Voice-Based Screening of Depression and Anxiety Using Machine Learning and Deep Learning: A Scoping Review of Methods and Clinical Readiness

Abstract

Despite the high global prevalence of depressive and anxiety disorders, access to early clinical assessment remains limited for many. Although this field has grown rapidly, existing reviews have focused primarily on technical performance, with limited systematic attention to whether current models meet the prerequisites for clinical implementation. This scoping review mapped methodological approaches across this domain and evaluated clinical readiness using five predefined indicators: sample size adequacy, external validation, prospective data collection, real-world evaluation, and model explainability. Searches of PubMed, Web of Science, and IEEE Xplore (March 2026) identified 2463 records; 34 studies (37 dataset evaluations) were included. The majority of studies (91%) were published from 2022 onwards. Depression was the primary target in 91% of studies, while only one study addressed anxiety. Hand-crafted acoustic features were the most frequent (57%), while classical machine learning was the most common model type (32%). External validation was conducted in only 32% of evaluations and real-world testing in 8%. Clinical readiness was classified as Low in 24%, Moderate in 65%, and High in 11% of evaluations. No evaluation met all five indicators simultaneously. These findings apply primarily to voice-based depression screening; 36 of 37 evaluations targeted depression, and the evidence base for anxiety disorders is limited to a single evaluation, precluding comparable characterisation for that condition. The principal challenges to clinical implementation are insufficient external validation, reliance on laboratory conditions, narrow linguistic coverage, and inconsistent metric reporting. The framework applied in this scoping review provides a replicable structure for assessing the clinical validity of AI-driven psychiatric screening tools.