Organizations: Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China. · Peking University Third Hospital, Beijing, China. · Baidu Inc., Beijing, China. · Chinese PLA General Hospital, Beijing, China. · Joint NTU-UBC Research Centre of Excellence in Active Living for the Elderly, Nanyang Technological University, Singapore.
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.
Figures & tables
Figure 1: DxEvolve: workflow-aligned diagnosis with experience-driven self-evolution. a , DxEvolve frames diagnosis as evidence-centered sequential reasoning, contrasting the static, single-pass inference typical of retrospective evaluations using complete records. b , Deep clinical research (DCR) workflow. From the patient history context, the agent iteratively plans the next step, requests evaluations (physical examination, laboratory tests and imaging) and, when necessary, consults external sources (guidelines and PubMed); only requested observations are revealed and are integrated into a compact high-salience encounter state to guide subsequent actions until final diagnosis. c , Diagnostic cognition primitives (DCPs). After each diagnosis reasoning, DxEvolve consolidates a DCP from the trajectory, consisting of a retrievable presentation pattern and evidence-linked guidance for investigation planning and diagnostic decision-making; DCPs are indexed in a repository and selectively reused in later encounters as an action like medical evaluation and searching external sources under the same DCR workflow. d , Cohorts and protocol. DCPs are built from a MIMIC-CDM accrual pool that is strictly non-overlapping with evaluation encounters, then assessed on a held-out in-distribution MIMIC-CDM cohort and an external hospital cohort for out-of-distribution evaluation.
Figure 2: Main diagnostic performance results on MIMIC-CDM. a , Diagnosis accuracy on the MIMIC-CDM evaluation cohort ( n =400), reported per pathology and as the average. For each base LLM (color), we compare the CDM baseline, DxEvolve without DCP retrieval (DxEvolve w/o DCP), and DxEvolve over multiple seeds. b , Accuracy improvement of DxEvolve over the CDM baseline stratified by encounter-level diagnostic burden (easy versus hard). Points show the stratum-specific improvement for each base LLM; annotations indicate the improvement in each stratum and the between-stratum difference. c , Diagnosis accuracy on a reader-study subset of MIMIC-CDM ( n =80). Bars report average diagnostic accuracy for CDM and DxEvolve distinguished by light and dark shades of the same color, together with single-pass full-information (FI) inference (hatched). Specialist medical LLMs with limited action compliance are reported under FI only. The clinician reference (Doctors) corresponds to the published reader-study subset with full information available Hager et al. (2023) .
Figure 3: External validation on an independent hospital cohort. a , Diagnostic accuracy on diagnoses overlapping with MIMIC-CDM (appendicitis, cholecystitis and pancreatitis) and their mean, evaluated using standardized English translations of the structured records. b , Category-level transfer on diagnoses that were never used for DCP accrual (liver abscess, urinary tract infection) and their mean, evaluated under the same protocol. c , Robustness to documentation with native institutional language, evaluated on the same external encounters using the original Chinese records.
Figure 4: Exposure-dependent self-evolution and provenance of retrieved experience. a , Overall diagnosis accuracy on the fixed MIMIC-CDM evaluation cohort ( n =400) as the DCP accrual pool increases, shown for three representative base LLM backbones. Accuracy improves with additional accrual encounters and then tapers, yielding a saturating learning curve. b , Provenance of retrieved experience during evaluation. Bars show the fraction of retrieved DCPs whose source accrual episode ended in an incorrect diagnosis (“incorrect experience rate”), computed separately for improvement cases and for all evaluation encounters pooled. P values indicate enrichment of incorrect-source DCPs among retrievals in improvement cases.
Figure 5: Maturation of accrued experience artifacts with encounter exposure. a , Blinded clinician ratings of diagnostic cognition primitives (DCPs) sampled from an early exposure window (encounters 1–300; n =20) and a late window (encounters 1700–2000; n =20). DCPs were scored for clinical correctness, actionability and generalizability, with the mean shown as an aggregate. Boxes denote interquartile range, centre line the median, and points individual DCPs; two-sided P values are shown (n.s., not significant). b , Inter-rater reliability of clinician ratings for the aggregate DCP score (ICC=0.81), supporting the reliability of the clinician assessment. c , Evaluation-time retrieval signal for late-stage DCPs, quantified as the fraction of retrieval events that involve DCPs in the late encounter window.
Figure 6: DxEvolve produces more workflow-consistent investigations and shows improved alignment with clinical guidelines. a , Workup consistency. Across the MIMIC-CDM evaluation cohort ( n =400), DxEvolve shows higher agreement with the documented investigation trace than the standard decision-making baseline CDM for each backbone, spanning whether a physical examination was performed, overlap with recorded laboratory testing, overlap with recorded imaging (modality and region), and concordance of the investigation ordering. Points are model-level means; grey lines connect paired results for DxEvolve versus CDM under the same backbone. b , Guideline adherence. Distributions of encounter-level guideline-compliance scores, derived from the mean adherence across three dimensions: physical examination, laboratory investigations, and imaging. Violin plots show score densities; embedded boxplots indicate the median and interquartile range; points mark the mean. P values are from paired two-sided comparisons.
{system_tag_start}
You are a senior physician. Your task is to perform stepwise diagnostic reasoning using ONLY the allowed tools. You must strictly follow one of the two output formats below at every step.
FORMAT A. INFORMATION GATHERING
Thought: [1-2 concise sentences: what you know + what uncertainty remains + why next action is needed]
You extract reusable diagnostic reasoning experience from completed clinical cases for future tool using agents.
Your goal:
- Do NOT retell the full case or reproduce chain of thought.
- Do NOT include treatment.
- Distill ONE Diagnostic Cognition Primitive (DCP): a short heuristic that improves future diagnosis.
Supplementary Table 2: Experience Construction Prompt.
Experience Pattern:
Post-cholecystectomy patient with acute RUQ/back pain, elevated liver enzymes and lipase. (Acute pancreatitis, DDx: Biliary pancreatitis vs. other etiologies)
Test Ordering Experience:
First, confirm pancreatitis with Laboratory Tests: Lipase, Amylase, CBC, CMP. Concurrently, order first-line biliary imaging: Imaging: modality=Ultrasound, region=Abdomen.
If ultrasound is negative for stones/dilation but liver enzymes remain elevated, escalate to definitive biliary evaluation (ERCP) per clinician orders; do not escalate to CT or MRCP without specific indications (e.g., concern for complications or failed ERCP).
Supplementary Table 3: Example DCP (Correct Case).
Experience Pattern:
Young to middle-aged female with acute right abdominal pain, sharp on palpation, bilious vomiting, chills, and history of prior abdominal surgeries (e.g., laparoscopies). Past medical history of endometriosis. (Correct: cholecystitis; Common misdiagnosis: adhesive small bowel obstruction)
Supplementary Table 4: Example DCP (Incorrect Case).
Patient History:
Elderly male patient with history of type 2 diabetes mellitus on insulin, hypothyroidism, hypertension, and prostate cancer status-post radiotherapy presented to the emergency department with acute onset abdominal pain. The patient reported that the pain began suddenly at approximately 3 AM, waking him from sleep. He described it as sharp, constant, and localized to the right side of the abdomen. When the pain persisted, he initially attempted to contact his primary care physician but was unable to secure an appointment, prompting him to call emergency services.
In the emergency department, initial vital signs were temperature 98°F, blood pressure 150/86 mmHg, heart rate normal, respiratory rate 16 breaths per minute, oxygen saturation 100% on room air. Initial laboratory studies showed lactate 2.3 mmol/L, glucose 279 mg/dL, normal liver function tests, and hematocrit 33.2%. Right upper quadrant ultrasound demonstrated extrahepatic biliary duct dilation with common bile duct measuring 10 mm and presence of gallstones. Computed tomography of the abdomen and pelvis confirmed these findings. He was treated with intravenous fluids, morphine for pain control, and ondansetron for nausea.
On admission to the hospital floor, the patient reported improved but persistent right upper quadrant abdominal pain. He endorsed dry heaves and constipation, with his last bowel movement occurring two days prior. He denied fever, chills, chest pain, or shortness of breath.
Supplementary Table 5: Example Diagnostic Reasoning Trace.
Tool-augmented large language model (LLM) agents can orchestrate specialist classifiers, segmentation models, and visual question-answering modules to interpret chest X-rays. However, these agents still solve each case in isolation: they fail to accumulate experience across cases, correct recurrent reasoning mistakes, or adapt their tool-use behavior without expensive reinforcement learning. While a radiologist naturally improves with every case, current agents remain static. In this work, we propose Evo-MedAgent, a self-evolving memory module that equips a medical agent with the capacity for inter-case learning at test time. Our memory comprises three complementary stores: (1)~Retrospective Clinical Episodes that retrieve problem-solving experiences from similar past cases, (2)~an Adaptive Procedural Heuristics bank curating priority-tagged diagnostic rules that evolves via reflection, much like a physician refining their internal criteria, and (3)~a Tool Reliability Controller that tracks per-tool trustworthiness. On ChestAgentBench, Evo-MedAgent raises multiple-choice question (MCQ) accuracy from 0.68 to 0.79 on GPT-5-mini, and from 0.76 to 0.87 on Gemini-3 Flash. With a strong base model, evolving memory improves performance more effectively than orchestrating external tools on qualitative diagnostic tasks. Because Evo-MedAgent requires no training, its per-case overhead is bounded by one additional retrieval pass and a single reflection call, making it deployable on top of any frozen model.
Weixiang Shen, Bailiang Jian, Jun Li +6
Technical University of Munich (TUM) · TUM University Hospital · LMU Munich +3
Interactive clinical agents operate under partial observability, so reliable care depends on reaching the correct diagnosis through evidence-grounded, safe interactions. Yet existing agents struggle to convert experience into reusable process knowledge with explicit provenance and authority. To address this gap, we introduce MediSkill-Evo, which self-evolves governed process knowledge without fine-tuning the backbone. It realizes this self-evolution by updating clinical, process, symbolic, and visual knowledge in four typed banks under type-specific validation and scope rules. The Process-Constrained Preference Harness then turns validated knowledge into action by grounding candidates in evidence and prioritizing safer decisions. We evaluate on 300 MIMIC-IV-derived FullChain encounters, 180 hard-isolation conditions covering six process obligations, and 100 multimodal NEJM image-diagnosis cases. On Qwen FullChain, MediSkill-Evo improves diagnosis accuracy by 7.81% and treatment-intent coverage by 70.67% over the best-performing prior agent, while reducing critical failures by 43.04%. Under stress, it improves the stress-process composite by 7.77% and required-action completion by 12.41% over the best-performing agent for each metric, with stronger patient-fact, temporal-evidence, and triage-red-flag recovery and no controller-scored errors in unavailable-evidence, treatment, and triage safety checks. On multimodal NEJM diagnosis, MediSkill-Evo with optional MedSAM localization improves diagnosis accuracy by 2.56% and core score by 18.96% over the best-performing memory agent. Code is available at https://anonymous.4open.science/r/mediskill-evo_anonymous-68E7.
Ruoyu Wu, Shenfu Xie, Yinqian Sun +2
Brain-inspired Cognitive AI Lab, Institute of Automation, Chinese Academy of Sciences · Beijing Key Laboratory of Safe AI and Superalignment · Beijing Institute of AI Safety and Governance +2
The application of large language models (LLMs) in clinical decision support faces significant challenges of "tunnel vision" and diagnostic hallucinations present in their processing unstructured electronic health records (EHRs). To address these challenges, we propose a novel chain-based clinical reasoning framework, called DxChain, which transforms the diagnostic workflow into an iterative process by mirroring a clinician's cognitive trajectory that consists of "Memory Anchoring", "Navigation" and "Verification" phases. DxChain introduces three key methodological innovations to elicit the potential of LLM: (i) a Profile-Then-Plan paradigm to mitigate cold-start hallucinations by establishing a panoramic patient baseline, (ii) a Medical Tree-of-Thoughts (Med-ToT) algorithm for strategic look ahead planning and resource aware navigation, and (iii) a Dialectical Diagnostic Verification procedure utilizing "Angel-Devil" adversarial debates to resolve complex evidence conflicts. Evaluated on two real world benchmarks, MIMIC-IV-Ext Cardiac Disease and MIMIC-IV-Ext CDM, DxChain achieves state-of-the-art performances in both diagnostic accuracy and logical consistency, offering a modular and reliable architecture for next-generation clinical AI. The code is at https://anonymous.4open.science/r/Dx-Chain.
Zhiqi Lv, Duofan Tu, Jun Li +4
School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei, Anhui, 230026, P.R. China · Suzhou Institute for Advanced Research, University of Science and Technology of China, Suzhou, Jiangsu, 215123, P.R. China · Key Lab of Intelligent Information Processing of Chinese Academy of Sciences (CAS), Institute of Computing Technology, CAS, Beijing, PR China +2