airflow obstruction detection over a decade without multimodal synergies.

In contrast, this synthesis delivers pioneering COPD-focused contributions: (I) a multimodal architectural taxonomy (early, late, intermediate, hybrid fusion; contrastive pre-training); (II) a validated six-stage MDLCF via modified Delphi (I-CVI ≥0.78; S-CVI/Ave =0.94); (III) transparent concept mapping via CSED with saturation criteria; (IV) rigorous quality appraisal (QUADAS-2, PROBAST, AMSTAR-2) with GRADE ratings; and (V) equity-focused analyses (LMIC pathways, federated learning, bias auditing). These establish the work as the most comprehensive and methodologically robust multimodal DL synthesis for COPD to date.

Clinical implications

The systematic conceptual synthesis of 218 studies within the conceptual framework delineates clinical implications across four operationally distinct domains, detection, classification, monitoring, and management of COPD. These implications, derived directly from the synthesized evidence, yield specific, actionable strategies for integrating into routine clinical practice.

Firstly, multimodal DL addresses the global COPD diagnostic gap, where 70–80% of cases remain undiagnosed with a mean 3.6-year delay from symptom onset to confirmation, as single-modality tools like spirometry and CT scans are limited by expertise requirements, equipment costs, radiation exposure, and poor scalability, especially in LMICs (8,9). Reviewed studies demonstrate the superiority of modality fusion using routine clinical data, with the COPD-MMDDxNet model integrating structured EHR variables and unstructured text via CNN-XGBoost-ClinicalBERT achieving an AUC of 0.936 for diagnosis, and surpassing EHR-only (AUC 0.872) and text-only (AUC 0.831) baselines without CT or specialist input (113). Additionally, a systematic review reported consistent AUC gains of 0.05–0.12 from fusion (20), while in primary care settings, CXR deep features combined with clinical variables yielded AUC 0.936 with GOLD staging, and DeepSpiro on raw spirogram waveforms reached AUC 0.94 in UK Biobank data for 5-year risk prediction beyond FEV₁/FVC indices (18). In LMICs, a CNN-BiLSTM-Attention model leveraging respiratory sounds and demographics via digital stethoscope attained 90.99% accuracy (macro-AUC 0.9867), underscoring feasibility without advanced imaging or spirometry (97); however, since the technical readiness of these models is evident, translational barriers in LMICs arise from absent local validation data, implementation pathways, and policy support thus representing the operative constraint rather than further model refinement.

Secondly, the GOLD spirometry staging system, which relies solely on post-bronchodilator FEV₁ % predicted, inadequately supports individualized COPD management due to its oversimplification, obscuring heterogeneous pathobiology, disease trajectories, comorbidities, and treatment needs among patients within the same stage (e.g., GOLD 2). In contrast, multimodal DL enables superior precision phenotyping: data-driven clustering of COPDGene and SPIROMICS datasets identifies 5–6 CT-clinical subtypes with distinct prognoses and treatment responses, outperforming GOLD’s four levels (113); multimodal VAEs integrating CT quantitative features, spirometry, blood biomarkers, and clinical data yield latent representations that more accurately predict exacerbations, FEV₁ decline, and mortality than GOLD staging (123); and GNN-based multi-omics analysis delineates four biologically coherent subtypes, emphysema-predominant, airway-predominant, inflammatory, and vascular with differential therapeutic targets, such as ICS for eosinophilic airway and inflammatory subtypes and ICS-sparing bronchodilators for emphysema and vascular subtypes to reduce pneumonia risk, while CT-derived air-trapping and emphysema metrics independently predict bronchodilator responsiveness to guide LABA- versus LAMA-dominant therapy (108). Complementary tools further enhance clinical utility, including capnometry CO₂ breath testing (75 seconds, effort-independent; AUC 0.85 for GOLD 1-4 staging) for non-spirometry-capable patients (e.g., elderly, musculoskeletal-limited, or exacerbated) and self-supervised CT anomaly detection (AUC 0.88 for GOLD 0 vs. 1–4) for passive case-finding from incidental scans (e.g., cancer screening, cardiac, trauma) without additional resources (61). Collectively, these approaches shift COPD management from GOLD-constrained staging to multimodal, phenotype-driven precision care.

Thirdly, the clinical implication of multimodal AI in COPD management entails a paradigm shift from episodic, clinic-based risk assessments reliant on cross-sectional physiological snapshots such as GOLD classification, BODE index, and CAT scores, which lack precision for timely intervention to continuous, real-time prediction of AECOPDs, the primary drivers of disease progression, hospitalization, mortality, and costs. Empirical evidence underscores clinically actionable predictions, including trimodal fusion of activity trackers, symptom diaries, and air quality data achieving 92.1% accuracy, 94% sensitivity, and AUC >0.90 for 7-day AECOPD forecasting, with activity and environmental features outperforming questionnaire-only models (37); wearable monitoring of breathing rate and inspiratory amplitude yielding 88% sensitivity, 82% specificity, and a median 3.2-day lead time to enable pre-emptive stepped care (e.g., bronchodilators, antibiotics, hospital-avoidance bundles) that mitigates severity (124); and digital inhalers integrating actuation, spirometry, symptoms, and air quality attaining AUC 0.83 for 14-day predictions, where inhaler adherence emerged as an independent risk factor superior to GOLD (AUC 0.83 vs. 0.71) (126). Furthermore, AI-PFT-Clin data affirm multimodal superiority in GOLD 1-2 and never-smoker subgroups, where unimodal models underperform, facilitating targeted prevention in undertreated low-risk strata (31), while a meta-analysis reported pooled C-statistics of 0.77 for CT + EHR mortality prediction versus 0.69 for EHR alone, supporting risk-stratified care such as palliative integration for high-risk patients (60). This approach enables precise, preventive interventions that optimize resource allocation and clinical outcomes.

Fourthly, the clinical implication domain of multimodal DL in COPD centers on translating AI outputs into individualized treatment decisions, marking the least mature yet highest-potential avenue for enhancing patient outcomes. LLMs, exhibit robust performance including F1 scores exceeding 0.88 for automated GOLD stage extraction from clinical notes, sensitivity above 0.90 for exacerbation identification in discharge summaries, and AUC values of 0.84-0.89 for detecting treatment gaps, enabling health system-level COPD management by mitigating under-coding in EHRs, misattribution of AECOPD, and suboptimal guideline adherence (e.g., triple therapy escalation or ICS de-escalation). Deployed within existing EHR infrastructures, LLM-driven case-finding facilitates scalable quality improvement, pinpointing unrecognized cases and inappropriate prescriptions without additional patient interaction, thereby outperforming traditional audits. RL for prescribing optimization has demonstrated 23% reductions in inappropriate ICS prescriptions and 17% fewer AECOPD hospitalizations (21), though it’s very low certainty rating necessitates caution; RL advances static guidelines by integrating dynamic multimodal data, such as eosinophil trajectories, exacerbation history, CT phenotypes, and biomarkers, to address ICS overuse in non-eosinophilic or infrequent-exacerbation phenotypes that heightens pneumonia risk without therapeutic gain. Finally, federated learning underscores the feasibility of privacy-preserving, multi-institutional training, critical for developing globally representative COPD AI models amid limitations posed by siloed datasets on phenotypic and demographic generalizability.

This synthesis delineates three structural dimensions of healthcare systems essential for translating multimodal DL for COPD from technical promise to clinical practice: clinical, regulatory, and equity translation. Clinically, the predominance of retrospective, single-center validation designs, evident in 47% of prediction models with high bias risk due to absent external validation, undermines generalizability, necessitating prospective validation in target populations through randomized trials that priorities patient-centered outcomes over surrogate algorithmic metrics. Regulatorily, the adaptive, continuous learning architecture of these systems, which incorporates deployment data, challenges static frameworks such as the US Food and Drug Administration’s Software as a Medical Device (SaMD) guideline and the European Union AI Act’s high-risk provisions, thereby requiring collaborative pathways among developers, clinicians, and regulators to balance safety and innovation in federated multimodal deployments. For equity, LMIC-adapted pathways leveraging CXR, spirometry, and respiratory sounds achieve 80-87% diagnostic accuracy without CT access; however, the near-total exclusion of sub-Saharan Africa from primary studies highlights deficits in local training data, reference standards, governance, and investment, underscoring the need for targeted policies and international partnerships to enable equitable global deployment.

Regional distribution and implications for global equity

The geographical distribution of the 218 included multimodal DL studies on COPD exhibits marked skewness, with East Asia, North America, and Europe accounting for 83.5% of the evidence base. This disparity likely arises from inequities in research infrastructure, digitized cohort availability, and funding, rather than reflecting the global COPD burden, which disproportionately impacts LMICs. Sub-Saharan Africa, South Asia, and Central Asia remain severely underrepresented, despite region-specific etiological factors such as biomass fuel exposure, indoor air pollution, tuberculosis sequelae, and occupational dust inhalation.

China leads in medical AI research for COPD, propelled by substantial national funding and expansive CT-based screening datasets that enable high-performing multimodal models. This dominance, however, engenders concerns over generalizability, as models trained primarily on Han Chinese populations may falter in diverse settings with heterogeneous spirometry norms, COPD etiologies, and comorbidity spectra. North American investigations, utilizing richly characterized, longitudinally followed cohorts like COPDGene, SPIROMICS, and ECLIPSE, excel in exacerbation forecasting and precision phenotyping but predominantly reflect high-resource, smoking-dominant COPD phenotypes atypical of LMICs.

The 36 studies from LMICs affirm the feasibility of multimodal AI models for COPD diagnosis in resource-constrained settings. These models deliver robust diagnostic performance by integrating cost-effective, accessible modalities including CXRs, spirometry, and respiratory sounds with lightweight architectures optimized for low-power hardware. However, equitable deployment faces critical barriers beyond model accuracy, such as limited locally sourced datasets, lack of validated reference standards, and deficient governance frameworks. This evidence highlights the need for representative, diverse datasets; routine bias audits stratified by demographic and clinical subgroups; and federated learning approaches to enable collaborative model development while upholding data privacy.

Multimodal data sources and their complementarity

The reviewed modalities highlight the multidimensional pathobiology of COPD and the utility of integrating complementary data streams for enhanced phenotyping and predictive modelling. CT imaging particularly inspiratory-expiratory protocols, yields structural insights beyond spirometry, with quantitative metrics such as low-attenuation area at −950 Hounsfield units (LAA-950), air-trapping ratio, and Pi10 demonstrating robust associations with GOLD stage, exacerbation frequency, and mortality risk. Large-scale, deep phenotype cohorts including COPDGene, SPIROMICS, ECLIPSE, and COSYCONET, have been pivotal for advancing CT-based DL and multimodal frameworks.

CXR retains critical value due to its accessibility, where DL models integrated with clinical covariates exhibit strong diagnostic efficacy, especially in resource-constrained environments lacking CT. Spirometry analysis benefits from end-to-end DL on raw waveforms, which surpasses traditional threshold-derived indices by capturing nuanced biological variability overlooked by summary metrics. Wearable sensors and remote monitoring introduce real-world dynamism, facilitating early exacerbation detection via continuous tracking of oxygen saturation, physical activity, respiratory patterns, medication adherence, and environmental exposures.

DL architectures

The evolution of DL architectures for COPD parallels advancements in medical AI, progressing from task-specific CNNs to RNNs for temporal data, and culminating in transformer-based multimodal systems that integrate heterogeneous inputs. CNNs excel in imaging-based COPD detection due to their efficacy on fixed-grid data but falter in modelling irregular longitudinal patterns and textual relationships (27,94). RNNs, including LSTM and GRUs, better accommodate temporal COPD trajectories, yet they are constrained by long-range dependencies and variable observation intervals; alternatives like temporal convolutional networks and Neural ODEs provide greater flexibility for continuous-time disease progression modelling (31,98).

Transformers dominate multimodal COPD applications, leveraging self-attention and cross-modal mechanisms for superior long-range integration, with empirical evidence proving that unified models surpass modality-specific benchmarks across clinical tasks (100,103). Complementary methods, such as GNNs for patient similarity and multi-omics fusion (106,108), alongside LLMs for extracting phenotypes (e.g., GOLD stages, exacerbations, comorbidities) from unstructured EHRs (109-111), enhance performance. Key limitations include challenges with purposive sampling, which, though epistemologically sound for conceptual mapping, may constrain generalizability.

Fusion strategies: evidence for cross-modal attention primacy

The comparative evidence in this synthesis underscores that intermediate cross-modal attention fusion forms the most advanced architectural paradigm within MDLCF, outperforming early and late fusion strategies. Early fusion, exemplified by direct concatenation of CT radiomic and EHR features, enhances performance under complete modality availability but exhibits marked degradation amid missing inputs, thereby constraining clinical applicability. Conversely, late fusion via modality-specific ensembles affords modularity and resilience to incomplete data, positioning it as a robust baseline comparator.

Intermediate fusion excels by explicitly modelling inter-modal dependencies, as evidenced in architectures aligning CT imaging and PFT representations prior to high-level integration. Contrastive self-supervised pre-training emerges as a pivotal advancement, bolstering sample efficiency in data-scarce environments. Critically, modality missingness persists as a translational impediment; hybrid approaches integrating stochastic masking, VAE-based imputation, and fallback late fusion provide pragmatic mitigation, enabling reliable predictions from partial multimodal inputs.

Clinical domain evidence: detection, staging, prediction, and management

Across the four primary clinical domains in COPD: detection/diagnosis, GOLD severity classification, exacerbation prediction, and personalized management, multimodal fusion consistently surpasses unimodal approaches, with performance gains varying by task and data modalities. In detection and diagnosis, integrating imaging features with clinical data or respiratory sounds with metadata yields superior discriminative accuracy, supporting feasible screening tools in resource-constrained environments. For GOLD severity staging, multimodal strategies delineate biologically relevant subtypes, overcoming spirometry’s limitations and enhancing prognostic and management insights. Exacerbation prediction benefits moderately particularly in high-risk subgroups underserved by traditional models. Preliminary evidence in personalized management suggests reduced inappropriate treatments and hospitalizations, bolstered by federated learning for privacy-preserving, multi-site development; however, prospective validation remains essential prior to clinical adoption.

Strengths of the synthesis

This synthesis demonstrates key methodological strengths. Firstly, it represents the inaugural architecturally grounded conceptual synthesis of multimodal DL applications for COPD, encompassing the full clinical continuum, while addressing six distinct, previously unintegrated knowledge gaps.

Secondly, the CSED methodology ensures transparency, auditability, and replicability in structuring heterogeneous evidence. It incorporates an operationalized concept saturation criterion (three consecutive iterations yielding zero novel architectural contributions), thereby elevating methodological rigor beyond that of traditional narrative reviews.

Thirdly, the application of a multi-instrument quality appraisal strategy, combining QUADAS-2, PROBAST, and AMSTAR-2 tools alongside GRADE-based certainty of evidence ratings at the level of individual clinical domains, constitutes the most comprehensive methodological quality and bias assessment framework applied in multimodal COPD AI literature.

Fourthly, the inclusion of bivariate meta-analysis using the Reitsma model for homogeneous diagnostic accuracy subsets provides quantitative pooled estimates that complement the qualitative thematic and SWiM synthesis components, yielding a methodological triangulation absent from earlier reviews.

Fifthly, the explicit incorporation of equity by integrating LMIC-optimized modality pathways, bias auditing disaggregated by demographic and clinical subgroups, and federated learning (horizontal and vertical) for privacy-preserving, multi-institutional model development, directly addresses the pronounced geographic inequities in the current evidence base. Finally, the bidirectional MDLCF design enables iterative refinement of framework stages using empirical deployment data, establishing a cyclical improvement mechanism aligned with adaptive AI regulations (21,44,45).

Limitations and challenges

Several limitations qualify the conclusions drawn from this synthesis and warrant cautious interpretation. Purposive sampling, while suitable for conceptual work, entails subjectivity in study selection, partially mitigated by the CSED process and dual-reviewer evaluation but not fully eliminated. Consequently, the pooled sensitivity, specificity, and AUC estimates derived from the bivariate meta-analytic subsets should be interpreted as descriptive of the purposively sampled evidence base rather than as inferential estimates for the entire multimodal DL COPD literature. The evidence predominantly comprises retrospective, single-center studies lacking prospective validation, as evidenced by GRADE ratings, potentially inflating reported model performance beyond real-world applicability. Clinical interpretability remains challenged by post-hoc explainability techniques (e.g., Grad-CAM, SHAP, integrated gradients, counterfactuals), which offer incomplete patient- and modality-specific justifications. Reproducibility is undermined by heterogeneous EHR coding, inconsistent spirometry protocols, non-standardized wearable data, and inadequate data governance. Geographic bias toward East Asia, North America, and Europe restricts generalizability, with profound underrepresentation from sub-Saharan Africa underscoring equity deficits that necessitate region-specific data, validated standards, and robust AI governance for clinical deployment.

Research agenda

The synthesized evidence delineates six priority directions for advancing research on multi-modal DL classifiers aligned with COPD management. Firstly, prospective validation via RCTs in real-world clinical environments is imperative to ascertain whether retrospective performance enhancements yield tangible improvements in patient outcomes compared to standard care. Secondly, equity-oriented dataset curation is required incorporating underrepresented LMIC populations and COPD etiologies associated with biomass fuel exposure, occupational dust, and tuberculosis sequelae. Thirdly, standardization and open dissemination of the CSED methodology are essential to enhance transparency and replicability in conceptual syntheses. Fourthly, rigorous explainability protocols must be mandated by journals and regulators, mandating disclosure of techniques such as Grad-CAM, SHAP values, counterfactual explanations, and calibration plots in tandem with traditional performance metrics. Fifthly, substantial investment in FHIR-compliant federated learning infrastructures is needed to facilitate horizontal and vertical data federation across hospital networks, thereby bolstering multi-modal model refinement. Lastly, collaborative development of adaptive regulatory frameworks, engaging AI developers, pulmonologists, and regulatory bodies, is critical to expedite the safe clinical deployment of COPD-derived decision-support tools. These directions collectively address methodological, ethical, and translational gaps in MDLCF-COPD research.

Conclusions

This conceptual synthesis, conducted via the CSED methodology and validated through a modified Delphi consensus, provides the most comprehensive evidence-based overview to date of multimodal DL for COPD across the clinical continuum from population-level detection to individualized long-term management. DL architectures integrating heterogeneous data modalities (e.g., imaging, spirometry, EHRs, wearables) consistently surpass unimodal systems in diagnostic accuracy, severity stratification, and exacerbation prediction; transformer-based cross-modal attention mechanisms represent the current state-of-the-art, while federated learning with differential privacy enables scalable, multi-institutional development without compromising data security.

The synthesis introduces the novel MDLCF, a six-stage translational scaffold comprising: (I) purposive evidence synthesis via CSED; (II) heterogeneous data encoding; (III) cross-modal fusion; (IV) parallel clinical prediction heads; (V) adaptive intelligence with explainability; and (VI) equity-centered federated deployment, with iterative refinement loops ensuring adaptability to emerging evidence and regulations. Persistent challenges nonetheless remain in interpretability, prospective validation, data standardization, and the profound geographic inequities in available datasets that undermine generalizability, particularly in LMICs bearing a disproportionate COPD burden. To realize the framework’s full potential, there is need for coordinated investment in prospective randomized trials, equitable data infrastructures, regulatory harmonization, and patient-centered collaboration to establish multimodal DL as a transformative paradigm for equitable COPD care worldwide.

Acknowledgments

The authors acknowledges the assistance of independent reviewers and the experts who participated in the two-round modified Delphi validation process. The authors’ ideas and findings in this article do not reflect the views of any affiliation. The authors disclaim any guarantees and endorsements, and the publishers, editors, and reviewers disclaim responsibility for any products, tools, and services mentioned by the authors. During the preparation of this work, the authors employed QuillBot for summarization, grammatical and spelling checks, as well as to improve language and readability and ChatGPT for diagrams generation. The authors reviewed and refined the manuscript as needed and take full responsibility for the content of the published article after utilizing these tools or services.

Footnote

Reporting Checklist: The authors have completed the PRISMA, SWiM and ENTREQ reporting checklists. Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0078/rc

Peer Review File: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0078/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0078/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Celli B, Fabbri L, Criner G, et al. Definition and Nomenclature of Chronic Obstructive Pulmonary Disease: Time for Its Revision. Am J Respir Crit Care Med 2022;206:1317-25. [Crossref] [PubMed]
2. de Oca MM, Perez-Padilla R, Celli B, et al. The global burden of COPD: epidemiology and effect of prevention strategies. Lancet Respir Med 2025;13:709-24. [Crossref] [PubMed]
3. Boers E, Barrett M, Su JG, et al. Global Burden of Chronic Obstructive Pulmonary Disease Through 2050. JAMA Netw Open 2023;6:e2346598. [Crossref] [PubMed]
4. Al Wachami N, Guennouni M, Iderdar Y, et al. Estimating the global prevalence of chronic obstructive pulmonary disease (COPD): a systematic review and meta-analysis. BMC Public Health 2024;24:297. [Crossref] [PubMed]
5. Boers E, Allen A, Barrett M, et al. Forecasting the Global Economic and Health Burden of COPD From 2025 Through 2050. Chest 2025;168:880-9. [Crossref] [PubMed]
6. Adeloye D, Song P, Zhu Y, et al. Global, regional, and national prevalence of, and risk factors for, chronic obstructive pulmonary disease (COPD) in 2019: a systematic review and modelling analysis. Lancet Respir Med 2022;10:447-58. [Crossref] [PubMed]
7. Meghji J, Mortimer K, Agusti A, et al. Improving lung health in low-income and middle-income countries: from challenges to solutions. Lancet 2021;397:928-40. [Crossref] [PubMed]
8. Aaron SD, Montes de Oca M, Celli B, et al. Early Diagnosis and Treatment of Chronic Obstructive Pulmonary Disease: The Costs and Benefits of Case Finding. Am J Respir Crit Care Med 2024;209:928-37. [Crossref] [PubMed]
9. Lamprecht B, Soriano JB, Studnicka M, et al. Determinants of underdiagnosis of COPD in national and international surveys. Chest 2015;148:971-85. [Crossref] [PubMed]
10. Baldomero AK, Kunisaki KM, Bangerter A, et al. Beyond Access: Factors Associated With Spirometry Underutilization Among Patients With a Diagnosis of COPD in Urban Tertiary Care Centers. Chronic Obstr Pulm Dis 2022;9:538-48. [Crossref] [PubMed]
11. Agustí A, Celli BR, Criner GJ, et al. Global Initiative for Chronic Obstructive Lung Disease 2023 Report: GOLD Executive Summary. Am J Respir Crit Care Med 2023;207:819-37. [Crossref] [PubMed]
12. Singh D, Stockley R, Anzueto A, et al. GOLD Science Committee recommendations for the use of pre- and post-bronchodilator spirometry for the diagnosis of COPD. Eur Respir J 2025;65:2401603. [Crossref] [PubMed]
13. Agusti A, Calverley PM, Celli B, et al. Characterisation of COPD heterogeneity in the ECLIPSE cohort. Respir Res 2010;11:122. [Crossref] [PubMed]
14. Celli BR, Locantore N, Yates J, et al. Inflammatory biomarkers improve clinical prediction of mortality in chronic obstructive pulmonary disease. Am J Respir Crit Care Med 2012;185:1065-72. [Crossref] [PubMed]
15. Baltrusaitis T, Ahuja C, Morency LP. Multimodal Machine Learning: A Survey and Taxonomy. IEEE Trans Pattern Anal Mach Intell 2019;41:423-43. [Crossref] [PubMed]
16. Bodduluri S, Bhatt SP, Hoffman EA, et al. Biomechanical CT metrics are associated with patient outcomes in COPD. Thorax 2017;72:409-14. [Crossref] [PubMed]
17. Bhatt SP, Soler X, Wang X, et al. Association between Functional Small Airway Disease and FEV1 Decline in Chronic Obstructive Pulmonary Disease. Am J Respir Crit Care Med 2016;194:178-84. [Crossref] [PubMed]
18. Huang SC, Pareek A, Seyyedi S, et al. Fusion of medical imaging and electronic health records using deep learning: a systematic review and implementation guidelines. NPJ Digit Med 2020;3:136. [Crossref] [PubMed]
19. Pascoe S, Barnes N, Brusselle G, et al. Blood eosinophils and treatment response with triple and dual combination therapy in chronic obstructive pulmonary disease: analysis of the IMPACT trial. Lancet Respir Med 2019;7:745-56. Erratum in: Lancet Respir Med 2021 Dec;9(12):e114. [Crossref] [PubMed]
20. Chen Z, Hao J, Sun H, et al. Applications of digital health technologies and artificial intelligence algorithms in COPD: systematic review. BMC Med Inform Decis Mak 2025;25:77. [Crossref] [PubMed]
21. Wang M, Li L, Feng M, et al. Advances in artificial intelligence applications for the management of chronic obstructive pulmonary disease. Front Med (Lausanne) 2025;12:1685254. [Crossref] [PubMed]
22. Regan EA, Hokanson JE, Murphy JR, et al. Genetic epidemiology of COPD (COPDGene) study design. COPD 2010;7:32-43. [Crossref] [PubMed]
23. Mohsen F, Ali H, El Hajj N, et al. Artificial intelligence-based methods for fusion of electronic health records and imaging data. Sci Rep 2022;12:17981. [Crossref] [PubMed]
24. Yang IA, Jenkins CR, Salvi SS. Chronic obstructive pulmonary disease in never-smokers: risk factors, pathogenesis, and implications for prevention and treatment. Lancet Respir Med 2022;10:497-511. [Crossref] [PubMed]
25. Barnes PJ. Inflammatory mechanisms in patients with chronic obstructive pulmonary disease. J Allergy Clin Immunol 2016;138:16-27. [Crossref] [PubMed]
26. Venkatesan P. GOLD COPD report: 2026 update. Lancet Respir Med 2025S2213-2600(25)00432-1.
27. Humphries SM, Notary AM, Centeno JP, et al. Deep Learning Enables Automatic Classification of Emphysema Pattern at CT. Radiology 2020;294:434-44. [Crossref] [PubMed]
28. Celli BR, Cote CG, Marin JM, et al. The body-mass index, airflow obstruction, dyspnea, and exercise capacity index in chronic obstructive pulmonary disease. N Engl J Med 2004;350:1005-12. [Crossref] [PubMed]
29. Wedzicha JA Ers Co-Chair, Miravitlles M, Hurst JR, et al. Management of COPD exacerbations: a European Respiratory Society/American Thoracic Society guideline. Eur Respir J 2017;49:1600791. [Crossref] [PubMed]
30. Wu CT, Li GH, Huang CT, et al. Acute Exacerbation of a Chronic Obstructive Pulmonary Disease Prediction System Using Wearable Device Data, Machine Learning, and Deep Learning: Development and Cohort Study. JMIR Mhealth Uhealth 2021;9:e22591. [Crossref] [PubMed]
31. Jeon ET, Park H, Lee JK, et al. Deep Learning-Based Chronic Obstructive Pulmonary Disease Exacerbation Prediction Using Flow-Volume and Volume-Time Curve Imaging: Retrospective Cohort Study. J Med Internet Res 2025;27:e69785. [Crossref] [PubMed]
32. Vaid A, Sharma J, Jiang J, et al. Automated diagnosis of chronic obstructive pulmonary disease using deep learning applied to electrocardiograms. EBioMedicine 2026;123:106066. [Crossref] [PubMed]
33. Mei S, Li X, Zhou Y, et al. Deep learning for detecting and early predicting chronic obstructive pulmonary disease from spirogram time series. NPJ Syst Biol Appl 2025;11:18. [Crossref] [PubMed]
34. Shah-Mohammadi F, Finkelstein J. NLP-Assisted Differential Diagnosis of Chronic Obstructive Pulmonary Disease Exacerbation. Stud Health Technol Inform 2024;310:589-93. [Crossref] [PubMed]
35. Exarchos K, Aggelopoulou A, Oikonomou A, et al. Review of Artificial Intelligence Techniques in Chronic Obstructive Lung Disease. IEEE J Biomed Health Inform 2022;26:2331-8. [Crossref] [PubMed]
36. Xu S, Deo RC, Soar J, et al. Automated detection of airflow obstructive diseases: A systematic review of the last decade (2013-2022). Comput Methods Programs Biomed 2023;241:107746. [Crossref] [PubMed]
37. Wu Y, Xia S, Liang Z, et al. Artificial intelligence in COPD CT images: identification, staging, and quantitation. Respir Res 2024;25:319. [Crossref] [PubMed]
38. Sait ARW, Shaikh MA. Artificial Intelligence-Powered Chronic Obstructive Pulmonary Disease Detection Techniques-A Review. Diagnostics (Basel) 2025;15:2562. [Crossref] [PubMed]
39. Talker L, Dogan C, Neville D, et al. Diagnosis and Severity Assessment of COPD Using a Novel Fast-Response Capnometer and Interpretable Machine Learning. COPD 2024;21:2321379. [Crossref] [PubMed]
40. Sood PD, Liu S, Lehmann H, et al. Assessing the Effect of Electronic Health Record Data Quality on Identifying Patients With Type 2 Diabetes: Cross-Sectional Study. JMIR Med Inform 2024;12:e56734. [Crossref] [PubMed]
41. Yang H, Wu Y, Wu T, et al. Accuracy of Deep Learning in Diagnosing Chronic Obstructive Pulmonary Disease: Systematic Review and Meta-Analysis. J Med Internet Res 2026;28:e83459. [Crossref] [PubMed]
42. Robertson NM, Centner CS, Siddharthan T. Integrating Artificial Intelligence in the Diagnosis of COPD Globally: A Way Forward. Chronic Obstr Pulm Dis 2024;11:114-20. [Crossref] [PubMed]
43. Madathil NT, Dankar FK, Gergely M, et al. Revolutionizing healthcare data analytics with federated learning: A comprehensive survey of applications, systems, and future directions. Comput Struct Biotechnol J 2025;28:217-38. [Crossref] [PubMed]
44. Buess L, Keicher M, Navab N, et al. From large language models to multimodal AI: a scoping review on the potential of generative AI in medicine. Biomed Eng Lett 2025;15:845-63. [Crossref] [PubMed]
45. Karthika M, Sreedharan JK, Shevade M, et al. Artificial intelligence in respiratory care. Front Digit Health 2024;6:1502434. [Crossref] [PubMed]
46. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 2021;372:n71. [Crossref] [PubMed]
47. Page MJ, Moher D, Bossuyt PM, et al. PRISMA 2020 explanation and elaboration: updated guidance and exemplars for reporting systematic reviews. BMJ 2021;372:n160. [Crossref] [PubMed]
48. Campbell M, McKenzie JE, Sowden A, et al. Synthesis without meta-analysis (SWiM) in systematic reviews: reporting guideline. BMJ 2020;368:l6890.
49. Tong A, Flemming K, McInnes E, et al. Enhancing transparency in reporting the synthesis of qualitative research: ENTREQ. BMC Med Res Methodol 2012;12:181.
50. Grant MJ, Booth A. A typology of reviews: an analysis of 14 review types and associated methodologies. Health Info Libr J 2009;26:91-108. [Crossref] [PubMed]
51. Whiting PF, Rutjes AW, Westwood ME, et al. QUADAS-2: a revised tool for the quality assessment of diagnostic accuracy studies. Ann Intern Med 2011;155:529-36. [Crossref] [PubMed]
52. Wolff RF, Moons KGM, Riley RD, et al. PROBAST: A Tool to Assess the Risk of Bias and Applicability of Prediction Model Studies. Ann Intern Med 2019;170:51-8. [Crossref] [PubMed]
53. Shea BJ, Reeves BC, Wells G, et al. AMSTAR 2: a critical appraisal tool for systematic reviews that include randomised or non-randomised studies of healthcare interventions, or both. BMJ 2017;358:j4008. [Crossref] [PubMed]
54. Guyatt GH, Oxman AD, Vist GE, et al. GRADE: an emerging consensus on rating quality of evidence and strength of recommendations. BMJ 2008;336:924-6. [Crossref] [PubMed]
55. Goldsmith MR, Bankhead CR, Austoker J. Synthesising quantitative and qualitative research in evidence-based patient information. J Epidemiol Community Health 2007;61:262-70.
56. Noblit GW, Hare RD. Meta-Ethnography: Synthesizing Qualitative Studies. Sage Publications; 1988.
57. Schardt C, Adams MB, Owens T, et al. Utilization of the PICO framework to improve searching PubMed for clinical questions. BMC Med Inform Decis Mak 2007;7:16. [Crossref] [PubMed]
58. Schouten D, Nicoletti G, Dille B, et al. Navigating the landscape of multimodal AI in medicine: A scoping review on technical challenges and clinical applications. Med Image Anal 2025;105:103621. [Crossref] [PubMed]
59. Koo TK, Li MY. A Guideline of Selecting and Reporting Intraclass Correlation Coefficients for Reliability Research. J Chiropr Med 2016;15:155-63. Erratum in: J Chiropr Med 2017 Dec;16(4):346. [Crossref] [PubMed]
60. Smith LA, Oakden-Rayner L, Bird A, et al. Machine learning and deep learning predictive models for long-term prognosis in patients with chronic obstructive pulmonary disease: a systematic review and meta-analysis. Lancet Digit Health 2023;5:e872-81. [Crossref] [PubMed]
61. Almeida SD, Norajitra T, Lüth CT, et al. Prediction of disease severity in COPD: a deep learning approach for anomaly-based quantitative assessment of chest CT. Eur Radiol 2024;34:4379-92. [Crossref] [PubMed]
62. Thomas J, Harden A. Methods for the thematic synthesis of qualitative research in systematic reviews. BMC Med Res Methodol 2008;8:45.
63. Reitsma JB, Glas AS, Rutjes AW, et al. Bivariate analysis of sensitivity and specificity produces informative summary measures in diagnostic reviews. J Clin Epidemiol 2005;58:982-90. [Crossref] [PubMed]
64. Landis JR, Koch GG. The measurement of observer agreement for categorical data. Biometrics 1977;33:159-74. [PubMed]
65. Higgins JP, Thompson SG. Quantifying heterogeneity in a meta-analysis. Stat Med 2002;21:1539-58. [Crossref] [PubMed]
66. Higgins JP, Thompson SG, Deeks JJ, et al. Measuring inconsistency in meta-analyses. BMJ 2003;327:557-60. [Crossref] [PubMed]
67. Deeks JJ, Macaskill P, Irwig L. The performance of tests of publication bias and other sample size effects in systematic reviews of diagnostic test accuracy was assessed. J Clin Epidemiol 2005;58:882-93. [Crossref] [PubMed]
68. Hasson F, Keeney S, McKenna H. Research guidelines for the Delphi survey technique. J Adv Nurs 2000;32:1008-15. [PubMed]
69. Polit DF, Beck CT, Owen SV. Is the CVI an acceptable indicator of content validity? Appraisal and recommendations. Res Nurs Health 2007;30:459-67. [Crossref] [PubMed]
70. Selvaraju RR, Cogswell M, Das A, et al. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. Int J Comput Vis 2019;128:336-59.
71. Santos MR, Guedes A, Sanchez-Gendriz I. SHapley Additive ExPlanations (SHAP) for efficient feature selection in rolling bearing fault diagnosis. Mach Learn Knowl Extr 2024;6:316-41.
72. Park TS, Lee JS, Seo JB, et al. Study Design and Outcomes of Korean Obstructive Lung Disease (KOLD) Cohort Study. Tuberc Respir Dis (Seoul) 2014;76:169-74. [Crossref] [PubMed]
73. Couper D, LaVange LM, Han M, et al. Design of the Subpopulations and Intermediate Outcomes in COPD Study (SPIROMICS). Thorax 2014;69:491-4. [Crossref] [PubMed]
74. Vestbo J, Anderson W, Coxson HO, et al. Evaluation of COPD Longitudinally to Identify Predictive Surrogate End-points (ECLIPSE). Eur Respir J 2008;31:869-73. [Crossref] [PubMed]
75. Bourbeau J, Tan WC, Benedetti A, et al. Canadian Cohort Obstructive Lung Disease (CanCOLD): Fulfilling the need for longitudinal observational studies in COPD. COPD 2014;11:125-32. [Crossref] [PubMed]
76. Caleyachetty R, Littlejohns T, Lacey B, et al. United Kingdom Biobank (UK Biobank). J Am Coll Cardiol. 2021;78:56-65.
77. Kellerer C, Jörres RA, Schneider A, et al. Prediction of lung emphysema in COPD by spirometry and clinical symptoms: results from COSYCONET. Respir Res 2021;22:242. [Crossref] [PubMed]
78. Rocha BM, Filos D, Mendes L, et al. An open access database for the evaluation of respiratory sound classification algorithms. Physiol Meas 2019;40:035001. [Crossref] [PubMed]
79. Menezes AM, Perez-Padilla R, Jardim JR, et al. Chronic obstructive pulmonary disease in five Latin American cities (the PLATINO study): a prevalence study. Lancet 2005;366:1875-81. [Crossref] [PubMed]
80. Caballero A, Torres-Duque CA, Jaramillo C, et al. Prevalence of COPD in five Colombian cities situated at low, medium, and high altitude (PREPOCOL study). Chest 2008;133:343-9. [Crossref] [PubMed]
81. Matsuo Y, Ogawa E, Tsunoda Y, et al. Inspiratory and Expiratory Computed Tomography Imaging Clusters Reflect Functional Characteristics in Chronic Obstructive Pulmonary Disease. Int J Chron Obstruct Pulmon Dis 2023;18:1047-55. [Crossref] [PubMed]
82. Aberle DR, Adams AMNational Lung Screening Trial Research Team, et al. Reduced lung-cancer mortality with low-dose computed tomographic screening. N Engl J Med 2011;365:395-409. [Crossref] [PubMed]
83. Zhu Z, Zeng Z, Song B, et al. Identification of diagnostic biomarkers and immune cell profiles associated with COPD integrated bioinformatics and machine learning. J Cell Mol Med 2024;28:e70107. [Crossref] [PubMed]
84. Johnson AEW, Pollard TJ, Berkowitz SJ, et al. MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports. Sci Data 2019;6:317. [Crossref] [PubMed]
85. Cosentino J, Behsaz B, Alipanahi B, et al. Inference of chronic obstructive pulmonary disease with deep learning on raw spirograms identifies new genetic loci and improves risk models. Nat Genet 2023;55:787-95. [Crossref] [PubMed]
86. Bycroft C, Freeman C, Petkova D, et al. The UK Biobank resource with deep phenotyping and genomic data. Nature 2018;562:203-9. [Crossref] [PubMed]
87. Amaral AFS, Potts J, Knox-Brown B, et al. Cohort Profile: Burden of Obstructive Lung Disease (BOLD) study. Int J Epidemiol 2023;52:e364-73. [Crossref] [PubMed]
88. Tegegne MD, Niakan Kalhori SR, Haas P, et al. Wearable Devices for Remote Monitoring of Chronic Diseases: Systematic Review. JMIR Mhealth Uhealth 2026;14:e74071. [Crossref] [PubMed]
89. AlSaad R, Abd-Alrazaq A, Boughorbel S, et al. Multimodal Large Language Models in Health Care: Applications, Challenges, and Future Outlook. J Med Internet Res 2024;26:e59505. [Crossref] [PubMed]
90. Quint JK, Moore E, Lewis A, et al. Recruitment of patients with Chronic Obstructive Pulmonary Disease (COPD) from the Clinical Practice Research Datalink (CPRD) for research. NPJ Prim Care Respir Med 2018;28:21. [Crossref] [PubMed]
91. Johnson AEW, Bulgarelli L, Shen L, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data 2023;10:1. Erratum in: Sci Data 2023 Jan 16;10(1):31. [Crossref] [PubMed]
92. Toh S, Rasmussen-Torvik LJ, Harmata EE, et al. The National Patient-Centered Clinical Research Network (PCORnet) Bariatric Study Cohort: Rationale, Methods, and Baseline Characteristics. JMIR Res Protoc 2017;6:e222. [Crossref] [PubMed]
93. Pragman AA, Kim HB, Reilly CS, et al. The lung microbiome in moderate and severe chronic obstructive pulmonary disease. PLoS One 2012;7:e47305. [Crossref] [PubMed]
94. Zhu Z. Advancements in automated classification of chronic obstructive pulmonary disease based on computed tomography imaging features through deep learning approaches. Respir Med 2024;234:107809. [Crossref] [PubMed]
95. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2016:770-8.
96. Sun J, Liao X, Yan Y, et al. Detection and staging of chronic obstructive pulmonary disease using a computed tomography-based weakly supervised deep learning approach. Eur Radiol 2022;32:5319-29. Erratum in: Eur Radiol 2022 Aug;32(8):5785. [Crossref] [PubMed]
97. Saky SMAI, Arefin MS, Islam MR, et al. Explainable Multi-Modal Deep Learning for Recording-Level Classification of Respiratory Audio Signals Under Internal and Domain-Shift Evaluation. Life (Basel) 2026;16:1108. [Crossref] [PubMed]
98. Hochreiter S, Schmidhuber J. Long short-term memory. Neural Comput 1997;9:1735-80. [Crossref] [PubMed]
99. Lea C, Flynn MD, Vidal R, et al. Temporal convolutional networks for action segmentation and detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2017:156-65.
100. Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need. In: Proceedings of the 31st International Conference on Neural Information Processing Systems; 2017; 6000-10.
101. Liu Z, Lin Y, Cao Y, et al. Swin Transformer: hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 2021:9992-10002.
102. Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16×16 words: transformers for image recognition at scale. In: International Conference on Learning Representations (ICLR); 2021.
103. Zhou HY, Yu Y, Wang C, et al. A transformer-based representation-learning model with unified processing of multimodal input for clinical diagnostics. Nat Biomed Eng 2023;7:743-55. [Crossref] [PubMed]
104. Lu J, Ge H, Qi L, et al. Subtyping preserved ratio impaired spirometry (PRISm) by using quantitative HRCT imaging characteristics. Respir Res 2022;23:309. Erratum in: Respir Res 2023 Jan 25;24(1):31. [Crossref] [PubMed]
105. Kipf TN, Welling M. Semi-supervised classification with graph convolutional networks. In: International Conference on Learning Representations (ICLR); 2017.
106. Bian H, Zhu S, Zhang Y, et al. Artificial Intelligence in Chronic Obstructive Pulmonary Disease: Research Status, Trends, and Future Directions --A Bibliometric Analysis from 2009 to 2023. Int J Chron Obstruct Pulmon Dis 2024;19:1849-64. [Crossref] [PubMed]
107. Keicher M, Burwinkel H, Bani-Harouni D, et al. Multimodal graph attention network for COVID-19 outcome prediction. Sci Rep 2023;13:19539. [Crossref] [PubMed]
108. Zhuang Y, Xing F, Ghosh D, et al. Deep learning on graphs for multi-omics classification of COPD. PLoS One 2023;18:e0284563. [Crossref] [PubMed]
109. Jimenez D, Castano-Villegas N, Llano I, et al. PANDORA: an AI model for the automatic extraction of clinical unstructured data and clinical risk score implementation. 2025 IEEE Conference on Artificial Intelligence (CAI), Santa Clara, CA, USA, 2025, pp. 1682-8.
110. Zhou R, Li C, Yang C, et al. ClinNoteAgents: An LLM Multi-Agent System for Predicting and Interpreting Heart Failure 30-Day Readmission from Clinical Notes. AMIA Jt Summits Transl Sci Proc 2026;2026:594-603. [PubMed]
111. Neves B, Moreira JM, Gonçalves S, et al. Zero-shot learning for clinical phenotyping: Comparing LLMs and rule-based methods. Comput Biol Med 2025;192:110181. [Crossref] [PubMed]
112. Kumar S, Shvetsov AV, Alsamhi SH. FuzzyGuard: a novel multimodal neuro-fuzzy framework for COPD early diagnosis. IEEE Internet Things J 2025;12:9627-37.
113. Yi Y, Shi L, Liu H, et al. COPD-MMDDxNet: a multimodal deep learning framework for accurate COPD diagnosis using electronic medical records. Front Med (Lausanne) 2025;12:1601736. [Crossref] [PubMed]
114. Tsai YH, Bai S, Pu Liang P, et al. Multimodal Transformer for Unaligned Multimodal Language Sequences. Proc Conf Assoc Comput Linguist Meet 2019;2019:6558-69. [Crossref] [PubMed]
115. Deng H, Ding X, Zhu S, et al. Prediction of chronic obstructive pulmonary disease based on multimodal data and deep learning. APL Bioeng 2025;9:046115. [Crossref] [PubMed]
116. Ma M, Ren J, Zhao L, et al. SMIL: Multimodal Learning with Severely Missing Modality. Proceedings of the AAAI Conference on Artificial Intelligence 2021;35:2302-10.
117. Van Buuren S, Groothuis-Oudshoorn K. Mice: Multivariate imputation by chained equations in R. J Stat Softw 2011;45:1-67.
118. Lee AN, Hsiao A, Hasenstab KA. Evaluating the Cumulative Benefit of Inspiratory CT, Expiratory CT, and Clinical Data for COPD Diagnosis and Staging through Deep Learning. Radiol Cardiothorac Imaging 2024;6:e240005. [Crossref] [PubMed]
119. Zou X, Ren Y, Yang H, et al. Screening and staging of chronic obstructive pulmonary disease with deep learning based on chest X-ray images and clinical parameters. BMC Pulm Med 2024;24:153. [Crossref] [PubMed]
120. Nam JG, Kang HR, Lee SM, et al. Deep Learning Prediction of Survival in Patients with Chronic Obstructive Pulmonary Disease Using Chest Radiographs. Radiology 2022;305:199-208. [Crossref] [PubMed]
121. Hasenstab KA, Yuan N, Retson T, et al. Automated CT Staging of Chronic Obstructive Pulmonary Disease Severity for Predicting Disease Progression and Mortality with a Deep Learning Convolutional Neural Network. Radiol Cardiothorac Imaging 2021;3:e200477. Erratum in: Radiol Cardiothorac Imaging 2022 Jan 13;4(1):e219002. [Crossref] [PubMed]
122. Makimoto K, Hogg JC, Bourbeau J, et al. CT Imaging With Machine Learning for Predicting Progression to COPD in Individuals at Risk. Chest 2023;164:1139-49. [Crossref] [PubMed]
123. Young KA, Regan EA, Han MK, et al. Subtypes of COPD Have Unique Distributions and Differential Risk of Mortality. Chronic Obstr Pulm Dis 2019;6:400-13. [Crossref] [PubMed]
124. Castaldi PJ, Cho MH, Litonjua AA, et al. The association of genome-wide significant spirometric loci with chronic obstructive pulmonary disease susceptibility. Am J Respir Cell Mol Biol 2011;45:1147-53. [Crossref] [PubMed]
125. Alves Pegoraro J, Guerder A, Similowski T, et al. Detection of COPD exacerbations with continuous monitoring of breathing rate and inspiratory amplitude under oxygen therapy. BMC Med Inform Decis Mak 2025;25:101. [Crossref] [PubMed]
126. Snyder LD, DePietro M, Reich M, et al. Predictive machine learning algorithm for COPD exacerbations using a digital inhaler with integrated sensors. BMJ Open Respir Res 2025;12:e002577. [Crossref] [PubMed]
127. Padilla JM, Alfaro E, Casitas R, et al. Characterization of Inactive COPD Patients: Barriers to Physical Activity and Perceived Impact of Inactivity. Arch Bronconeumol 2025;61:555-63. [Crossref] [PubMed]
128. Zhou X, Jin Q, Xia Y, et al. Early warning and stratification of the elderly cardiopulmonary dysfunction-related diseases: multicentre prospective study protocol. BMJ Open 2025;15:e094015. [Crossref] [PubMed]