Foundation Agents Meet Agentic Deep Research: Evidence-Grounded Clinical Code Forecasting
ICD-Deepresearch is a novel DeepResearch workflow that combines structured EHR foundation models (SparseEHR) with language foundation models (GPT-5) and medical search to predict future ICD diagnosis codes The system uses a two-path candidate generation strategy: an EHR Prior from SparseEHR feeds into bounded Research Expansion rounds, while GPT-5 provides independent direct forecasts Final Selection validates, deduplicates, and jointly ranks candidates from both paths, with a separate rationale
Analysis
TL;DR
- ICD-Deepresearch is a novel DeepResearch workflow that combines structured EHR foundation models (SparseEHR) with language foundation models (GPT-5) and medical search to predict future ICD diagnosis codes
- The system uses a two-path candidate generation strategy: an EHR Prior from SparseEHR feeds into bounded Research Expansion rounds, while GPT-5 provides independent direct forecasts
- Final Selection validates, deduplicates, and jointly ranks candidates from both paths, with a separate rationale-writing module that does not alter predictions
- Achieves patient-averaged precision/recall of 24.60/35.09% on MIMIC-III and 25.14/48.32% on MIMIC-IV, outperforming registered local comparators
- Physicians rated 51% and 68% of retrieved documents as useful, significantly higher than standalone GPT-5 web search (22%/39%) and Medical Deep Research (32%/41%)
Why It Matters
This work represents a significant step toward evidence-grounded clinical AI systems that combine the pattern-recognition strengths of structured EHR models with the hypothesis-generation capabilities of large language models. For AI practitioners working in healthcare, it demonstrates a practical architecture for multi-source candidate generation and validation that could generalize to other complex forecasting tasks where no single model modality is sufficient.
Technical Details
- Architecture: A two-path ensemble combining SparseEHR (structured EHR foundation model capturing recurrence and temporal progression) with GPT-5 (language model generating flexible diagnostic hypotheses), integrated through a DeepResearch workflow with medical search and ICD dictionaries
- Candidate Generation: SparseEHR produces an EHR Prior that initializes two bounded Research Expansion rounds; an independent GPT-5 Direct Forecast supplies complementary candidates, operating under a fixed top-K budget
- Selection & Rationale: Final Selection validates, deduplicates, and jointly ranks both candidate paths; a separate module writes rationales without modifying predictions, ensuring explanation fidelity
- Evaluation: Tested on MIMIC-III and MIMIC-IV datasets for next-encounter multi-label ICD code forecasting, with physician-rated document usefulness as a human-centered evaluation metric
- Key Innovation: The system addresses the fundamental challenge that no source reveals the future code set, evaluating candidates by linking patient evidence, external clinical relations, and exact code semantics
Industry Insight
- The hybrid architecture of combining structured EHR models with LLM-based reasoning and web search represents a scalable pattern for clinical AI systems, suggesting that future healthcare AI should prioritize multi-modal evidence integration over single-model approaches
- The physician usefulness ratings (51%/68%) significantly outperforming standalone search systems indicate that domain-specific grounding in EHR data is critical for clinical tool adoption, reinforcing the need for healthcare AI to be evaluated on clinician-centric metrics rather than purely algorithmic ones
- The separation of prediction from rationale generation is a notable design choice that could reduce hallucination risks in clinical settings, offering a template for building trustworthy AI systems where explanation fidelity must be decoupled from decision-making
Disclaimer: The above content is generated by AI and is for reference only.