MIT researchers' ICD-Deepresearch predicts future diagnoses 2.2x more accurately

A team from the University of Massachusetts and MIT introduced **ICD-Deepresearch**, a novel AI system that predicts future diagnosis codes (ICD codes) for patients by analyzing their longitudinal electronic health records (EHRs). Unlike traditional models that focus only on structured data, this approach combines **SparseEHR foundation models**—designed to capture temporal progression in patient records—with **GPT-5** to generate flexible diagnostic hypotheses. The system employs a two-stage candidate generation process: SparseEHR produces an EHR Prior, followed by two bounded Research Expansion rounds, while GPT-5 contributes complementary candidates through direct forecasting. A Final Selection module then validates, deduplicates, and ranks these candidates using clinical evidence and ICD semantics.

The method was evaluated on the MIMIC-III and MIMIC-IV datasets, achieving **24.60% precision and 35.09% recall on MIMIC-III**, and **25.14% precision and 48.32% recall on MIMIC-IV**—demonstrating a marked improvement over baseline models. In clinical validation, physicians found **68% of ICD-Deepresearch’s retrieved documents useful**, compared to just **39% for GPT-5’s web search** and **41% for Medical Deep Research**, highlighting its ability to deliver more clinically relevant evidence. The research underscores the potential of foundation agents and agentic deep research workflows in improving predictive accuracy in high-stakes healthcare applications.