Explainable AI for Mental Health Emergency Returns: Integrating LLMs with Predictive Modeling

LLMs can make AI predictions more understandable to doctors: The study found that integrating large language models with traditional AI achieved 99% accuracy in translating model predictions into clinically relevant explanations that doctors can actually use in practice, addressing a major barrier to AI adoption in healthcare.

AI shows promise for identifying high-risk mental health patients: The integrated LLM approach improved prediction accuracy for 30-day ED returns from an AUC of 0.74 to 0.76, which could help clinicians identify the roughly 24-27% of mental health patients who return to the emergency department within a month and provide them with better preventive care.

Social factors matter, but AI struggles with some of them equally: While the AI successfully identified social determinants of health (like substance abuse) with 96% accuracy, it performed notably worse at detecting exercise habits and home environment issues (70-67% accuracy), suggesting AI tools need human oversight on certain social factors that affect mental health outcomes.

Smaller, specialized AI models can work as well as larger ones: The study used LLaMA 3 with only 8 billion parameters rather than massive models, achieving 88% accuracy in understanding patient complaints, indicating that responsible AI implementation doesn't always require the largest or most resource-intensive systems.

This approach was tested on real patient data at scale: The findings are based on analysis of over 42,000 actual emergency department visits from nearly 28,000 unique patients, making this a practical demonstration rather than a theoretical exercise.