Integrating Large Language Models and Machine Learning for Enhanced Catatonia Phenotyping: A Study on Clinical Data...

Catatonia, a complex syndrome with diagnostic challenges, was studied using a novel approach combining LightGBM and GPT-4 to enhance phenotyping from electronic health record (EHR) data. LightGBM, trained on structured data, achieved superior performance (AUROC 0.713) compared to GPT-4 (best AUROC 0.709 with retrieval-augmented generation (RAG) few-shot and ML predictions). While LightGBM excelled in raw metrics, GPT-4 added value through enhanced interpretability, effectively analyzing contextual information like the Bush-Francis Catatonia Rating Scale from clinical notes. This integration demonstrates the potential of combining ML and GPT-4 to improve catatonia identification and phenotyping accuracy.