Relapse prediction after remission in major depressive disorder: From single classification to sequential clinical decisions

Relapse after remission is a major obstacle in managing major depressive disorder, and scheduled follow-up may be too infrequent to detect early deterioration.Artificial intelligence has been applied to relapse prediction across neuroimaging, electronic health records, wearable sensing, and speech and language data, yet performance has remained modest, endpoints heterogeneous, and clinical integration weak.We argue that the bottleneck is less model quality or the choice of data modality than how the prediction task has been specified.Relapse has usually been treated as a single distal classification, whereas prevention requires a sequence of clinical decisions: Who needs closer surveillance, when behaviour drifts from a personal baseline, and whether that drift warrants action.On current evidence, these modalities are partial answers to different questions rather than competing solutions to one.We organise this into the Stratify-Monitor-Confirm framework: Baseline data stratify vulnerability and set monitoring intensity; passive sensing, brief repeated self-report, or both detect sustained withinperson deviation rather than relapse itself; and triggered assessment confirms whether the signal reflects clinically meaningful worsening.We present Stratify-Monitor-Confirm as a proposed, evidence-informed conceptual architecture for shifting artificial intelligence from isolated relapse prediction towards clinically actionable post-remission monitoring: Untested as a sequence and weakest at Confirm, but testable as specified.