Digital phenotyping with large language models to detect depressive state changes in patients.

Area of Science:

• Digital health
• Artificial intelligence in mental health
• Computational psychiatry

Background:

• Digital phenotyping uses smartphone and wearable data for continuous mental health monitoring.
• Early detection and personalized care for mood disorders are crucial for effective treatment.
• Data heterogeneity in digital phenotyping poses significant modeling challenges.

Purpose of the Study:

• To evaluate the effectiveness of large language models (LLMs) in detecting depression severity changes using digital phenotyping data.
• To compare different LLM strategies, including in-context learning and fine-tuning.
• To assess the performance of LLMs against traditional baseline models.

Main Methods:

• Utilized digital phenotyping data from individuals with major depressive episodes.
• Implemented and compared few-shot prompting, in-context learning, and fine-tuning LLM strategies.
• Evaluated embedding-only and QLoRA fine-tuning techniques.

Main Results:

• Both few-shot prompted and fine-tuned LLMs significantly outperformed traditional baseline models.
• Embedding-only fine-tuning excelled with individual digital phenotyping features.
• QLoRA fine-tuning demonstrated superior performance when integrating combined input features.

Conclusions:

• LLMs show significant potential for integrating heterogeneous behavioral data from digital phenotyping for mental health analysis.
• LLM approaches offer a promising avenue for real-time depression severity monitoring.
• Clinical validation and robust ethical frameworks are essential for deploying these technologies in practice.