Abstract
BACKGROUND: Depression diagnosis faces challenges of subjectivity and delay. Speech features offer potential objective biomarkers, but a systematic comparison of traditional machine learning (TML) and deep learning (DL) models is lacking. OBJECTIVE: To evaluate and compare the diagnostic accuracy of TML and DL models for depression detection using speech features, and to examine subgroup effects across sample size, validation strategy, language, and diagnostic criteria. METHODS: Following PRISMA guidelines, we systematically searched 9 databases (PubMed, Medline, Embase, PsycINFO, Scopus, IEEE, Cochrane, ACM Digital Library, and Web of Science) from inception to April 2025. Eligible studies included clinically diagnosed patients with depression and healthy controls, assessed using speech-based TML or DL models, and reporting sensitivity, specificity, or the area under the curve (AUC). Risk of bias was evaluated using the diagnostic Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2). Random-effects bivariate models pooled diagnostic performance, and heterogeneity, subgroup, and sensitivity analyses were conducted. RESULTS: Twenty-five studies met the inclusion criteria (9 TML, 16 DL). TML models showed pooled sensitivity of 0.82 (95% CI: 0.74-0.88), specificity 0.83 (95% CI: 0.75-0.90), and AUC 0.89 (95% CI: 0.86-0.92). DL models achieved pooled sensitivity of 0.83 (95% CI: 0.77-0.88), specificity 0.86 (95% CI: 0.80-0.90), and AUC 0.91 (95% CI: 0.89-0.93). Subgroup analyses indicated that diagnostic performance varied by sample size, validation strategy, language, and diagnostic criteria. CONCLUSION: Both TML and DL models demonstrate good diagnostic accuracy in speech-based depression detection. The marginal but consistent superiority of DL models supports their potential use in secondary care settings for confirmatory diagnosis, while TML remains valuable for primary care screening.. CLINICAL TRIAL NUMBER: Not applicable.