LLM-based split learning predicts mental distress across heterogeneous surveys

A new arXiv paper proposes a schema-aware split learning approach that uses LLMs to predict mental distress from survey data while keeping sensitive records private. The method is designed to work across surveys with differing structures and questions, which is a common obstacle when pooling mental health data from schools, employers, and clinics. The work targets privacy-preserving collaboration, so data stays local rather than being centralized.