Screening for depression risk via smartphone narratives with fully fine-tuned WavLM

Scalable, low-burden tools are needed to identify individuals at risk of depression before progressing to a clinically significant level of depressive disorder. We evaluated a smartphone framework for subclinical-depression risk using self-supervised speech representations and a field-ready data-collection protocol. Participants (N = 119) were stratified into high-risk (PHQ-9 ≥ 10; n = 64) and low-risk (n = 55) groups. A mobile app elicited two 1-minute narrative recordings using negative and positive mood-induction tasks, separated by a 10-minute neutral interval. We compared four models: extreme gradient boosting on handcrafted features; a convolutional neural network–recurrent neural network (CNN–RNN) on mel-spectrograms; head-only WavLM-MLP; and fully fine-tuned WavLM (WavLM-FT). Performance was estimated using 5-fold cross-validation (CV) and out-of-fold (OOF) aggregation. In a 5-fold CV, WavLM-FT achieved the highest area under the receiver operating characteristic curve (ROC-AUC) of 0.90, area under the precision–recall curve of 0.90, F1 of 0.73, accuracy of 0.68, recall of 0.89, and precision of 0.65. In OOF subject-level predictions, WavLM-FT led (ROC-AUC 0.86; accuracy 0.79) and outperformed CNN–RNN and WavLM-MLP. Results suggest full-model adaptation captures informative paralinguistic cues within standardized smartphone recordings. A brief, ecologically valid protocol with self-supervised learning may enable scalable, non-invasive depression-risk screening.