Researchers have developed a new artificial intelligence (AI) approach that improves the prediction of depression severity from clinical interview transcripts, even when data is limited and comes from different languages and clinical environments. This is important because accurately assessing how severe a patient’s depression is can guide treatment decisions, but collecting large, consistent datasets for training AI models is difficult, especially across diverse languages and healthcare systems. The new method leverages knowledge learned from one dataset in English to help analyze another in Chinese, demonstrating a promising way to build more adaptable mental health tools.
Key Takeaways
- The study introduces a sequential transfer learning technique that fine-tunes a large language model first on English depression interview data, then adapts it to Chinese clinical data with a different depression scale.
- This approach improves prediction accuracy on the smaller, Chinese dataset compared to training solely on that target data or using traditional methods.
- Using the original Chinese transcripts works better than relying on machine-translated English, highlighting the value of native language data.
- While promising, this is an early, exploratory study focused on internal evaluation and does not yet establish clinical diagnostic use.
The research tackles a common challenge in AI for mental health: data scarcity and variability. Depression severity can be measured using different clinical rating scales, such as PHQ-8 (Patient Health Questionnaire-8) and HAMD-17 (Hamilton Depression Rating Scale-17), which are not directly comparable. Additionally, datasets may come from different languages and clinical settings, making it hard for a single AI model to perform well across them. The team addressed this by using a large language model called Qwen3 as a backbone, which they first fine-tuned on an English dataset (DAIC-WOZ) containing avatar-mediated clinical interviews scored with PHQ-8. Then, they adapted this model to a smaller Chinese dataset (PDCH), consisting of real clinical consultations scored with HAMD-17.
To achieve this, the researchers employed a technique known as sequential low-rank adaptation (LoRA). LoRA is a method that allows AI models to be fine-tuned efficiently by updating only a small number of parameters in a structured way, rather than retraining the entire model. This makes it feasible to transfer learning from one dataset to another without requiring huge computational resources or large amounts of new data. The model’s “regression head”—the component responsible for predicting the depression severity score—was designed to be bounded, ensuring predictions stayed within realistic clinical ranges. For the target Chinese dataset, this head was reinitialized to match the specific scoring scale used there.
The team evaluated their approach using rigorous cross-validation, a method that tests the model’s performance on different splits of the data to ensure reliability. They compared their sequential transfer learning method to models trained only on the target Chinese data and to other baseline methods that do not use large language models. Results showed that their approach achieved lower mean absolute error (MAE) and root mean squared error (RMSE), meaning it made more accurate predictions. Additionally, it attained better macro-F1 scores, indicating improved balance between precision and recall when classifying depression severity levels.
Importantly, the study found that using the original Chinese transcripts led to better results than translating them into English before analysis, emphasizing the importance of native language data in mental health AI. They also tested variations like reversing the order of training or using unsupervised learning exposure, which did not yield clear benefits. The authors caution that this is an exploratory, single-site study and does not yet demonstrate clinical screening or diagnostic utility. It also does not isolate the effects of language, rating scale, or clinical setting changes individually.
Looking ahead, this research suggests that sequential transfer learning with efficient adaptation methods like LoRA could help AI models generalize better across diverse clinical populations and languages, addressing a major hurdle in mental health technology. Future work will need to validate these findings on larger, multi-site datasets and explore how to integrate such models into real-world clinical workflows. By bridging gaps between different depression assessment tools and languages, this approach moves us closer to AI systems that can support clinicians worldwide in understanding and treating depression more effectively.
Based on research published on arXiv by Wenjie Feng, Sahba Zojaji, Satoshi Nakamura.