Mortality prediction for ICU patients with mental disorders using large language models ensemble and unstructured medical notes

Abstract

Assessing mortality risk in the intensive care unit (ICU) is crucial for improving clinical outcomes and management strategies. Conventional artificial intelligence studies often neglect vital clinical information contained in unstructured medical notes. Recently, large language models (LLMs) have achieved leading-edge performance in natural language processing tasks, though each model has limitations stemming from its architecture and pre-training. The ensemble of heterogeneous language models, including both conventional LMs and LLMs, effectively addresses these constraints. The study introduces a predictive ensemble classifier using a decision fusion approach of diverse medical LLMs and LMs, including Asclepius, Meditron, GatorTron, and PubMedBERT. These models, fine-tuned with multimodal data from the medical records of 11,914 individuals diagnosed with various mental disorders from the MIMIC-IV dataset, enhance the diversity of the resulting ensemble model. The performance of our multimodal ensemble model was rigorously evaluated, delivering superior results compared to individual LLM and LM models based on single modalities. Our study underscores the substantial influence of language models on mental health management in the ICU, advocating for advanced clinical decision-making techniques that integrate unstructured medical texts with language models to enhance patient outcomes.