PHO-Agents: A Large Language Model-Powered Multi-Agent System for Predicting Health Outcomes

Predicting health outcomes from electronic health records (EHRs) is challenging because traditional models rely on structured data and often ignore external medical knowledge. We propose an approach that integrates structured EHR with text‑based clinical evidence to improve prediction and interpretability.

We introduce PHO-Agents, a multi-agent system powered by large language models (LLMs) for health outcome prediction. Structured EHR sequences are encoded to produce attention-based representations and initial logits, which are converted into patient summaries by a data agent. A retrieval agent gathers relevant clinical guidelines. Research and practical doctor agents independently assess the patient, and a leader agent synthesizes their analyses. Outputs from the EHR-based model and the LLM agents are fused to generate final predictions and explanation reports. PHO-Agents was evaluated on three real-world cohorts: acute kidney injury (AKI) patients (in-hospital mortality), chronic kidney disease patients (AKI onset within two years), and cancer patients receiving immune checkpoint inhibitors (immune-related adverse events within one year).

PHO-Agents outperformed single-agent and multi-agent LLM baselines across all cohorts. In the AKI mortality task, it achieved a PR-AUC of 90.20 ± 2.07, compared with 56.46 ± 2.98 for the best single-agent baseline. Similar gains were observed in the ICI and CKD cohorts. Ablation studies showed that both multi-agent reasoning and logit-level fusion contributed to performance improvements, and case analyses demonstrated clinically consistent explanations.

PHO-Agents integrates longitudinal EHR modeling with collaborative LLM reasoning, improving predictive performance, interpretability, and robustness across diverse clinical tasks. This hybrid approach offers a trustworthy strategy for real-world clinical decision support.