[Paper Review] Emulating Clinician Cognition via Self-Evolving Deep Clinical Research

Ruiyang Ren, Yuhao Wang+9 more|arXiv (Cornell University)|Mar 11, 2026

Artificial Intelligence in Healthcare and Education0 citations

TL;DR

DxEvolve is a self-evolving diagnostic agent that uses a deep clinical research workflow to convert encounters into reusable diagnostic cognition primitives, achieving clinician-level accuracy and cross-institution portability without parametric retraining.

ABSTRACT

Clinical diagnosis is a complex cognitive process, grounded in dynamic cue acquisition and continuous expertise accumulation. Yet most current artificial intelligence (AI) systems are misaligned with this reality, treating diagnosis as single-pass retrospective prediction while lacking auditable mechanisms for governed improvement. We developed DxEvolve, a self-evolving diagnostic agent that bridges these gaps through an interactive deep clinical research workflow. The framework autonomously requisitions examinations and continually externalizes clinical experience from increasing encounter exposure as diagnostic cognition primitives. On the MIMIC-CDM benchmark, DxEvolve improved diagnostic accuracy by 11.2% on average over backbone models and reached 90.4% on a reader-study subset, comparable to the clinician reference (88.8%). DxEvolve improved accuracy on an independent external cohort by 10.2% (categories covered by the source cohort) and 17.1% (uncovered categories) compared to the competitive method. By transforming experience into a governable learning asset, DxEvolve supports an accountable pathway for the continual evolution of clinical AI.

Motivation & Objective

• Motivate the gap between static AI diagnosis and dynamic clinical reasoning with auditable governance.
• Propose a workflow-aligned diagnostic agent that evolves from clinical encounters through exposed experience artifacts.
• Enable cross-institution and cross-language portability of learned diagnostic heuristics via DCPs.
• Demonstrate improved diagnostic accuracy and clinician-level performance on public benchmarks and external cohorts.
• Provide an auditable, governance-friendly path for continual improvement of clinical AI.

Proposed method

• Introduce DxEvolve, an interactive deep clinical research (DCR) workflow that guides evidence acquisition and hypothesis refinement.
• Lightweightly integrate exploration of examinations, lab tests, imaging, and external sources (guidelines, PubMed) within an evidence-centered reasoning loop.
• Distill encounter trajectories into diagnostic cognition primitives (DCPs), a reusable, indexable repository of clinical experiences.
• Evaluate DxEvolve on MIMIC-CDM with fixed accrual pools, comparing against a CDM baseline and a non-DCP ablation.
• Validate portability with external validation on a Chinese PLA General Hospital cohort using translated and native records.

Experimental results

Research questions

• RQ1 Can a workflow-aligned diagnostic agent achieve clinician-level accuracy under interactive evidence acquisition constraints?
• RQ2 Do diagnostic cognition primitives enable auditable, governable, and transferable learning across institutions and languages?
• RQ3 How does self-evolution affect diagnostic performance as encounter exposure increases and as errors drive learning?
• RQ4 Is external retrieval of guidelines and PubMed data essential, or can the core workflow scaffolding and DCPs suffice for gains?
• RQ5 What is the cross-institutional and cross-language generalizability of DxEvolve’s experiential learning?

Key findings

• DxEvolve yields an 11.2% mean accuracy gain over the CDM baseline across base LLM backbones on MIMIC-CDM.
• DxEvolve reaches 90.4% accuracy on a reader-study subset, surpassing the human expert benchmark of 88.8%.
• External validation shows a 10.2% gain over the CDM baseline on the external Chinese cohort, with 17.1% gains in categories absent from the initial repository.
• DCP-guided self-evolution provides portable, domain-agnostic heuristics that transfer across translations and institutions (including Chinese documentation).
• Later-stage DCPs receive higher clinician-rated scores and are retrieved more often in error-correcting episodes, indicating maturation of experience artifacts.
• DxEvolve improves workflow-consistency and guideline-adherence in evidence acquisition compared to the baseline.