Artificial intelligence and machine learning for the diagnosis of Huntington disease: a narrative review

Introduction

Huntington’s disease (HD) is a progressive, degenerative neurological disorder with a prevalence of approximately 4.88 cases per 100,000 people in Western countries as of 2022 (1). It is characterized by by a gradual decline in motor, cognitive, and psychiatric functions, typically leading to death within 15–20 years after symptom onset (2,3). HD initially presents motor symptoms like chorea, speech impairment (dysarthria), and balance disruptions (postural instability), and over time progresses to affect cognitive and psychiatric functions. This disease significantly affects patients’ quality of life, often with grave financial implications (4). In the context of this review, HD diagnostics refer to the procedural methods or tests employed to determine the presence or absence of the disease, its progression states, or its phenotypic characteristics, using both invasive and non-invasive approaches. This typically includes genetic testing, considering familial history, and identifying motor symptoms (5). HD clinical assessments, commonly derived from evaluations of patients’ motor functions, cognitive capacities, or psychiatric status, serve to ascertain the current state of the disease (5). However, such metrics can vary significantly among patients, leading to diagnostic challenges, and are still in the process of being validated (6). In addition, non-motor symptoms, while often present, are not easily well defined. Moreover, such traditional clinical assessment scores, assessments lack patient-to-patient consistency and suffer from practitioners’ diagnostic discrepancy, limiting their diagnostic use (7,8). Recently, artificial intelligence (AI) models played a crucial role in advancing the diagnosis and management of HD due to the variability, complexity and progressive nature of the disorder. Also, such models, particularly those trained on large, multi-dimensional datasets, can offer significant improvements by accurately analyzing diverse forms of data—including genetic, neuroimaging, and biofluid markers. Not only that, but they could also take the diagnosis levels up by their capability of identifying subtle patterns in data that might be difficult for clinicians to detect, enabling earlier diagnosis, more precise phenotyping, and better monitoring of disease progression. Furthermore, AI can help integrate these different diagnostic modalities, providing a more holistic view of the disease and its varying manifestations across patients. Additionally, as HD progresses over time, AI tools could play a pivotal role in monitoring symptom changes, predicting future health trajectories, and ultimately guiding personalized treatment approaches, all of which are crucial for improving patient care and quality of life. That is why recently there has been a growing interest in using AI to aid in the diagnosis of HD and its various states, capable of interpreting various data types and identifying relevant key biomarkers. However, despite these trends, a comprehensive reference source, encompassing the breadth of AI and machine learning (ML) utilization in HD diagnosis has been lacking.

Existing literature primarily focuses on medical imaging and motor features (9), overlooking potential applications of omics and biofluids data. Additionally, there is a scarcity of reviews examining ML algorithm characteristics and dataset properties in the context of HD diagnosis.

This study addresses these gaps by offering a comprehensive synthesis and analysis of ML utilization in HD diagnosis, differentiation from other NDDs, employed ML models and categories, useful data sources and modalities, preprocessing techniques, ML validation methods and metrics. By synthesizing these aspects from an extensive literature review, the review aims to answer this question: “How can ML and AI models be effectively applied to diagnose HD, identify its various states, distinguish it from control group or other neurodegenerative diseases (NDDs), and uncover key diagnostic biomarkers, considering the limitations of traditional diagnostic methods?”. Therefore, this review will provide researchers and practitioners with insights that not only advance understanding in ML employments for HD diagnostics but also serves as a valuable resource for researchers within the same field who seek to reproduce the work for relevant purpose and using datasets curated from other resources. Additionally, this review is beneficial for researchers in diverse fields, especially relevant to neurodegeneration, facilitates knowledge transfer to those who aim to incorporate ML advancements into the diagnosis of other diseases manifestations. This broad applicability underscores the review’s potential impact on improving symptomatic and presymptomatic diagnostic practices across a wide range of medical fields.

The paper is structured to present the methodology, survey results, discussion, and concluding remarks in a logical flow of information. We present this article in accordance with the Narrative Review reporting checklist (available at https://jmai.amegroups.com/article/view/10.21037/jmai-25-13/rc).

Methods

To achieve the aim of this review, we sought to synthesize a range of key aspects related to the application of AI in HD diagnostics. This included analyzing the study aims, datasets utilized across studies, their characteristics (such as modality, source, and study design), the most used biomarkers and clinical scores, as well as the various ML models employed in HD diagnosis. Additionally, we examined the types and methods of data preprocessing, validation techniques, and performance evaluation methods. Lastly, we explored the research objectives, primary use cases of AI in HD, and the outcomes and future directions of ML techniques in this domain.

To gather relevant papers to answer the review question and achieve the aforementioned aims, multiple resources were searched, including PubMed, IEEE Explore, and Heriot-Watt University Discovery (university’s digital library). In terms of the scope of included studies, we focused on English and peer-reviewed articles and conference proceedings published between January 1, 2010, and February 20, 2023, that explicitly utilized ML for HD diagnostics. The review specifically concentrated on diagnostic approaches, narrowing its focus to papers that dealt with the use of AI in diagnosing HD and its stages, as opposed to those focused on disease progression, prognosis, or treatment. The inclusion of studies was based on the explicit mention of both “Huntington disease” and “Machine Learning” or “Artificial Intelligence” in the title or abstract, ensuring that the papers were directly relevant to the subject at hand. To achieve that, The Boolean search query and process appeared in Figure 1 were applied.

Figure 1 A PRISMA flow diagram illustrating the steps of study identification, review, assessment of eligibility, and final selection for inclusion.

We excluded non-peer-reviewed works, duplicate studies, and those that did not employ ML methods. Studies that focused only on statistical analysis, patient stratification, or biomarker discovery were also excluded, as were those concentrating on HD’s progression or treatment. This selective approach allowed for a more targeted analysis of ML methodologies, specifically within the context of diagnostic applications, while avoiding the inclusion of studies that fell outside the core objective of understanding how AI can improve the diagnosis of HD. Table 1 summarizes the followed search strategy.

By focusing on studies that meet these inclusion criteria, this review aims to provide a comprehensive and in-depth synthesis of the ways in which ML is being used to enhance HD diagnostics. The inclusion of various datasets, features, and models enables a more thorough understanding of the diagnostic landscape in HD research. The findings contribute to ongoing efforts to refine diagnostic practices and improve the accuracy and timeliness of HD detection. This progress was reviewed, and conflicts were resolved with the support of the co-authors.

Results

The PRISMA flowchart illustrated in Figure 1 details the study selection process. Out of an initial 414 studies, 45 duplicates and 258 non-relevant studies were removed. Also excluded were HD studies that focused on prognostics, treatment monitoring, patient stratification, and related biomarker discovery diagnostics-focused studies, as they will be the subject of another review paper. The results were 54 papers remained. Nonetheless, the content contained therein was deemed adequate for inclusion in this context. Figure 2 demonstrates that the interest in using ML for HD diagnostics has been increasing over time, with noticeable surges observed in 2018, 2020, and 2022.

Figure 2 Reviews’ count by years.

Dataset characteristics

Wide insights about existing HD datasets and their characteristics were extracted in this review. This includes source, sample size (here human participants only is counted), data modalities, and types of data collection design of the included studies. Its worth noting that researchers had utilized 40 dataset variations derived from 22 main datasets, with some datasets employed in multiple studies, and the rest were from unknown sources (datasets details and utilization counts are depicted in Table S1 in the appendix). The most commonly utilized datasets pertained to physiological signals, predominantly sourced from the PhysioNet database, and medical imaging data, mostly acquired from Harvard Medical School. The table will also display the relevant papers for each dataset, along with the associated extracted information. The prevalent use of these data sources is attributed mainly to the fact that motor characteristics are the predominant visible symptoms distinguishing HD. To understand the underlying causes of these symptoms and the nature of the disease, researchers often analyze specific brain areas using medical imaging or leverage medical images for gene expression analysis. This inclination arises from the indispensable role that ML plays in the intricate analysis of physiological data. This stands in stark contrast to clinical data, for which conventional statistical methods remain a viable and common analytical approach. The design of data collection studies differs across various research: 87% used cross-sectional designs, 2% longitudinal designs, and the design of the rest were unknown. This diversity be credited to the nature of cross-sectional studies; they are simple in design and execution, efficient in cost and work overhead, and available, making them more commonly utilized than longitudinal studies.

The number of participants in the surveyed literature ranged from 2 to 15,301, with the most having fewer than 100 participants, illustrated in Figure 3. It is notable that 15 out of the total studies examined were specifically dedicated to exploring HD patients in conjunction with control groups. Additionally, 11 studies exclusively focused on HD patients or related samples. Conversely, the remaining subset of surveyed papers, totaling 23 studies, conducted comparisons with other patient diseases, thereby providing valuable insights into the diverse research landscape within the field of HD.

Figure 3 Distribution of participants size in HD diagnostics paper. HD, Huntington’s disease.

The datasets encompassed different data types, broadly categorized into electronic health records (EHR), medical imaging, omics, and others. EHR data, often mentioned explicitly by researchers, was a mix of structured and unstructured representation of patients’ characteristics. Medical imaging, primarily neuroimaging such as structural magnetic resonance imaging (Structural MRI), functional MRI, and Resting state MRI, were either sourced from existing database or taken during the research period. Omics, a large-scale biological study covering genomics, proteomics, and metabolomics, aimed to understand the disease at a molecular level. Physiological signals included measures like speech, electrocardiography (ECG), electroencephalography (EEG), and motor function data, providing neurological and cognitive status of patients. The ‘Others’ category comprised data modalities like biofluids, chemical compounds, and proteins, typically used alongside other data types for a holistic disease view. Figure 4 shows that among various data modalities, 52% of research papers have utilized mostly physiological signals for HD diagnostics. Conversely, it highlights a temporal trend in the increasing utilization of various data modalities, specifically, physiological signals.

Figure 4 Distribution of data modality usage counts in surveyed literature. EHR, electronic health records.

Besides, it presents a noticeable surge in EHR data utilization initiating from the year 2015.

Attributes employed in ML diagnostics

Different categories of participants’ features (individual characteristics), biomarkers (disease-informed features), and clinical scores (disease states identifiers) were used to study HD diagnostics. These include control groups’ and HD patients’ demographics, clinical history, neurophysiological signals, biochemical indicators in body fluids and tissues, and medical images. Besides, complex omics, which includes genomics expression, proteomics, and metabolomics were also often utilized. More than 177 studies utilized biomarkers and features were explicitly mentioned and used in the surveyed papers, with the most frequently used being (in descending order) age, total functional capacity (TFC) score, total motor score (TMS), and cytosine-adenine-guanine (CAG) repeats, as depicted in Figure S1 in Appendix. It is worth noting that statistical features like kurtosis, standard deviation, mean, and skewness were common parameters utilized mainly to describe the raw data of recorded motor signals (10,11) (such as gravitational force (GF) and left/right foot gait analysis) and speech signals (12). On the other hand, structural MRI features, like putamen and caudate volumes, were frequently identified as key indicators for HD diagnostics, as presented in Figure S2 in the Appendix, together with other features considered to be the most reliable characteristics for disease identification (symptomatic and presymptomatic). It is important to mention that T1 and T2 weighted MRI images (where T1 enhances fatty tissue signal and suppresses the water signal, while T2 does the opposite) were employed as is in ML algorithms. Besides, in some inaccessible papers, the abstracts highlight the utilization of MRI brain images, without giving more specification about its characteristics, that is why they were named as such in the figures (13). Clinical scores, measures for HD states and severity, assess different disease aspects and commonly fall into motor, cognitive and behavioral, and functional assessment scores categories (7,8,14). Worth mentioning, the usage of these measures began to emerge in 2014, and as depicted in Figures 5, beside Figures S1,S2 in the appendix, cognitive and behavioral scores, particularly TFC, followed by motor assessment scores, mainly TMS and symbol digit modalities test (SDMT), were most frequently employed and counted as efficient diagnostics and disease states indicators in HD research, accounting for 55% of all 34 HD assessment scores usage.

Figure 5 Paper counts of key clinical scores use in HD diagnostics research. HD, Huntington’s disease; TMS, total motor score; TFC, total functional capacity; SDMT, symbol digit modalities test; cUHDRS, composite unified huntington’s disease rating scale; SIT, stroop interference test; SWR, stroop word reading.

Besides, a new unified assessment that represents all assessment scores classes, called composite unified huntington’s disease rating scale (cUHDRS), has been witnessed in 2021 and 2022. The increased use of HD attributes by time, specifically clinical scores since 2014, underscores HD domain’s progression. Overall, the current emphasis on specific attributes like age and TFC score, and MRI features, highlights their importance as biomarkers in diagnosis. Age was often found correlated with the disease onset, TFC score mostly indicated patient’s motor abilities status by time and informed disease’s stage, and MRI features revealed unique HD related brain changes. Thorough details about the review’s applications for diagnostics will be detailed in the next section.

Research objectives, outcomes, and future perspectives

HD diagnosis in the context of other NDDs

ML techniques have been examined to distinguish not just HD patients from healthy individuals, but also for various other NDD diseases patients (like Parkinson’s, and amyotrophic lateral sclerosis). These investigations used various data modalities and produced a range of model performance outcomes. These modalities, including EEG (15), choreatic and gait patterns features (16-25), MRI (13,26,27), signals of speech (28), and blood transcripts analysis (29), have been employed for the classification of NDD patients. These modalities have consistently demonstrated high classification accuracy and have the capability to assess the severity of each case within each disease. However, further investigation and exploration of the potential of EEG signals are needed (15). In addition, plenty of future work recommended by researchers in the field of ML utilization for NDD disease classification. For example, further utilizations of deep learning models, larger datasets, and further preprocessing techniques may elevate the value of EEG signals in producing high NDD classification results (15). Furthermore, larger gait datasets may excel in the automation of gait analysis and aid in developing more generalized models with accurate diagnosing of movement disorders (17). On the other hand, identifying abnormal brain images related to abnormal NDD disease states using T1-weighted MRI is another interesting field to be explored (26). Also, evaluating the NDD disease classification algorithm for different NDD disease stages is another recommended future work (22). Expanding the computer-aided framework for gait fluctuations detection to include additional diseases and other patient-specific factors is recommended (10). Last, exploring the potential of transfer learning to enhance the accuracy of NDD disease classification, specifically for speech disorders across different languages, is another area worthy of investigation (26).

Motor-based diagnostics for HD

Several studies have investigated HD patients’ movement patterns extracted from clinician assessment and recorded signals to diagnose HD. One study (30) employed support vector machines (SVM) to classify oculomotor features obtained from a four-task psychophysical experiment, successfully distinguishing controls from pre-manifested HD participants using selected features within each task. Another study (31) utilized meta-classifiers and reduced number of movement sensors (accelerometer and gyroscope) placed at the ankles to improve classification accuracy of HD patients’ gait data. Also, the study (32) aimed to develop an automated system for evaluating upper limb movement impairment in HD, using ensembles, and eliminating inconsistencies in clinical assessments. In addition, in this study (33), a computerized behavioral model based on artificial neural networks and fuzzy logic successfully predicted the impaired reaction condition (functional capacity level) in HD patients. Another study (34) developed an automatic method for diagnosing HD using gait dynamics information, achieving an average accuracy of 100%. Moreover, Paper (35) proposed a wireless sensing technology-based monitoring method for HD patients, with SVM and random forest algorithms achieving prediction accuracies of up to 98.0% and 96.7%, respectively. Other interesting work (35), researchers focused on distinguishing left and right foot contacts using multiple features from a gyroscope located at the lower back, achieving high classification accuracy for different health conditions. These studies collectively contribute to the advancement of HD diagnostics and monitoring through the utilization of various techniques and technologies. Some future perspectives, researchers (30) highlight the classification of oculomotor features into disease stages using the SVM algorithm, emphasizing the need for task-specific features. Study (31) reveals that utilizing only two movement sensors per ankle may improve HD patient classification, while study (32) recommend the developing of low-cost automated systems for assessing movement impairment. However, study (33) introduces a computerized behavioral model for predicting impaired reaction conditions, calling for validation with more big datasets. Additionally, study (34) suggests automatic methods for HD diagnosis using gait dynamics, study (11) presents further exploration needs for remote monitoring using wireless sensing technology, and study (35) aims for the future to enhance the classification of gait data from lower back sensors. These findings highlight the potential requirements for improved motor assessment and monitoring for proper diagnosis of HD.

Imaging of brain structures and functions for HD diagnostic

The employment of functional and structural MRI images discovered multiple potentials for enhancing the diagnosis of HD. For example, in study (36), researchers focused on evaluating imaging biomarkers and successfully diagnosed pre-manifestation HD from healthy controls using segmented grey matter features from the fronto-striatal and basal ganglia regions with up to 76% accuracy. Besides, researchers have successfully differentiated HD from Healthy control utilizing from grey matter, caudate nucleus (CAU), and putamen volumes images which found influential in the classification task (37,38). Also, structural, and diffusion-weighted imaging achieved high accuracy in study (39) in distinguishing early manifested HD from healthy controls (38), with key regions being the putamen and caudate. On the other hand, in study (40), researchers explored the potential of quantitative EEG (qEEG) as a biomarker in HD and found it correlated with HD clinical scores. Finally, gene expressions in the CAU prior to HD clinical onset were analyzed in study (41) and similarities were identified between HD CAU and prefrontal cortex samples, suggesting shared disease processes between these regions. Researchers in study (36) propose utilizing longitudinal MRI data to predict individual disease progression based on cross-sectional MRI data. They also suggest exploring multimodal neuroimaging classification to enhance classification model accuracy (38). However, study (37) highlighted that the focus is required for finding the imaging marker for precise clinical diagnosis in HD, considering the use of estimated time to disease onset as a continuous variable. Additionally, future work entails conducting longitudinal studies to evaluate the utility of qEEG indices and investigating progressive changes in HD brain function (40). Furthermore, study (41) proposes further investigation of gene expression patterns in the CAU and prefrontal cortex (BA9) to comprehend shared disease processes and explore existing biological pathways.

Cognitive assessments for HD diagnostics

A computational cognitive biomarker approach utilizing anti-saccade behavioral assessment data and executive control-related features, such as the drift-rate parameter, was effective in detecting subtle differences between non-symptomatic HD subjects and controls, as well as distinguishing between different stages of pre-manifestation HD (42). These features showed correlations with disease progression measures, TFC and TMC. Furthermore, despite that age, CAG repeat number, and the disease burden score revealed a constrained ability to discern between normal cognitive measures and those indicating the impact of the disease, significant genetic association between better cognitive performance and minor alleles of NCOR1 and ADORA2B were found in pre-manifestation HD individuals (43). In the field of dementia classification, ML algorithms, particularly Gradient Boost, achieved a high accuracy of 0.96 in classifying patients into ‘demented’, ‘non-demented’, and ‘converted’ groups (12).

Omics analysis in HD diagnostics

Utilizing Artificial neural network and data extracted from blood tests, detecting subtle differences between HD subjects and controls was feasible, but this is not the case for non-symptomatic HD (44). Metabolic alterations were identified as an important part of HD etiology, with similarities observed in the response of striatum, liver tissue, and plasma to the CAG repeat allele associated with HD, suggesting tissue-specific metabolic changes that also vary under different dietary conditions (45). Besides, mutant huntingtin (mHtt) found to be disruptive to critical processes involved in the pathogenesis of HD, including ribosomal DNA condensation and DNA repair, resulting in the accumulation of DNA damage and neuronal cell apoptosis. It is also important to highlight the importance of informative value (IV) of motifs in predicting interactors of mHtt protein (46). Deep learning techniques were employed for multi-label classification of stem cell microscopy images, facilitating the identification and categorization of different cell colony subtypes related to the HD phenotype (47). The findings of this study provide a foundation for future research in the field of automated detection of changes in cellular behavior. Additionally, ensemble consensus-guided unsupervised feature selection improved disease-associated gene prediction accuracy and stability, with enrichment analysis indicating the involvement of postsynaptic density and membrane, synapse, and cell junction during HD progression (48). Further gene expression microarray studies are required to identify potential genetic biomarkers involved in HD progression (49).

Miscellaneous applications for HD diagnostics

Here, other various diverse ML applications related to diagnosing HD are presented. For example, the investigation of psychosis in HD revealed the presence of alcohol use disorders, depression, violent/aggressive behavior, and perseverative/obsessive behavior as characteristics linked to the unique phenotype of HD patients presenting with psychosis (50). Furthermore, a study focusing on weight loss in the premanifest stage of HD found a strong association between severe weight loss and specific cognitive, psychiatric, and functional abilities (51). In addition, the development of a specialized text predictor for patients with HD demonstrated satisfactory performance even with words of varying lengths, providing support for individuals with HD. Additionally, several studies in the field of speech and language focus on HD and dysarthria. One study introduces a new set of acoustic features to estimate hypernasality in dysarthria, which demonstrates generalizability across different dysarthria corpora (23). Another study presents an optimized approach using voice recordings to detect the presence of HD, achieving high precision and accuracy above 99% with a prediction time below 1 second (52). Additionally, an investigation into emotion expression in HD reveals difficulties in expressing emotions through voice and language for HD patients compared to pre-manifestation HD participants and controls, indicating reduced emotional expression in HD patients (53).

ML models and dataset pre-processing techniques

This section will provide detailed insights into the ML techniques and the variety of preprocessing methods employed within the scope of this study. Various types of ML methods have been applied in literature, and they can be generally categorized into several groups, including supervised, semi-supervised, unsupervised, and reinforcement learning (54). Supervised learning employs labeled data to develop a model capable of predicting the output variable for novel input data and can include deep learning methods that utilize multi-layered artificial neural networks to interpret complex data. Semi-supervised learning blends supervised and unsupervised learning and leverages both labeled and unlabeled data (54). Unsupervised learning algorithms identify and extract patterns and features from unlabeled data and reinforcement learning involves an agent learning through interaction with the environment to maximize a reward signal (54). Besides these well-known ML categories, they were also accompanied by ensemble and probabilistic models (or undirected graphical models) in recent years and that is depicted in Figure 6. To describe each in short, ensemble methods like bagging, boosting, or stacking merge various models to enhance accuracy and mitigate overfitting, while voting uses majority rule for predictions (55). Probabilistic graphical techniques express relationships among variables graphically, including models like hidden markov models (HMMs) and event based models (EBMs) (54). As illustrated in Figure 6, the employment of supervised ML techniques was the most prevalent, accounting for 84% of all techniques in the surveyed studies. This was followed by unsupervised and ensemble models with each representing 7% and 6% of all utilized models, respectively. Each ML type has common usage scenarios, including classification, regression, feature selection, dimensionality reduction, event/rule extraction, visualization, reversal or experiential learning, and discerning feature/state relationships, which started gaining attention from 2016 onwards (Figure 6). Figure S3 shows that classification stands as the most common application throughout all ML categories, specifically for supervised problem types. For example, unsupervised learning models were used mainly for clustering, features’ grouping and identification. More elaboration on their applications for HD diagnosis are summarized in ML applications section. HD diagnosis was performed in many ways and for different purposes. These include studying the impact of omics, brain structure and functions, psychiatric, cognitive, speech and text in diagnostics of HD patients, their variabilities at different disease stages, and their abilities to differentiate them from other NDDs disease patients and control groups.

Figure 6 The spread of ML types in HD diagnostics search, with yearly publication counts. HD, Huntington’s disease; ML, machine learning.

Table 2 summarizes these applications aspects, utilized ML types and use cases. Around 58 distinct ML algorithms, or combinations of algorithms, were employed for HD diagnostics, and as depicted in Figure S3, decision tree-based algorithms with all their variations were the most commonly used such as random survival forests, random forest, boosted trees model, gradient boosting decision tree, light and extreme gradient boosting decision tree (XGBoost), and extremely randomized trees. Next were SVM and k-nearest neighbor algorithms were recurrently utilized in the literature. Eleven deep learning algorithms were utilized, mostly convolutional neural networks (CNNs)-based algorithms (65), such as one-dimensional and two-dimensional CNN (i.e., 1D-CNN and 2D-CNN, in respectively), Residual Network 18 (ResNet-18), Visual Geometry Group (VGG), and AlexNet.

Various preprocessing techniques were employed across multiple data modalities, with physiological signals being the predominant modality compared to others, as shown in Figure 7. Those preprocessing groups, as defined in study (66) are mainly data transformation (changing data format), data scaling (modifying the data values into standard range), data reduction (simplifying data while retaining its core effect), data cleaning (eliminating instances flaws and errors), data partitioning (subsetting dataset), and data augmentation (increasing dataset volume).

Figure 7 Distribution of various preprocessing techniques by year and across different data modalities.

Figure 7 indicates that in the last five years there has been a focus on multiple preprocessing techniques, and the adoption of data transformation and partitioning techniques for HD diagnostics started to gain traction. Table 3 presents these techniques, grouped according to their preprocessing categories, and the corresponding data modalities.

Validation techniques and performance evaluation metrics

In the reviewed literature, the evaluation of model performance depended on employing the right validation techniques to understand model’s adaptability to generalize to new, unseen data, rather than the risks associated with overfitting. Researchers have interchangeably employed cross-validation and datasets splitting into training and testing subsets, with a slight preference towards cross-validation techniques. This makes sense since, as previously reported, the utilized sample sizes were quite small. Among these techniques, k-fold cross-validation, specifically leave-one-out cross-validation and 10-fold-cross-validation approaches are the most prevalent. By contrast, when exploring dataset splitting practices, a recurring pattern emerges: a train-test split average ratio approximately of 75% for training and 25% for testing.

Moving to ML evaluation metrics, Figure 8 and Table 4 show commonly used ML evaluation metrics for HD diagnostics in the surveyed articles. They are in descending order: accuracy (29%), sensitivity (16%), and specificity (14%). Over all data modalities, signals and omics applied most of the metrics in comparison to others. Other metrics like average accuracy, precision, F1 score, root mean square error (RMSE), mean absolute error (MAE), and area under the curve (AUC), area under the receiver operating characteristic curve (AUC-ROC) were also observed. These findings underscore the significance of diverse evaluation scales in validating ML models for HD diagnostics related tasks. Notable accuracy and sensitivity were the most used performance evaluation metrics across all data modalities, specifically with physiological signals and imaging data modalities. Advanced metrics like AUC, ROC, F1 score, and error-based measures (MSE, MAE) were used less frequently and primarily in physiological signal and omics studies. As previously noted, HD diagnosis has been conducted through various methods and for diverse objectives. This encompasses investigating the power of omics, the structure and functionality of the brain, psychiatric evaluations, cognitive assessments, and speech and text analyses in detecting the present of the disease, and its onset. Besides, such diagnostics’ features were examined to classify HD patients at various disease stages, compare them to control groups, and distinguish them from patients with other NDD diseases.

Figure 8 Distribution of utilized ML evaluation metrics (top right chart), by year (top left chart), and by data modalities (bottom chart) in HD diagnostics. AUC, area under the curve; HD, Huntington’s disease; ML, machine learning; MAE, mean absolute error; MSE, mean squared error; ROC, receiver operating characteristic; RMSE, root MSE.

Discussion

Since 2010, there has been a significant evolution in the application of ML algorithms for the diagnostics of HD, with a particular emphasis on supervised ML techniques. Most of the research has focused on classical ML approaches, with a predominant utilization of physiological signals data modalities. This focus is logical, given that HD manifestation is primarily characterized by motor disorders, which are phenotypically measurable in a noninvasive manner. In addition, those modalities that are employed by institutions engaged in the manipulation of physiological signals, were found also the most inclined to employ ML methodologies. Brain and omics image datasets have proven to be valuable resources for understanding the progression and neurodegeneration process to a certain extent, highlighting how it influences various parts of the brain. These studies also centered around specific datasets, with putamen and caudate brain volumes and motor clinical score categories (primarily the TMS score) emerging as the most significant biomarker for tracking the progression or predicting the onset of different symptomatic states of the disease. Among the various supervised ML algorithms, decision tree algorithms and their variants have been the most utilized and performant models in this context. Their reliability, efficiency, and interpretability ability by medical experts are well-established, as also noted by study (64). In addition to the choice of algorithms, data transformation has been extensively applied as a preprocessing technique. This step helps in extracting representative patterns from signals or images, which are essential for successful downstream tasks like classification or regression. Deep learning methods, particularly CNNs and long short-term memory (LSTM) networks, have become prominent in this area due to their superior ability to extract meaningful patterns from complex data modalities through representation learning (68).

Furthermore, accuracy remains the prevalent performance metric, and cross-validation techniques have been widely adopted for model performance validation, especially in scenarios with small datasets. The logic behind the widespread use of cross-validation is its proven efficiency in small dataset contexts and its effectiveness in estimating ML model performance and superiority compared to the hold-out set method, even for big datasets (69). For larger datasets, although cross-validation may require more computational resources than simpler methods like hold-out validation, its use of multiple representative data splits provides more reliable performance estimates and better generalization (69).

Researchers have made notable efforts to ensure interpretation of their results by aligning findings with clinical experts’ interpretations and existing literature. Despite that, there remains a gap in the generalizability and interpretability of the ML algorithms utilized. Future work should focus on enhancing the interpretability of these algorithms to ensure their clinical utility and acceptance. Moreover, further exploration of omics data is essential for a deeper understanding of HD. Additionally, the application of advanced deep learning models and the use of longitudinal data across various modalities hold significant promise for understanding the progression and course of disease states and remain an intriguing topic for future investigations.

From clinical practice perspectives, the findings from this review suggest that while current ML tools show promise for HD diagnostics, the clinical trust adoption of such technologies is still in the early stages. Additionally, policy efforts to support the use of such datasets have begun, but they need to be expanded to ensure greater data accessibility, generalizability, and reliability employment of these models in various clinical settings. Further, care providers should consider incorporating these methods, especially noninvasive ones, like those focused on motor assessments and neuroimaging, but after multiple clinical trails with a keen understanding of their limitations due to dataset variability and the absence of large, diverse samples. Furthermore, the integration of multi-modal data, including EHRs, omics, and neuroimaging, could lead to more precise diagnostic frameworks that assist clinicians in identifying HD in its early stages. For future research, the effort should be more toward include larger, more diverse participants from longitudinal studies to validate and refine diagnostic models, and unlock various disease states. Additionally, exploring advanced ML techniques such as deep learning models, transformers, reinforcement learning and transfer learning could further enhance the capability of these models to generalize across various NDDs, improving the detection not only of HD but also other conditions with overlapping symptoms. Implementing all the aforementioned strategies could ultimately help decipher the abnormal progression of such diseases, understand their dynamic developments, and, as a result, facilitate more tailored treatments and improved care and management of HD patients.

Conclusions

This paper presents a review of studies that have employed ML techniques across various facets of HD diagnostics. A notable limitation was the restricted accessibility of several key papers, though their abstracts highlighted interesting details that were covered in this literature (12,13,18-20,34,56). This study carefully gathered important information on the ML use cases for HD diagnostics, adopted ML and dataset characteristics, data preprocessing strategies, potential biomarkers and clinical measures, and ML validation and evaluation techniques. The insights from this review substantially enrich the current comprehension of ML’s role in diagnosing HD and provide guidance for forthcoming studies. Leveraging ML in HD diagnostics has significant implications, promising advancements in disease detection, comprehension, clinical trial facilitation, and ultimately, improving patients’ quality of life.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://jmai.amegroups.com/article/view/10.21037/jmai-25-13/rc

Peer Review File: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-25-13/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-25-13/coif). M.V. reports grant support from EPSRC for unrelated projects in ALS, Huntington’s disease, and lung cancer. She is a co-inventor on a patent related to lung cancer diagnostics (unrelated to the current manuscript) and received travel support from a company for an event not related to this work. She also serves as an Academic Editor for Nature Communications Biology and receives a modest honorarium. The other 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. Medina A, Mahjoub Y, Shaver L, et al. Prevalence and Incidence of Huntington’s Disease: An Updated Systematic Review and Meta-Analysis. Mov Disord 2022;37:2327-35. [Crossref] [PubMed]
2. Naarding P, Kremer HP, Zitman FG. Huntington’s disease: a review of the literature on prevalence and treatment of neuropsychiatric phenomena. Eur Psychiatry 2001;16:439-45. [Crossref] [PubMed]
3. Solberg OK, Filkuková P, Frich JC, et al. Age at Death and Causes of Death in Patients with Huntington Disease in Norway in 1986-2015. J Huntingtons Dis 2018;7:77-86. [Crossref] [PubMed]
4. Roos RA. Huntington’s disease: a clinical review. Orphanet J Rare Dis 2010;5:40. [Crossref] [PubMed]
5. Reilmann R, Leavitt BR, Ross CA. Diagnostic criteria for Huntington’s disease based on natural history. Mov Disord 2014;29:1335-41. [Crossref] [PubMed]
6. Mestre TA, Bachoud-Lévi AC, Marinus J, et al. Rating scales for cognition in Huntington’s disease: Critique and recommendations. Mov Disord 2018;33:187-95. [Crossref] [PubMed]
7. Winder JY, Achterberg WP, Marinus J, et al. Assessment Scales for Patients with Advanced Huntington’s Disease: Comparison of the UHDRS and UHDRS-FAP. Mov Disord Clin Pract 2018;5:527-33. [Crossref] [PubMed]
8. Winder JY, Roos RAC, Burgunder JM, et al. Interrater Reliability of the Unified Huntington’s Disease Rating Scale-Total Motor Score Certification. Mov Disord Clin Pract 2018;5:290-5. [Crossref] [PubMed]
9. Tăuţan AM, Ionescu B, Santarnecchi E. Artificial intelligence in neurodegenerative diseases: A review of available tools with a focus on machine learning techniques. Artif Intell Med 2021;117:102081. [Crossref] [PubMed]
10. Fraiwan L, Hassanin O. Computer-aided identification of degenerative neuromuscular diseases based on gait dynamics and ensemble decision tree classifiers. PLoS One 2021;16:e0252380. [Crossref] [PubMed]
11. Zhu Q, Guan L, Khan MB, et al. Monitoring of Huntington’s Disease Based on Wireless Sensing Technology. Appl Sci. 2020;10:870. [Crossref]
12. Arora A, Gourisaria MK, Chatterjee R. Classification and Analysis of Dementia using Machine Learning Algorithms. In: 2022 IEEE International Conference on Electronics, Computing and Communication Technologies (CONECCT). IEEE; 2022. p. 1-6.
13. Ahmad F, Hameed Z, Rehman S, et al. Normal/Abnormal Magnetic Resonant Brain Images Detection Based on Supervised Classifiers. In: 2020 International Conference on Computing and Information Technology (ICCIT-1441). IEEE; 2020. p. 1-4.
14. Oosterloo M, de Greef BTA, Bijlsma EK, et al. Disease Onset in Huntington’s Disease: When Is the Conversion? Mov Disord Clin Pract 2021;8:352-60. [Crossref] [PubMed]
15. Rodrigues, Miguel P, Teixeira, et al. Classification of Electroencephalogram signals using Artificial Neural Networks. In: 2010 3rd International Conference on Biomedical Engineering and Informatics. IEEE; 2010. p. 808-12.
16. Berke Erdaş Ç, Sümer E, Kibaroğlu S. CNN-based severity prediction of neurodegenerative diseases using gait data. Digit Health 2022;8:20552076221075147. [Crossref] [PubMed]
17. Banaie M, Pooyan M, Mikaili M. Introduction and application of an automatic gait recognition method to diagnose movement disorders that arose of similar causes. Expert Syst Appl 2011;38:7359-63. [Crossref]
18. Mannini A, Trojaniello D, Cereatti A, et al. A Machine Learning Framework for Gait Classification Using Inertial Sensors: Application to Elderly, Post-Stroke and Huntington’s Disease Patients. Sensors (Basel) 2016; [Crossref] [PubMed]
19. Vajiha Begum SA, Rani MP. Recognition of Neurodegenerative Diseases with Gait Patterns using Double Feature Extraction Methods. In: 2020 4th International Conference on Intelligent Computing and Control Systems (ICICCS), Madurai, India, 2020, pp. 332-8.
20. Heikal M, Eldawlatly S. An Ensemble Classification Technique of Neurodegenerative Diseases from Gait Analysis. In: 2020 15th International Conference on Computer Engineering and Systems (ICCES), Cairo, Egypt, 2020, pp. 1-6.
21. Negi PC, Negi S, Sharma N. Gait Analysis-Based Identification of Neurodegenerative Diseases Using Machine Learning Techniques. In: 2022 International Conference on Advances in Computing, Communication and Materials (ICACCM), Dehradun, India, 2022, pp. 1-6.
22. Nam Nguyen QD, Liu AB, Lin CW. Development of a Neurodegenerative Disease Gait Classification Algorithm Using Multiscale Sample Entropy and Machine Learning Classifiers. Entropy (Basel) 2020; [Crossref] [PubMed]
23. Saxon M, Tripathi A, Jiao Y, et al. Robust Estimation of Hypernasality in Dysarthria with Acoustic Model Likelihood Features. IEEE/ACM Trans Audio Speech Lang Process 2020;28:2511-22. [Crossref] [PubMed]
24. Marziyeh Ghoreshi Beyrami S, Ghaderyan P. A robust, cost-effective and non-invasive computer-aided method for diagnosis three types of neurodegenerative diseases with gait signal analysis. Measurement 2020;156:107579. [Crossref]
25. Liu Y, Oubre B, Duval C, et al. A Kinematic Data-Driven Approach to Differentiate Involuntary Choreic Movements in Individuals With Neurological Conditions. IEEE Trans Biomed Eng 2022;69:3784-91. [Crossref] [PubMed]
26. Zhang Y, Wu L. An MR brain images classifier via principal component analysis and kernel support vector machine. Progress In Electromagnetics Research 2012;130:369-88. [Crossref]
27. Neffati S, Taouali O. An MR brain images classification technique via the Gaussian radial basis kernel and SVM. In: 2017 18th International Conference on Sciences and Techniques of Automatic Control and Computer Engineering (STA). IEEE; 2017. p. 611-6.
28. Vásquez-Correa JC, Rios-Urrego CD, Arias-Vergara T, et al. Transfer learning helps to improve the accuracy to classify patients with different speech disorders in different languages. Pattern Recognit Lett 2021;150:272-9. [Crossref]
29. Huseby CJ, Delvaux E, Brokaw DL, et al. Blood Transcript Biomarkers Selected by Machine Learning Algorithm Classify Neurodegenerative Diseases including Alzheimer’s Disease. Biomolecules 2022; [Crossref] [PubMed]
30. Miranda Â, Lavrador R, Júlio F, et al. Classification of Huntington’s disease stage with support vector machines: A study on oculomotor performance. Behav Res Methods 2016;48:1667-77. [Crossref] [PubMed]
31. Acosta-Escalante FD, Beltran-Naturi E, Boll MC, et al. Meta-classifiers in huntington’s disease patients classification, using iPhone’s movement sensors placed at the ankles. IEEE Access 2018;6:30942-57.
32. Bennasar M, Hicks YA, Clinch SP, et al. Automated Assessment of Movement Impairment in Huntington’s Disease. IEEE Trans Neural Syst Rehabil Eng 2018;26:2062-9. [Crossref] [PubMed]
33. Lauraitis A, Maskeliūnas R, Damaševičius R. ANN and Fuzzy Logic Based Model to Evaluate Huntington Disease Symptoms. J Healthc Eng 2018;2018:4581272. [Crossref] [PubMed]
34. Paula Felix J, Henrique Teles Vieira F, Da Silva Vieira G, et al. An automatic method for identifying huntington’s disease using gait dynamics. In: 2019 IEEE 31st International Conference on Tools with Artificial Intelligence (ICTAI), Portland, OR, USA, 2019, pp. 1659-63.
35. Ullrich M, Kuderle A, Reggi L, et al. Machine learning-based distinction of left and right foot contacts in lower back inertial sensor gait data. Annu Int Conf IEEE Eng Med Biol Soc 2021;2021:5958-61. [Crossref] [PubMed]
36. Rizk-Jackson A, Stoffers D, Sheldon S, et al. Evaluating imaging biomarkers for neurodegeneration in pre-symptomatic Huntington’s disease using machine learning techniques. Neuroimage 2011;56:788-96. [Crossref] [PubMed]
37. Mason SL, Daws RE, Soreq E, et al. Predicting clinical diagnosis in Huntington’s disease: An imaging polymarker. Ann Neurol 2018;83:532-43. [Crossref] [PubMed]
38. Chu C, Bandettini P, Ashburner J, et al. Classification of neurodegenerative diseases using Gaussian process classification with automatic feature determination. In: 2010 First Workshop on Brain Decoding: Pattern Recognition Challenges in Neuroimaging, Istanbul, Turkey, 2010, pp. 17-20.
39. Lavrador R, Júlio F, Januário C, et al. Classification of Huntington’s Disease Stage with Features Derived from Structural and Diffusion-Weighted Imaging. J Pers Med 2022; [Crossref] [PubMed]
40. Odish OFF, Johnsen K, van Someren P, et al. EEG may serve as a biomarker in Huntington’s disease using machine learning automatic classification. Sci Rep 2018;8:16090. [Crossref] [PubMed]
41. Agus F, Crespo D, Myers RH, et al. The caudate nucleus undergoes dramatic and unique transcriptional changes in human prodromal Huntington’s disease brain. BMC Med Genomics 2019;12:137. [Crossref] [PubMed]
42. Wiecki TV, Antoniades CA, Stevenson A, et al. A Computational Cognitive Biomarker for Early-Stage Huntington’s Disease. PLoS One 2016;11:e0148409. [Crossref] [PubMed]
43. White A, Li Y, Kuan WL, et al. F31 Understanding cognitive heterogeneity beyond disease burden in Huntington’s disease. Journal of Neurology, Neurosurgery, and Psychiatry 2018;89:A50.3-A51.
44. Alfonso Perez G, Caballero Villarraso J. Neural Network Aided Detection of Huntington Disease. J Clin Med 2022; [Crossref] [PubMed]
45. Carroll JB, Deik A, Fossale E, et al. HdhQ111 Mice Exhibit Tissue Specific Metabolite Profiles that Include Striatal Lipid Accumulation. PLoS One 2015;10:e0134465. [Crossref] [PubMed]
46. Lokhande S, Koo S, Ray A. Supervised machine learning with protein structural and network topological features predicts physical interactors of the human huntingtons disease protein. In: 2018 International Conference on Computational Science and Computational Intelligence (CSCI), Las Vegas, NV, USA, 2018, pp. 1362-7.
47. Witmer A, Bhanu B. Multi-label Classification of Stem Cell Microscopy Images Using Deep Learning. In: 2018 24th International Conference on Pattern Recognition (ICPR), Beijing, China, 2018, pp. 1408-13.
48. Guo X, Jiang X, Xu J, et al. Ensemble Consensus-Guided Unsupervised Feature Selection to Identify Huntington’s Disease-Associated Genes. Genes (Basel) 2018; [Crossref] [PubMed]
49. Lovrecic L, Slavkov I, Dzeroski S, et al. ADP-ribosylation factor guanine nucleotide-exchange factor 2 (ARFGEF2): a new potential biomarker in Huntington’s disease. J Int Med Res 2010;38:1653-62. [Crossref] [PubMed]
50. Rocha NP, Mwangi B, Gutierrez Candano CA, et al. The Clinical Picture of Psychosis in Manifest Huntington’s Disease: A Comprehensive Analysis of the Enroll-HD Database. Front Neurol 2018;9:930. [Crossref] [PubMed]
51. Khan W, Alusi S, Tawfik H, et al. The relationship between non-motor features and weight-loss in the premanifest stage of Huntington’s disease. PLoS One 2021;16:e0253817. [Crossref] [PubMed]
52. Guimaraes MT, Medeiros AG, Almeida JS, et al. An Optimized Approach to Huntington’s Disease Detecting via Audio Signals Processing with Dimensionality Reduction. In: 2020 International Joint Conference on Neural Networks (IJCNN), Glasgow, UK, 2020, pp. 1-8.
53. Gallezot C, Riad R, Titeux H, et al. Emotion expression through spoken language in Huntington disease. Cortex 2022;155:150-61. [Crossref] [PubMed]
54. Hastie T, Tibshirani R, Friedman J. The Elements of Statistical Learning: Data Mining, Inference, and Prediction, Second Edition. Part of the book series: Springer Series in Statistics 2009;27.
55. Matloob F, Ghazal TM, Taleb N, et al. Software defect prediction using ensemble learning: A systematic literature review. IEEE Access 2021;9:98754-771.
56. Erdas ÇB, Sümer E. A Deep Learning-Based Approach to Detect Neurodegenerative Diseases. In: 2020 Medical Technologies Congress (TIPTEKNO), Antalya, Turkey, 2020, pp. 1-4.
57. Vipani R, Hore S, Basak S, et al. Gait signal classification tool utilizing Hilbert transform based feature extraction and logistic regression based classification. In: 2017 Third International Conference on Research in Computational Intelligence and Communication Networks (ICRCICN), Kolkata, India, 2017, pp. 57-61.
58. Klomsae A, Auephanwiriyakul S, Theera-Umpon N. String Grammar Unsupervised Possibilistic Fuzzy C-Medians for Gait Pattern Classification in Patients with Neurodegenerative Diseases. Comput Intell Neurosci 2018;2018:1869565. [Crossref] [PubMed]
59. Zhang H, Huo X, Guo X, et al. A disease-related gene mining method based on weakly supervised learning model. BMC Bioinformatics 2019;20:589. [Crossref] [PubMed]
60. Gelšvartas J, Simutis R, Maskeliūnas R. User Adaptive Text Predictor for Mentally Disabled Huntington’s Patients. Comput Intell Neurosci 2016;2016:3054258. [Crossref] [PubMed]
61. Riad R, Lunven M, Titeux H, et al. Predicting clinical scores in Huntington’s disease: a lightweight speech test. J Neurol 2022;269:5008-21. [Crossref] [PubMed]
62. Ye Q, Xia Y, Yao Z. Classification of Gait Patterns in Patients with Neurodegenerative Disease Using Adaptive Neuro-Fuzzy Inference System. Comput Math Methods Med 2018;2018:9831252. [Crossref] [PubMed]
63. Vajiha Begum SA, Pushpa Rani M. Diagnosis of Neurodegenerative Diseases by Gait Analysis using Triblock CNN and Deep RQA Techniques. In: 2022 International Conference on Automation, Computing and Renewable Systems (ICACRS), Pudukkottai, India, 2022, pp. 762-70.
64. Podgorelec V, Kokol P, Stiglic B, et al. Decision trees: an overview and their use in medicine. J Med Syst 2002;26:445-63. [Crossref] [PubMed]
65. Chaki J, Woźniak M. Deep learning for neurodegenerative disorder (2016 to 2022): A systematic review. Biomedical Signal Processing and Control 2023;80:104223. [Crossref]
66. Fan C, Chen M, Wang X, et al. A Review on Data Preprocessing Techniques Toward Efficient and Reliable Knowledge Discovery From Building Operational Data. Frontiers in Energy Research. 2021; [Crossref]
67. Gao S, Wang Y, Wang J, et al. Leveraging explainable deep learning methodologies to elucidate the biological underpinnings of Huntington’s disease using single-cell RNA sequencing data. BMC Genomics 2024;25:930. [Crossref] [PubMed]
68. Tzelepi M, Nousi P, Passalis N, et al. Representation learning and retrieval. In: Deep Learning for Robot Perception and Cognition. Elsevier; 2022. p. 221-41.
69. Yadav S, Shukla S. Analysis of k-Fold Cross-Validation over Hold-Out Validation on Colossal Datasets for Quality Classification. In: 2016 IEEE 6th International Conference on Advanced Computing (IACC). IEEE; 2016. p. 78-83.