Healthcare Agents: Large Language Models in Health Prediction and Decision-Making

Abstract

Large Language Models (LLMs) are transforming healthcare, yet utilizing them for clinical applications presents significant challenges. In this thesis, we explore two critical aspects in healthcare AI: (1) leveraging LLMs for multimodal health prediction from wearable sensor data and (2) developing collaborative AI framework for medical decision-making. We first introduce a Health-LLM framework that performs multimodal fusion of temporal physiological signals from wearable devices with contextual metadata to predict health outcomes. By implementing novel context enhancement strategies, our framework demonstrates significant improvements in prediction accuracy across multiple health domains compared to existing benchmarks. Furthermore, we present MDAgents, an adaptive framework that optimizes multi-agent LLM collaboration for complex medical reasoning tasks. MDAgents dynamically configures agent roles and interaction patterns based on task complexity, implementing a hierarchical consensus mechanism that emulates clinical team dynamics. Through comprehensive evaluation on medical diagnosis and reasoning tasks, MDAgents exhibits superior performance in multimodal medical reasoning compared to single-agent approaches. Our findings demonstrate that LLMs, when architected for multimodal integration and strategic collaboration, can serve as robust agents in healthcare systems, advancing both preventive medicine through continuous health monitoring and clinical decision support through distributed AI reasoning.