Diagnostic Validity, External Validation, and Clinical Utility of Multimodal Artificial Intelligence for Depression and Anxiety Detection in Adults: A Critical Review

Abstract

Multimodal artificial intelligence (AI) combines behavioural, linguistic, physiological, questionnaire and contextual signals to identify depression or anxiety. Although reported discrimination is frequently promising, diagnostic accuracy, transportability and clinical benefit represent different evidential questions. This critical narrative review examines adult-focused multimodal systems, the validity of their reference standards, the strength of internal and external evaluation, and the evidence required before clinical deployment. Literature published from January 2014 to 23 August 2026 was examined alongside earlier methodological and measurement studies. Sources were selected for relevance to diagnostic constructs, study design, model assessment, external validation and clinical implementation. The synthesis distinguishes clinician-assessed disorders from questionnaire thresholds, symptom trajectories, induced anxiety states and benchmark-derived labels. Recent multimodal depression studies include a geographically independent validation with strong case–control discrimination, whereas a naturalistic speech study using structured diagnostic interviews showed substantially more modest performance and little incremental benefit beyond conventional rating scales. Syntheses of model performance remain vulnerable to heterogeneous reference standards, retrospective study designs, selective cohorts and dependence on repeatedly reused datasets. Evidence for anxiety is particularly fragmented: physiological and smartphone models often predict momentary arousal or elevated questionnaire scores rather than independently diagnosed anxiety disorders. Apparent gains from fusion must therefore be evaluated against simple clinical comparators, missing-modality patterns, calibration, subgroup reliability, data burden and downstream consequences. A staged translational framework is proposed, separating analytical validity, diagnostic validity, external reproducibility and patient-centred utility. Priority studies should prospectively recruit consecutive or representative clinical populations, employ blinded structured assessments, prespecify thresholds, report calibration and decision consequences, and compare multimodal models with feasible care pathways. Multimodal AI has credible research value as an adjunct to assessment and symptom monitoring; evidence for autonomous diagnosis or improved adult clinical outcomes remains insufficient.