Uncertainty-Guided Latent Diagnostic Trajectory Learning for Sequential Clinical Diagnosis

Xuyang Shen; Haoran Liu; Dongjin Song; Martin Renqiang Min

cs.AIApr 6, 2026

Uncertainty-Guided Latent Diagnostic Trajectory Learning for Sequential Clinical Diagnosis

Authors: Xuyang Shen, Haoran Liu, Dongjin Song, Martin Renqiang Min

Organizations: University of Connecticut · Texas A&M University · NEC Laboratories America

Abstract

Clinical diagnosis requires sequential evidence acquisition under uncertainty. However, most Large Language Model (LLM) based diagnostic systems assume fully observed patient information and therefore do not explicitly model how clinical evidence should be sequentially acquired over time. Even when diagnosis is formulated as a sequential decision process, it is still challenging to learn effective diagnostic trajectories. This is because the space of possible evidence-acquisition paths is relatively large, while clinical datasets rarely provide explicit supervision information for desirable diagnostic paths. To this end, we formulate sequential diagnosis as a Latent Diagnostic Trajectory Learning (LDTL) framework based on a planning LLM agent and a diagnostic LLM agent. For the diagnostic LLM agent, diagnostic action sequences are treated as latent paths and we introduce a posterior distribution that prioritizes trajectories providing more diagnostic information. The planning LLM agent is then trained to follow this distribution, encouraging coherent diagnostic trajectories that progressively reduce uncertainty. Experiments on the MIMIC-CDM benchmark demonstrate that our proposed LDTL framework outperforms existing baselines in diagnostic accuracy under a sequential clinical diagnosis setting, while requiring fewer diagnostic tests. Furthermore, ablation studies highlight the critical role of trajectory-level posterior alignment in achieving these improvements.

Figures & tables

Figure 1: A real-world example of the sequential clinical diagnosis system.

Figure 2: The proposed LDTL framework within a sequential diagnosis system. A planning LLM agent sequentially selects diagnostic tests and a diagnostic LLM agent updates the disease prediction and decides whether to stop based on the accumulated patient and test information.

	Accuracy					F1-Score
Method	Appendicitis	Cholecystitis	Diverticulitis	Pancreatitis	Overall	F1-Score
SFT-all	97.9	93.1	90.4	90.7	94.2	95.8
Fixed info (history) (ZS)	47.9	34.6	73.1	79.6	54.1	58.8
Fixed info (history + one test) (ZS)	50.5	47.7	76.9	72.2	58.3	60.5
Fixed info (history + two tests) (ZS)	60.9	60.7	78.8	80.5	67.2	68.2
All info (ZS)	73.9	70.0	82.7	87.9	76.9	76.7

Table 1: Performance comparison of our method LDTL and baseline methods. We report class-wise accuracies and F1-scores. ZS = zero-shot.

Method	Acc	F1	Appendicitis	Cholecystitis	Diverticulitis	Pancreatitis
RS	85.4 $\pm$ 1.3	84.9 $\pm$ 1.3	83.5 $\pm$ 1.1	85.8 $\pm$ 1.5	90.9 $\pm$ 2.8	85.4 $\pm$ 2.1
w/o LP	87.9 $\pm$ 0.7	87.8 $\pm$ 0.8	94.0 $\pm$ 1.3	90.9 $\pm$ 1.4	81.1 $\pm$ 1.7	78.1 $\pm$ 0.9
LDTL	93.0 $\pm$ 0.4	92.4 $\pm$ 0.4	97.6 $\pm$ 0.5	95.2 $\pm$ 1.0	88.4 $\pm$ 2.5	86.0 $\pm$ 2.4

Table 2: Ablation study on Llama3-8B. We compare random selection (RS), a state-conditioned planner without latent path regularization (w/o LP), and our full model with trajectory-level posterior alignment over five seeds.

Figure 3: Case studies comparing our method LDTL with the variant without latent path regularization (w/o LP).

Figure 4: Distribution of diagnostic termination steps on the test cases. Numbers indicate how many cases terminate after one, two, or three actions.

Method	1-step Acc	2-step Acc	3-step Acc
RS	85.2 $\pm$ 1.3	87.4 $\pm$ 4.0	83.4 $\pm$ 3.4
w/o LP	88.9 $\pm$ 1.0	90.9 $\pm$ 2.2	80.5 $\pm$ 2.8
LDTL	93.2 $\pm$ 0.4	95.1 $\pm$ 2.3	84.7 $\pm$ 2.2

Table 3: Step-conditioned accuracy and efficiency over five seeds. The table reports diagnostic accuracy conditioned on the termination step along with final accuracy.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

	Accuracy					F1-Score
Method	Appendicitis	Cholecystitis	Diverticulitis	Pancreatitis	Overall	F1-Score
SFT-all (llama-3-8b)	97.9	93.1	90.4	90.7	94.2	95.8
gpt-4o OpenAI (2024b)	99.4	98.4	88.4	88.8	95.6	95.7
gpt-4o-mini OpenAI (2024a)	94.7	91.5	82.6	86.1	90.5	90.6
claude-sonnet-4 Anthropic (2025)	99.3	98.4	90.3	88.8	95.8	95.8
gemini-3-flash Google DeepMind (2025)	99.4	94.6	90.3	90.7	95.2	95.3

Appendix

Table 4: Performance comparison of our proposed LDTL with additional commercial models. (For proprietary LLMs, we perform evaluation through inference-only API calls without using the data for model training, fine-tuning, or retention. This protocol follows the PhysioNet data use agreement and ensures that no protected clinical data are incorporated into external model training.)

Figure 5: Distribution of diagnostic termination steps on the test cases. Numbers indicate how many cases terminate after one, two, or three actions.

Medical Diagnosis

Medical diagnosis uses clinical data, tests, and images to identify diseases, assess risk, and guide care.

See all 261 papers

MedAction: Towards Active Multi-turn Clinical Diagnostic LLMs
May 8, 2026Hsin-Ling Hsu, Zizheng Wang, Donghua Zhang +9Healthcare Medical Diagnosis
National Chengchi University · Georgetown University · University of Michigan +3
Timely Clinical Diagnosis through Active Test Selection
Oct 21, 2025Silas Ruhrberg Estévez, Nicolás Astorga, Mihaela van der SchaarClinical Decision Support Clinical Decision-Making
University of Cambridge Cambridge, UK
From Uncertainty to Clinical Risk: Severity-Aware Conformal Planning for Interactive Medical Diagnosis
Aug 28, 2026Yue Zhou, Haiyang Zhou, Jin Zhang +4Risk-Sensitive Planning Medical Diagnosis
Lanzhou University, China · Hunan University, China · National University of Singapore, Singapore +2

May 8, 2026J/K move · Enter open · S save

MedAction: Towards Active Multi-turn Clinical Diagnostic LLMs

Hsin-Ling Hsu, Zizheng Wang, Donghua Zhang, Nai-Chia Chen+8

National Chengchi University · Georgetown University · University of Michigan+3

Healthcare Medical Diagnosis Clinical Reasoning LLM Agent Training Differential Diagnosis Medical LLMs

Open paper AskPDF arXiv

Most existing LLM diagnoses are evaluated on static, single-turn settings where complete patient information is provided upfront, an oversimplification of real clinical practice. We study active diagnosis: the real-life clinical process of starting from initial observation, ordering tests, interpreting results, and updating a differential diagnosis across multiple turns. Through systematic analysis, we identify three recurring failure modes in current LLMs: ungrounded test ordering, unreliable diagnostic update, and degraded multi-turn coherence. Together, these failures reveal a core deficit: existing medical training data teaches models to reason from complete information but not to act under evolving, partial evidence. To address this gap, we introduce MedAction, a tree-structured distillation pipeline that synthesizes diverse and high-quality multi-turn diagnostic trajectories via LLM-environment interaction. We propose two knowledge-graph-grounded metrics to filter trajectory quality: Disease Trajectory Consistency (DTC), which tracks whether the model's hypothesis converges toward the correct diagnosis, and Reasoning-Action Consistency (RAC), which verifies that belief updates are driven by gathered evidence. Using this pipeline, we construct MedAction-32K, a dataset of 32,681 trajectories from 2,896 PMC cases. Fine-tuning an 8B model on MedAction-32K achieves state-of-the-art performance among open-source models on both MedR-Bench and our curated MedAction-300-Hard benchmark, pushing the edge for open-source medical LLMs.

Abstract

Medical Diagnosis

Explore similar work

MedAction: Towards Active Multi-turn Clinical Diagnostic LLMs

Timely Clinical Diagnosis through Active Test Selection

From Uncertainty to Clinical Risk: Severity-Aware Conformal Planning for Interactive Medical Diagnosis