Reducing false negatives and false positives in predictive AI algorithms: a methodological refinement

In artificial intelligence (AI)-assisted medical screening, autonomous algorithms can efficiently and reliably estimate the probability of the occurrence of cancer through screening images. In particular, physicians strive to eliminate false negatives (FNs) as much as possible, which, inadvertently, leads to the misclassification of patients. This article proposes a technique to uncover latent FNs while demonstrating that this technique with the AI algorithm alone yields comparable clinical and superior economic results compared with radiologists reassessing the algorithm’s output. As illustrative examples, the article references Sybil, a lung cancer screening protocol developed at MIT. This study accomplishes two purposes. Firstly, the proposed method re-processes the negative results at a higher specificity level than in the initial assessment, that is, at a lower two-year risk cutoff, calibrated to cull high risk patients. Secondly, the analysis demonstrates the economic effectiveness of bypassing human review without endangerment of patient welfare. The recalibration of the classified negatives flushed out seven of the 44 FNs concealed among the true negatives (TNs). This reversion of FNs into true positives (TPs) secures additional patient welfare and boosts the effectiveness of the AI process. The second sorting minimizes the number of false positives (FPs) as well. The findings confirmed the viability of the second reading technique of the residual negatives from the initial reading. By increasing the specificity from 96.5% in the initial scan to 99.613% from a sample of 10,587 patients, the second scan converted seven FNs to TPs at a cost of $709,807 with 433 FPs. The experimental case with radiologists evaluating the negatives, though discovering seven TPs more than the algorithm alone, would have incurred $1,001,651 with 520 FPs. The second reading technique revealed seven misclassified FNs, exactly the same as the combined reading, with $291,844 in savings and 87 fewer FPs. A secondary processing of the negative elements computationally isolates suspicious borderline cases, reclassifying them as positives to successfully recover obscured FNs. This supplemental process avoids punitive costs of examining and treating excessive FPs, which can accumulate with unassisted physician intervention.