ICD-Deepresearch: A Groundbreaking AI Workflow for Future Clinical Diagnosis Coding
A paper available on arXiv (identifier 2608.17075v1) presents ICD-Deepresearch, a workflow designed for deep research that predicts future International Classification of Diseases (ICD) codes utilizing electronic health records (EHRs). The study, named 'Foundation Agents Meet Agentic Deep Research: Evidence-Grounded Clinical Code Forecasting,' integrates structured EHR foundation models with language foundation models to forecast diagnosis codes for upcoming clinical visits based on past data. It employs SparseEHR for the initial EHR prior and leverages GPT-5 to generate a variety of candidate codes. This methodology seeks to improve the accuracy and interpretability of clinical coding, potentially benefiting decision support and healthcare management. Specific performance metrics and validation information are not included.
Key facts
- Paper is available on arXiv with identifier 2608.17075v1 and announcement type 'cross'.
- Introduces ICD-Deepresearch, a DeepResearch workflow for next-encounter ICD code forecasting.
- Task is prospective and multi-label, predicting multiple future diagnosis codes from historical EHR data.
- Combines structured EHR foundation models (SparseEHR) with language foundation models (GPT-5).
- SparseEHR generates an EHR prior that initializes two bounded research expansion rounds.
- GPT-5 supplies an independent direct forecast with complementary candidates.
- System evaluates candidate transitions using patient evidence, external clinical relations, and code semantics under a fixed top-K budget.
- No performance metrics or validation results are reported in the abstract.
Entities
—