Trustworthy AI for Neurodegenerative Disease Screening: Explainability, Clinical Accountability, and Human–AI Collaboration
Main article
Abstract
Neurodegenerative diseases such as Parkinson's disease and Alzheimer's disease are placing a growing burden on ageing societies, while traditional clinical screening still depends heavily on subjective expert evaluation and long diagnostic pathways. Recent deep-learning models on resting-state functional MRI and other neuroimaging data have improved discrimination accuracy, but their adoption in clinical practice remains limited because they offer little insight into how a decision is reached, who is accountable when it is wrong, and how it should be combined with the judgement of an experienced clinician. This article proposes a sociotechnical framework for trustworthy AI in neurodegenerative disease screening that integrates three pillars — explainability, clinical accountability, and human–AI collaboration — and connects them into a single deployable workflow. We characterise faithful explanation methods for graph-based brain-network models, including saliency, concept-based attribution, subgraph rationale, and uncertainty quantification, and discuss the role of regulatory frameworks (FDA SaMD, EU AI Act, NMPA) in defining audit trails, post-market surveillance, and liability allocation. We then formalise a collaborative diagnostic workflow in which clinicians retain final authority while AI provides risk scores, calibrated confidence, and structured rationale, and we describe a closed feedback loop that links clinical override events to bias monitoring and model retraining. An empirical evaluation across four hospital cohorts (640 participants) shows that the framework preserves screening performance (pooled AUROC 0.881) while improving clinician trust calibration and reducing decision time. The findings suggest that trustworthy AI for neurodegenerative screening is achievable only when explainability, accountability, and collaboration are designed jointly rather than treated as independent technical add-ons.
