Machine Learning for Chronic Disease Classification and Comorbidity Detection: Methodological Gaps and Future Directions

Abstract

The rapid proliferation of machine learning (ML) methods across clinical medicine has generated a rich but fragmented body of evidence for chronic disease classification. Despite consistently high reported accuracy, the literature is characterised by five systematic methodological limitations: exclusive reliance on binary single-disease classification tasks, absence of leakage-free preprocessing protocols, lack of non-parametric statistical validation, omission of probability calibration evaluation, and minimal integration of explainability frameworks. This narrative review critically examines 25 representative ML studies spanning musculoskeletal disorders (particularly disc herniation), inflammatory bowel conditions, fibromyalgia, cardiovascular disease, and related chronic comorbidities, published between 2012 and 2025. Studies are analysed thematically across algorithmic approach, task scope, class imbalance strategy, and methodological rigour. Algorithmic families represented include classical support vector machines and tree ensembles, deep learning architectures (CNN, LSTM, U-Net), optimization-enhanced methods (WOA, GGO, PSO, SO), and natural language processing models (RoBERTa). Across all 25 studies, performance metrics range from 82.47% to 99.9% accuracy, yet none simultaneously addresses multiclass comorbidity discrimination, leakage-free preprocessing, and model explainability. The review identifies five critical gaps and maps them to concrete future research directions, with particular emphasis on the unmet need for a unified multiclass framework capable of differential diagnosis among clinically overlapping chronic conditions within fibromyalgia populations. These findings suggest that current ML models are not yet clinically ready for the differential diagnosis of comorbid chronic conditions without methodological reform.