Probabilistic Chain-of-Thought Agent for Clinical EHR Reasoning

A new AI framework, COTCAgent (Probabilistic Chain-of-Thought Completion Agent), addresses critical flaws in large language models' reasoning over longitudinal electronic health records (EHR). Current LLMs hallucinate clinical trends due to lack of fine-grained statistical reasoning and struggle with non-uniform time series and scarce labels. COTCAgent is a hierarchical reasoning framework with three core modules: Temporal-Statistics Adapter (TSA) for quantitative evidence extraction, Probabilistic Chain-of-Thought (P-CoT) for uncertainty-aware reasoning, and a Completion module for handling missing data. The system aims to improve preventive consultation and clinical decision support by enabling reliable temporal reasoning from EHR data. The paper is available on arXiv (2605.15016).