Screening for depressive symptoms in primary and secondary school students based on speech features: A one-year...

This study aims to develop and validate an interpretable prediction model for recognizing depression risk in primary and secondary school students based on acoustic features of speech using multiple machine learning methods.

This study focused on the Jiangsu Province School-based Evaluation Advancing Response for Child Health (SEARCH) program. The baseline survey was conducted on 10,926 individuals in December 2022. Longitudinal follow-up was conducted with 4321 students who had no depressive symptoms at baseline and were available for the second and third follow-up surveys conducted in June 2023 and December 2023, respectively. Four machine learning methods were used, and the Shapley additive explanations (SHAP) method was used to rank and interpret the significance of acoustic speech features. Separate prediction models were constructed to assess the risk of depression in primary and secondary school students based on the number and types of acoustic speech features. The final prediction model was selected based on the optimal number of acoustic speech features. Additionally, participants were analyzed through subgroup stratification by academic level and gender, with the predictive effectiveness of the final model assessed across different groups.

Of the 4321 students without depressive symptoms at baseline, 452 (10.5 %) had developed symptoms by the second follow-up, and 512 (11.8 %) by the third. Of the four machine learning models tested, the Light Gradient Boosting Machine (LightGBM) achieved the best validation performance, with an AUC of 0.718 (95 % CI: 0.695-0.741), a PPV of 0.940, and an NPV of 0.196. Feature selection yielded an interpretable LightGBM model based on ten acoustic predictors. SHAP analysis indicated that the most influential were spectral features (mel spectrograms, spectral contrast, and MFCCs), with one additional prosodic feature. Validation of the final model indicated stable performance (AUC 0.699, 95 % CI: 0.675-0.722). Subgroup analysis further revealed higher predictive efficacy in primary school students than in middle and high school students.

Acoustic speech features can effectively identify depression risk in primary and secondary school students and may provide a novel approach for individualized assessment of depressive symptom risk in student populations.