Multimodal fusion of speech and PHQ-9 for machine learning-based adolescent depression screening

Traditional screening for adolescent depression primarily relies on self-reported symptoms. However, this method is vulnerable to under-reporting and stigma, which can delay identification and referral. Integrating objective speech biomarkers with machine learning into screening workflows offers a potential adjunctive approach to adolescent depression screening.

A total of 421 adolescents participated in this single-site cross-sectional study, completing speech tasks and the Patient Health Questionnaire-9 (PHQ-9). All participants underwent semi-structured interviews based on the 17-item Hamilton Depression Rating Scale (HAMD-17) and were categorized into control (≤ 7, n = 364) and depression (> 7, n = 57) groups. Following acoustic feature extraction, we performed feature selection via recursive feature elimination with cross-validation. The selected acoustic features and PHQ-9 item scores were concatenated through feature-level multimodal fusion to train and test machine learning models. Model interpretation was conducted using SHAP.

Compared to the standard threshold-based PHQ-9 approach, the 30-dimensional multimodal feature set demonstrated superior performance across four machine learning models. Notably, the balanced random forest (BRF) model achieved a balanced accuracy of 0.875 and an AUROC of 0.957 on the training set, with corresponding values of 0.835 and 0.896 on the hold-out test set.

The main limitations of this study include the absence of longitudinal follow-up and a single-site recruitment design.

Machine learning-based multimodal fusion of acoustic features and self-reports enhances adolescent depression screening performance and supports the technical potential for scalable adjunctive screening of current depressive symptoms.