Auditable Clinical Support with Decision Agents

Tiancheng Bao, Riga Wu, Jan Wienhold, Sebastian Billig, Patrick Winnersbach, Hsuan-Chih Chen, Akis Giannoukos, Jiayang Li, Li Shen, Gary Weissman, Eduardo J. Mortani Barbosa, Dyke Ferber, Jakob Nikolas Kather, Christos Davatzikos, Daniel Truhn, Tianyu HanarXiv 2026 · arXiv

TL;DR

Researchers have developed Decision2Care, a clinical "decision agent" that performs medical work-ups, selects treatments, and predicts mortality without generating a single word of free-text reasoning. By using a probabilistic framework of "typed decisions," the agent achieves diagnostic accuracy comparable to state-of-the-art Large Language Models (LLMs) while being 40 times faster, significantly cheaper, and—most importantly—fully auditable.

What Problem Does This Paper Address?

In recent years, LLM-based agents have shown they can match physicians in diagnostic accuracy. However, these models have a "black box" problem: their written explanations (rationales) don't always reflect the actual factors that led to a decision—a phenomenon known as "unfaithful reasoning." Furthermore, generating long strings of text makes these agents slow and expensive to run at scale.

In a high-stakes environment like an Emergency Department (ED) or Intensive Care Unit (ICU), clinicians need to know exactly what information was considered at each step and why a diagnosis changed. If an agent misses a pulmonary embolism, a reviewer needs to see if the model ignored a critical CT report or if the data itself was misleading. Current text-based agents make this level of granular auditing difficult and costly.

Methods

Decision2Care replaces free-text generation with a system of typed decisions. Instead of writing a paragraph, the model answers two specific types of questions:

  1. Choices: The model assigns a probability to each item in a predefined list (e.g., "Which imaging study is most appropriate?").
  2. Nouls: The model returns the probability that a specific statement is true (e.g., "Should we order a lipase test now?").
Figure 1: Overview of Decision2Care. Decision2Care puts 2 kinds of typed questions into its decision model: for a Choice, the model returns a probability for each named option, and for a Noul, it returns the probability that a statement is true. With these questions, Decision2Care codes the final diagnosis (a), follows the differential diagnosis (b), runs the work-up (c) with tools (d) and recommends treatment (e). a, The final diagnosis. At the end of the work-up, the decision model answers a Choice at each level of the ICD-10-CM tree and then among the billable codes, gives I26.94. b, The differential diagnosis. After every action, a Choice over the diseases of the evaluation and ‘other’ gives the probability of each diagnosis without changing the work-up. In the patient shown, the probability of pulmonary embolism was 0.45 after the abdominal CT and 0.96 after the chest CT. c, The work-up loop. In each round, the decision model reads the current clinical state and answers a Choice over the available actions. The controller then runs the selected tool, which retrieves the recorded findings and adds them to the state. The loop continues until the Choice selects the final diagnosis. d, Four of the 5 tools (the review of the home medication list is not shown). For the physical examination and the vital signs, a Noul is asked for each item, and the tool retrieves every item with a probability of at least 0.5. For laboratory tests, a Choice first ranks the groups of tests and a Noul then selects specific tests within a group. For imaging, a Choice selects one study, defined by its modality and body region, or the option that no further imaging is indicated. e, Treatment recommendation. After the final diagnosis, a Noul for each medication and procedure category selects the indicated options. Choices then select the drug regimen with its dose, route and frequency, the action on each home medication and the ICD-9-CM procedure.

A deterministic controller then applies fixed rules to these probabilities. For instance, if a "Noul" for a specific test reaches a probability of 0.5, the controller executes that action. This architecture separates the "thinking" (the decision model) from the "doing" (the controller), ensuring that every step is recorded as a set of quantifiable probabilities.

Datasets

The authors evaluated the agent using:

  • MIMIC-IV: Records from 546 emergency admissions across 12 diseases (e.g., appendicitis, meningitis, heart failure).
  • MIRA-v2 Benchmark: An existing benchmark for medical reasoning used to compare Decision2Care against LLMs like GPT-4 and Claude.
  • INSPECT: A dataset from Stanford Health Care used for external validation of mortality predictions in pulmonary embolism and pneumonia.
  • MIMIC-III/IV Mortality Cohorts: Used to test the agent's ability to predict death within 7, 30, and 90 days.

Key Insights and Ideas

  • Text is Optional for Reasoning: The paper proves that clinical agents don't need to "talk" to be smart. By restricting the model to probabilistic choices, the researchers eliminated the risk of "hallucinated" tests or unfaithful rationales.
  • The "Aha!" Moment is Quantifiable: Because the agent tracks the "differential diagnosis posterior" at every step, you can see exactly which test result caused the model to pivot. In 10 of 12 diseases, the largest jump in diagnostic probability occurred exactly at the test recommended by medical guidelines.
  • Auditability as a Safety Feature: When the agent made a high-confidence error, the authors could "debug" the decision. They traced errors back to specific misleading imaging reports. By removing just that one report and re-running the work-up, the agent often corrected itself.
  • Efficiency Gains: By avoiding text generation, Decision2Care is a speed demon. A full work-up takes a median of 2.2 seconds, compared to over 90 seconds for text-based LLM agents.

Key Results

Decision2Care matched or exceeded the performance of much larger, more expensive models across several metrics:

  • Diagnostic Accuracy: It reached 86.5%, placing it squarely in the range of top-tier LLMs (86.1% to 91.7%) and significantly higher than the 77.6% accuracy of board-certified physicians on the same benchmark.
  • Cost and Speed: It was 40 times faster and 16 to 61 times cheaper than the MIRA-v2 LLM workflow.
  • Treatment Adherence: Its drug orders were judged to be guideline-adherent more often than those of the LLM comparators and even the documented care provided by human clinicians on the first day of admission.
  • Mortality Prediction: In the ICU, Decision2Care’s predictions from just 2–6 hours of data were as good as or better than standard 24-hour clinical severity scores (like APS III or OASIS).
Table 1: Comparison of computational efficiency and disease-specific diagnostic accuracy between Decision2Care and MIRA-v2

Limitations and Future Work

While the results are impressive, the authors note several caveats:

  • Retrospective Nature: The study used historical data. Real-world prospective studies are needed to see how clinicians actually interact with these agents.
  • Calibration: The diagnostic confidence scores are useful for ranking but need better calibration to ensure a "90% confidence" score truly means a 90% chance of being correct in a diverse patient population.
  • Confidently Wrong: The agent can still be misled by a single "dirty" data point (like a radiologist's report that uses the word "compatible with" a certain disease). Future versions might need better ways to weigh conflicting evidence.

Ultimately, Decision2Care suggests a future where AI clinical support is not a chatty assistant, but a transparent, probabilistic engine that clinicians can audit, verify, and trust.

Summary updated

Abstract

Clinicians reviewing a diagnostic work-up will ask what was considered at each step and why the assessment changed. Agents powered by free text can reach physician-level diagnostic accuracy, but their written reasoning does not always reflect the factors that went into the decision. Here, we present Decision2Care, a decision agent that works up the acutely ill patient without producing any text. A decision model assigns a probability to each permitted EHR action, and fixed rules carry out the most probable one. We evaluated Decision2Care on the records of 546 emergency admissions across 12 diseases. Decision2Care reached a diagnostic accuracy of 86.5%, compared with 86.1% for the MIRA-v2 agent, and required a median of 2.2 s per work-up, compared with 92-100 s. Each work-up tracked which options Decision2Care was considering at each step. In 10 of the 12 diseases, the differential diagnosis changed most at the test on which the guideline bases the diagnosis. Where Decision2Care was most confident after a single work-up, all but four of its diagnoses were correct (98.5%). We traced each of these four errors back to a single imaging report. Three of them were corrected when that report was removed. The same kind of decision was used to select the treatments, and Decision2Care's drug orders were judged guideline-adherent more often than MIRA-v2's. In emergency patients, its prediction of death was comparable to that of three intensivists on the same cases. In five of the eight intensive care cohorts, it exceeded the best of the three 24-h severity scores within 2-6 h of records. Our results show that decision agents can carry out clinical work-ups while recording the options, probabilities, and execution rules behind each action for audit.
BibTeX
@misc{bao2026auditable,
  author = {Tiancheng Bao and Riga Wu and Jan Wienhold and Sebastian Billig and Patrick Winnersbach and Hsuan-Chih Chen and Akis Giannoukos and Jiayang Li and Li Shen and Gary Weissman and Eduardo J. Mortani Barbosa and Dyke Ferber and Jakob Nikolas Kather and Christos Davatzikos and Daniel Truhn and Tianyu Han},
  title = {{Auditable Clinical Support with Decision Agents}},
  year = {2026},
  eprint = {2611.23700},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.23700},
}

Loading related content...

Figure