Beyond explainable AI: Enhancing trust and robustness in machine learning for sleep apnea diagnosis.

Kilic et al. reviewed machine learning (ML) and deep learning (DL) for sleep apnea detection, emphasizing explainable AI (XAI) while noting challenges like Apnea-Hypopnea Index (AHI) discrepancies. This paper extends their critique, arguing that XAI tools like SHAP inherit model biases, and high prediction accuracy does not guarantee reliable feature importances, which inherently lack ground truth validation. To overcome these limitations and build clinical trust, we advocate for a comprehensive approach combining unsupervised ML (e.g., feature agglomeration, highly variable gene selection) with nonlinear nonparametric statistical methods (e.g., Spearman's correlation). This strategy robustly evaluates variable relationships and p-values, particularly for monotonic associations, mitigating misapplications stemming from assumption violations and inadequate interpretation of model ground truth, thus fostering real-world applicability.