Comparative study of an ai-based visual psychophysiological analysis platform and self-report scales for screening depression and anxiety: a single-center prospective diagnostic study.

Background: Depression and anxiety are among the most prevalent psychiatric disorders in clinical practice. Their high comorbidity and the inherent subjectivity of self-report screening tools have motivated efforts to identify objective, physiology-based digital phenotypes. Objectives: To rigorously evaluate the diagnostic performance of an artificial intelligence visual analysis platform based on head-neck micro-vibration signals for screening depression and anxiety, to compare its differences and complementarities with traditional self-report scales, and to develop and explore the potential utility of a combined "AI broad screening + scale refinement" approach. Methods: We conducted a single-center prospective diagnostic study enrolling 98 outpatients. A psychiatrist-administered structured interview grounded in DSM-5 served as the clinical diagnosis. All participants completed Self-Rating Depression Scale (SDS) and Self-Rating Anxiety Scale (SAS) assessments in parallel with testing by the AI psychophysiological analysis system. We constructed confusion matrices, calculated F1 scores, and generated receiver operating characteristic curves and decision curve analyses to quantify and compare the screening and stratification performance of each tool and of the combined models. Results: For depression-risk screening, the AI tool demonstrated very high sensitivity (95.9%), exceeding that of the SDS (83.6%). The combined "AI + SDS" model further increased sensitivity to 98.6%, demonstrating a minimized false-negative rate in this cohort. For anxiety, integrating AI with the SAS increased recall by 50.0% (to 69.2%) and improved the F1 score by 25.4%. In-depth analyses revealed that the AI system was particularly effective at identifying "silent patients" with alexithymia or prominent somatization, whereas the scales aligned more closely with clinical judgment for fine-grained severity grading. ROC and decision curve analyses consistently showed that the combined "AI + SDS/SAS" model achieved the best overall discrimination and greatest net clinical benefit. Conclusions: This study demonstrates that an AI tool based on head-neck micro-vibration signals can serve as a high-sensitivity, objective sentinel, mitigating the risk of missed cases associated with subjective self-report scales in specific populations. AI and self-report measures capture complementary facets of psychopathology. A tiered workflow of "AI broad screening + scale refinement" may constitutes a translationally promising paradigm to facilitate earlier, more objective, and efficient screening and to support more precise interventions in psychiatric disorders.