Emulating Clinician Cognition via Self-Evolving Deep Clinical Research

1 Introduction

• DxEvolve addresses process and developmental gaps in clinical AI: static full-information prediction and a lack of mechanisms to distill longitudinal practice into inspectable experience.These gaps limit clinical and procedural governance of accumulated behavior.Source: §1 · p. 2
• Its deep clinical research (DCR) workflow frames diagnosis as active inquiry, while diagnostic cognition primitives (DCPs) capture reusable presentation patterns and actionable workup guidance from encounter trajectories.The workflow requests evidence iteratively and grounds decisions in traceable observations.Source: §1 · pp. 2–3, Figure 1 · p. 4
• 11.2% mean accuracy gain over the competitive baseline system was reported on MIMIC-CDM; DxEvolve reached 90.4% accuracy versus the 88.8% human expert on the reader-study subset.The reported external-cohort gains were 10.2% on translated records, 11.9% on raw Chinese documentation, and 17.1% for categories absent from the initial repository.Source: §1 · p. 3
• DxEvolve’s experience artifacts reportedly mature with exposure, while heuristics from prior diagnostic failures are associated with greater subsequent performance gains.The paper describes a transparent pathway for clinician-led oversight and continuous improvement.Source: §1 · p. 3
• The paper frames clinical AI competence as reliable improvement with exposure during procedural evidence acquisition, rather than snapshot performance alone.It presents workflow faithfulness and governance as supports for inspection, curation, and controlled updating.Source: §1 · p. 3

2.1 Experimental design and the DxEvolve framework

• The evaluation uses a held-out MIMIC-CDM cohort of 400 encounters, non-overlapping DCP accrual data, an 80-encounter reader-study subset, and an independent Chinese PLA General Hospital cohort of 293 encounters.The external cohort includes diagnostic categories both overlapping with and absent from MIMIC-CDM.Source: §2.1 · p. 5

2.2 DxEvolve achieves clinician-level diagnostic performance

• 11.2% mean accuracy gain over the established CDM baseline and 9.1% over DxEvolve w/o DCP were reported on MIMIC-CDM; DxEvolve reached 90.4% versus the 88.8% clinician reference on the reader-study subset.The 90.4% versus 88.8% comparison uses workflow-constrained DxEvolve and a clinician reference diagnosed with full information, not matched information access.Source: §2.2 · pp. 5, 7, Figure 2 · p. 6

2.3 External validation supports cross-institution portability of experiential gains

• 10.2% mean accuracy gain over the CDM baseline on translated external records, 17.1% on categories absent from DCP accrual, and 11.9% on original Chinese records were reported.Accuracy decreased for appendicitis and cholecystitis but improved for pancreatitis; broader transferability across heterogeneous syndromes warrants further investigation.Source: §2.3 · pp. 7–9

2.4 Self-evolution shows exposure-dependent scaling behavior and error-driven correction

• 8.97% mean accuracy gain followed accrual over the first 0–1,000 encounters, with a further 0.9% gain over 1,000–2,000; weaker backbones plateaued while more capable models continued incremental growth.Retrieved DCPs from prior diagnostic failures were enriched in encounters where DxEvolve succeeded and its baseline failed.Source: §2.4 · p. 10

2.5 Self-evolution is accompanied by progressive maturation of experience

• Late-stage DCPs received higher mean clinician ratings than early-stage DCPs, 4.47 vs 4.17 on a 5-point scale, and their retrieval prevalence rose from 12.4%–13.5% overall to 13.9%–15.9% in error-correcting episodes.Clinician ratings covered correctness, actionability, and generality; aggregate-score inter-rater reliability was ICC=0.81.Source: §2.5 · pp. 11–12, Figure 5 · p. 11

2.6 DxEvolve’s evidence acquisition aligns with clinical workflows and clinical guidelines

• DxEvolve’s mean overall trajectory consistency with recorded workups was 0.89 versus 0.68 for the standard workflow-aligned baseline, and it achieved higher guideline compliance than CDM across evaluated backbones.Consistency measures included physical examination, laboratory and imaging sets, and action-order concordance.Source: §2.6 · p. 12

3 Discussion

• DxEvolve reached clinician-comparable performance under interactive diagnosis and showed consistent DCP-enabled gains on an external cohort from a distinct healthcare system.The discussion describes this as supporting portability under cross-institutional shift.Source: §3 · pp. 13–15
• DxEvolve turns procedural evidence acquisition and hypothesis refinement into reusable diagnostic cognition primitives that can guide later workups without parameter updates.These primitives are inspectable, curatable, and retractable, making experience-driven adaptation reviewable.Source: §3 · pp. 12, 14
• Later-accrued experiences scored higher on clinical correctness, actionability, and generality, and were retrieved across more encounters and more often in error-correcting episodes.The authors attribute gains to experience refinement as well as accumulation.Source: §3 · pp. 14–15
• DxEvolve showed higher agreement with recorded workups and improved guideline alignment, providing an auditable basis for calibrating investigation intensity and reviewing recurrent failures.The figure caption covers physical examination, laboratory and imaging overlap, investigation ordering, and guideline-compliance score distributions.Source: Figure 6 · p. 13, §3 · p. 15
• Evaluation used de-identified EHR records; prospective studies should incorporate clinician–patient interaction and broaden evaluation across case mixes and clinical contexts.The current action schema emphasizes diagnostic-relevant actions, with richer actions proposed for specific deployments.Source: §3 · p. 15

4.1 DxEvolve framework

• DxEvolve iteratively requests evidence, updates an explicit encounter state and hypotheses, and reuses compatible prior experience without updating the base LLM.Completed trajectories are consolidated into diagnostic cognition primitives (DCPs) containing a presentation pattern, test-ordering guidance, and diagnostic decision experience.Source: §4.1 · pp. 16–17

4.2 Data sources

• MIMIC-CDM contains 2,400 acute-abdominal-pain presentations across four diagnoses, with examinations, laboratory results, and imaging revealed only when requested.Inputs excluded diagnosis fields and label-bearing metadata; evidence items were retained in record order and revealed only after the corresponding request.Source: §4.2 · p. 18

4.3 Evaluation cohorts

• The study used a fixed 400-encounter MIMIC-CDM evaluation cohort, a separate 80-encounter reader-study subset, and an independent 293-encounter Chinese hospital cohort.The external cohort included three diagnoses represented in MIMIC-CDM and two additional diagnoses; it was used only for out-of-distribution evaluation, not DCP accrual.Source: §4.3 · pp. 18–19

4.4 Ethics approval and governance

• MIMIC-derived and hospital records were de-identified and handled under their respective data-use agreements, institutional policies, and ethics approval.The hospital’s institutional ethics committee approved the external cohort study and waived informed consent where applicable.Source: §4.4 · p. 19

4.5 Models and implementation

• Experiments used open-weight LLM backbones, a shared prompt contract, and local inference; medical-domain models were evaluated only in single-pass full-information mode.Those models did not reliably follow the structured action-calling protocol required for workflow-aligned evaluation.Source: §4.5 · pp. 19–20

4.6 Evaluation and analysis

• Episodes ended with a valid primary diagnosis or after 20 interaction steps, and the primary endpoint was encounter-level diagnostic correctness.The interactive regime revealed only requested evidence; single-pass full-information inference was evaluated only on the reader-study subset (n=80).Source: §4.6 · p. 21
• Exposure analyses varied DCP accrual while holding the 400-encounter evaluation cohort fixed, and improvement cases required correct DxEvolve diagnoses, incorrect DCP-free diagnoses, and DCP retrieval.Provenance analyses compared source outcomes of DCPs retrieved in improvement cases with those retrieved across the full evaluation cohort.Source: §4.6 · p. 22
• The clinician assessment sampled 40 early- and late-exposure DCPs, rated correctness, actionability, and generality, and found high aggregate-score agreement (weighted κ=0.83, ICC=0.81).Process analyses also compared requested investigations with the structured record and assessed trajectory consistency and guideline-adherence proxies.Source: §4.6 · pp. 22–23

• MIMIC-CDM is available through its original release under MIMIC-IV access terms, and preprocessing and cohort-splitting scripts are available on GitHub.The external cohort is not public; qualified researchers may request the minimum data needed for reproduction subject to hospital governance approval and a data-use agreement.Source: Data availability · p. 24

• The shared diagnostic prompt template enables medical examinations, experience retrieval, clinical guidelines, and PubMed search across the reported experiments.Template variables are replaced with model-specific chat delimiters at runtime.Source: Diagnostic Prompt Template · p. 28
• Before diagnosing, the prompt requires experience and guideline searches and directs the agent to integrate retrieved guidance only when relevant and consistent with objective patient data.It also instructs the agent to stop when a confident diagnosis is possible or provide the best-supported diagnosis when uncertainty remains without a high-yield action.Source: Diagnostic Prompt Template · p. 29
• The prompt directs the agent to perform stepwise reasoning in an action loop, using iterative information-gathering outputs before a final-diagnosis format.The interaction uses a Thought → Action → Observation sequence for gathering information.Source: Diagnostic Prompt Template · p. 30

• After each completed case, the construction template distills the diagnostic trajectory into a DCP for storage and retrieval in future cases.Its inputs include the case, diagnostic steps, model and ground-truth diagnoses, correctness, process feedback, and clinician test orders.Source: B Experience Construction Prompt… · pp. 30, 32
• Recommended tests and imaging settings must be supported by clinician orders, successful tool calls, rule-based feedback, or retrieved guidance; the prompt forbids inventing unsupported tests.It prioritizes covering the clinician orders and successful tool calls before adding other supported recommendations.Source: B Experience Construction Prompt… · p. 31
• Correctness feedback shapes consolidation: correct cases yield reusable patterns and ordering heuristics, while incorrect diagnoses are treated as negative examples and corrected using ground truth and feedback.The template includes case context, intermediate steps, diagnoses, correctness, rule-based process feedback, and charted clinician orders.Source: B Experience Construction Prompt… · pp. 31–32
• Each DCP contains an experience pattern for retrieval, prioritized test-ordering guidance, and a concise diagnostic decision rule for weighing findings.The prompt permits ranking actions and specifying escalation criteria while expressing lower-priority tests conditionally rather than as blanket prohibitions.Source: B Experience Construction Prompt… · pp. 31, 33

Correct-case DCP

• For suspected biliary pancreatitis, the example recommends confirming pancreatitis with laboratory tests and starting biliary imaging with ultrasound.If ultrasound is negative but liver enzymes remain elevated, it recommends ERCP; CT or MRCP require specific indications.Source: C Example Diagnostic Cognition Primitive · p. 34
• The example decision experience treats post-cholecystectomy pancreatitis with transaminitis or hyperbilirubinemia as biliary until proven otherwise, despite a negative initial ultrasound.It identifies microlithiasis or sphincter dysfunction as possible causes.Source: C Example Diagnostic Cognition Primitive · p. 34
• A correct-case DCP contains an experience pattern, test-ordering experience, and diagnostic decision experience.These encode discriminative presentation cues, actionable workup guidance, and evidence-linked implications for hypothesis refinement and final decisions.Source: C Example Diagnostic Cognition Primitive · pp. 33–34, Table 3 · p. 33

DCP examples

• The two example DCPs show how both correctly diagnosed and misdiagnosed trajectories can be consolidated into reusable experience artifacts.The incorrect-case DCP uses the ground truth and evaluator feedback to label the misdiagnosis and provide a corrected reasoning pathway.Source: C Example Diagnostic Cognition Primitive · pp. 33–35, Table 3 · p. 33, Table 4 · p. 34

Case context

• The trace concerns an elderly patient with diabetes and acute right upper-quadrant pain, ultimately diagnosed with acute calculous cholecystitis.The annotation reports right upper-quadrant tenderness and neutrophilia alongside normal liver enzymes and lipase.Source: D Example Diagnostic Reasoning Trace · pp. 35, 36, 43

Diagnostic reasoning trace

• Initial ultrasound showed gallstones, 12 mm common bile duct dilation, and trace pericholecystic fluid; the later MRCP showed progressive gallbladder wall edema and enhancement.The common bile duct dilation with abrupt tapering was interpreted as suggestive of papillary stenosis from previously passed stones.Source: D Example Diagnostic Reasoning Trace · pp. 38, 40, 41
• The trace’s final diagnosis was acute calculous cholecystitis, matching the ground-truth label.The trace describes the diagnosis as supported by MRCP findings and retrieved clinical experiences and guidelines.Source: D Example Diagnostic Reasoning Trace · pp. 42–43
• The agent proceeded from examination and laboratory testing to ultrasound, then escalated to MRCP for more definitive biliary assessment.The trace reports normal liver enzymes and lipase, while MRCP showed gallbladder wall edema, pericholecystic fluid, and wall enhancement.Source: D Example Diagnostic Reasoning Trace · pp. 40, 41, 43