Smartphone AI Detects Teen Depression via Speech with 95% Accuracy

How Smartphone Speech Analysis Can Improve Early Detection of Depression in Adolescents

Yes, a smartphone can detect teen depression with strong accuracy, as smartphone speech analysis offers an objective way to screen for major depressive disorder through structured voice tasks recorded on everyday phones. A large multicenter study tested the approach in 1,838 Chinese teens aged 13–18 and showed robust performance that may help clinicians in settings with limited mental health resources, and these findings align with broader trends in AI-powered mental health tools, including conversational agents and patient-centered screening systems.

Two models drove the results: one used a fine-tuned Whisper deep-learning system, while the other relied on handcrafted acoustic features, and both performed well across age, sex, and symptom-severity groups.

Key Performance Metrics

• Whisper-FT model: AUROC 0.95 in development, 0.88 in external validation.
• ComParE-OS-CB model: AUROC 0.84 in development, 0.82 in external validation.
• Fusion with PHQ-8 scores raised AUROC to 0.99 and 0.97 respectively.

SHAP analysis highlighted spectral shape, cepstral features, and jitter as the most useful signals, and decision-curve analysis confirmed added clinical value beyond questionnaires alone.

Ad

Study Design and Methods

Researchers enrolled 981 adolescents with major depressive disorder and 857 healthy controls, with diagnoses following MINI-KID interviews. Thirteen speech tasks were recorded at 16 kHz through a smartphone app, and models were validated with leave-one-site-out cross-validation and tested again in 145 real-world participants.

Practical Workflow Benefits

A two-stage process works best: first, the PHQ-8 screens everyone, and speech analysis then reviews borderline scores of 5–9. This step improves case-finding efficiency while keeping high accuracy, and reduced task sets still retained most of the full battery’s performance.

These findings mark the first large-scale test of speech-based major depressive disorder detection in adolescents. The tools support smartphone screening as a complement to existing methods, and they may reduce reliance on subjective reports and aid health economics research through earlier identification. Read the full study in npj Digital Medicine for detailed methodology and results.

Frequently Asked Questions

How does the two-stage workflow affect screening?

It raises efficiency by using the PHQ-8 first and speech analysis only for borderline cases.

Do the models perform equally across groups?

Yes. Both models showed stable results without major fairness gaps by age, sex, or symptom severity.

What limits wider use of these tools?

Limits include high anxiety comorbidity, variable recording settings, medication effects, Mandarin-only data, and the cross-sectional design.