Explaining the Unseen: AI-Driven Insights from Sleep Motion for Explainable Alzheimer Disease’s Screening

Early detection of Alzheimer’s disease (AD) is critical for effective intervention, yet current diagnostic methods are often invasive, costly, or inaccessible. Emerging research identifies sleep psychomotor disturbances as potential digital biomarkers for screening. However, deploying artificial intelligence (AI) in high-risk healthcare settings requires more than predictive performance; it demands transparency to ensure clinical trust and accountability. This research monitored the nocturnal psychomotor activity of 219 female participants (118 diagnosed with AD and 101 healthy controls) using wrist-worn inertial sensors. To bridge the gap between algorithmic accuracy and clinical interpretability, we employed X-ROCKET (Explainable Random Convolutional Kernel Transform). Unlike traditional black-box deep learning models, our proposed approach allows for the extraction of understandable features that validate the model’s decision-making process. Our results achieved a sensitivity (recall) of 88.16% and an accuracy of 82.42%, performance levels comparable to several invasive clinical benchmarks. The explainability analysis revealed that the model correctly prioritizes low-frequency movements, physiologically consistent with disruptions in sleep architecture (e.g., REM atonia), while identifying wrist pronation/supination as non-discriminative. These findings suggest that integrating explainable AI into psychomotor analysis can provide a new approach for developing robust, non-invasive, and trustworthy tools for large-scale cognitive impairment screening.

Direct access at: https://doi.org/10.1007/s10796-026-10736-0