Summary
Predicting how severe someone’s depression is can help doctors provide better care, but there isn’t always enough data for training computer models. The authors tried teaching a model first on English data with one type of depression scale, then adjusted it to work on Chinese data with a different scale. This step-by-step approach worked better than training only on the smaller Chinese dataset or using other methods. They also found that using native language transcripts helped more than machine-translated text. However, this study is an early test and not meant for clinical decision making yet.
What this means in practice
- •For clinical data scientists: Improve depression severity prediction models by transferring knowledge from larger English datasets to scarce Chinese clinical data using sequential fine-tuning.
- •For multilingual medical ai developers: Build better AI tools that adapt depression assessment models across different languages and clinical scales with limited target data.
Tested on one dataset.
Abstract
This work addresses continuous depression-severity score prediction from clinical interview transcripts under data scarcity. We propose a sequential low-rank adaptation (LoRA) protocol for cross-scale transfer: a Qwen3 backbone with a bounded regression head is first fine-tuned on the English DAIC-WOZ dataset (189 avatar-mediated sessions, PHQ-8), and the adapter then initializes fine-tuning on the Chinese PDCH dataset (100 real clinical consultations, HAMD-17), where a reinitialised, scale-specific head predicts the clinician-assigned score. All configurations use patient-level stratified 5-fold, 2-repeat cross-validation. On the data-scarce HAMD-17 target, the sequential protocol attains the best point-estimate MAE , RMSE, and macro-$F_1$ on both 0.6B and 1.7B backbones, outperforming target-only training and non-LLM baselines---4.96/6.59/0.36 with Qwen3-0.6B and 4.38/5.62/0.46 with Qwen3-1.7B. Ablations suggest that correctly aligned source supervision gives the best point estimates (unsupervised exposure and shuffled-label controls also show partial gains), that native-Chinese target input outperforms machine-translated English input, and that the reversed order yields no clear gain within run-to-run variance. The study is an exploratory, single-site internal evaluation: it does not establish screening or diagnostic utility, nor separately identify the contribution of the scale, language, or paradigm shifts. To our knowledge, no prior study evaluates this specific DAIC-WOZ-to-PDCH sequential transfer setting.