Automatic depression severity assessment with deep learning using parameter-efficient tuning.

Area of Science:

• Artificial Intelligence
• Mental Health Technology
• Computational Linguistics

Background:

• Accurate depression assessment is crucial for mental health care providers.
• Deep learning models show promise for automatic depression severity assessment from clinical interview transcriptions.
• A significant bottleneck in applying deep learning to mental health is the scarcity of large, high-quality datasets.

Purpose of the Study:

• To develop a novel, data-efficient deep learning approach for automatic depression severity assessment.
• To address the challenge of limited datasets in mental health applications using parameter-efficient tuning.
• To evaluate the efficacy of prefix vectors in adapting large language models for depression prediction.

Main Methods:

• A novel approach leveraging pretrained large language models (LLMs) and parameter-efficient tuning (prefix vectors) was proposed.
• The method adapted a small set of tunable parameters (prefix vectors) to guide LLMs in predicting Patient Health Questionnaire (PHQ)-8 scores.
• Experiments were conducted on the Distress Analysis Interview Corpus - Wizard of Oz (DAIC-WOZ) dataset, evaluating performance on training, development, and test sets.

Main Results:

• The proposed prefix vector model achieved state-of-the-art performance on the DAIC-WOZ test set, outperforming previous methods with a root mean square error (RMSE) of 4.67 and mean absolute error (MAE) of 3.80.
• Prefix-enhanced models demonstrated reduced overfitting compared to conventionally fine-tuned models, utilizing significantly fewer training parameters (<6% relatively).
• The model's performance was robust, with consistent results across multiple runs on the development and test sets.

Conclusions:

• Prefix-tuning offers an effective strategy for adapting pretrained LLMs to specific downstream tasks like depression assessment, even with limited data.
• The flexibility of prefix vector size allows for fine-grained control over model learning capacity, enhancing performance.
• This research provides strong evidence for the utility of prefix-tuning in developing advanced tools for automatic depression assessment.