Comparative Evaluation of Pretrained Large Language Models for Suicide Risk Prediction from Clinical Notes in U.S. Veterans

A study compared the ability of artificial intelligence to predict suicide risk in 27,241 U.S. veterans by analyzing their clinical notes. Researchers tested modern language models (LLMs) and compared them with traditional text processing methods. Results showed that modern language models were more effective in seven of nine tested scenarios. When using text alone, they achieved maximum accuracy of 0.644, but when combined with structured clinical data, accuracy increased to 0.748. The models identified specific language related to suicide, particularly in notes from the last 30 days among high-risk patients. The study suggests that modern language models can improve suicide risk prediction by analyzing clinical narratives.