Authors: 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 Han
Organizations: Department of Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA. · Department of Anesthesiology, Medical Faculty, RWTH Aachen University, Aachen, Germany. · Department of Intensive Care and Intermediate Care, Medical Faculty, RWTH Aachen University, Aachen, Germany. · Department of Biostatistics, Epidemiology and Informatics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA. · Center for Artificial Intelligence and Data Science for Integrated Diagnostics (AI2D), Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA. · Pulmonary, Allergy, and Critical Care Division, Department of Medicine, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA. · Else Kröner Fresenius Center for Digital Health, Faculty of Medicine and University Hospital Carl Gustav Carus, TUD Dresden University of Technology, Dresden, Germany. · Department of Medical Oncology, National Center for Tumor Diseases (NCT), University Hospital Heidelberg, Heidelberg, Germany. · Institute for Medical AI (IMAI), Heidelberg University Hospital and Medical Faculty of Heidelberg University, Heidelberg, Germany. · Department of Diagnostic and Interventional Radiology, University Hospital RWTH Aachen, Aachen, Germany.
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.
Figures & tables
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.
Measure
n
Decision2Care
MIRA-v2 workflow
This work
GPT-OSS-120B
GLM-5
Claude Sonnet 5
Computational efficiency
Median time per work-up, s a
—
2.2 (1.7–2.7)
92.7 (74.3–116.6)
92.4 (75.5–121.8)
100.1 (87.5–119.1)
Cost per case, 1 work-up, US$ b
546
0.0024
0.0376
0.1266
0.1460
Total cost, 1 work-up, US$ b
546
1.31
20.5
69.1
79.7
Total cost, 5 work-ups, US$ c
546
—
102.6
345.5
398.5
Table 1: Comparison of computational efficiency and disease-specific diagnostic accuracy between Decision2Care and MIRA-v2
Figure 2: The differential diagnosis after every action. a – c , The differential diagnosis during 3 work-ups of the baseline run. The stacked areas show the probability of each disease and of none of the testing diseases (grey) on the history alone (H) and after each step. The true disease is shown in teal, and the line traces its probability. The cards give, for the history and the circled steps, the action, the findings returned and the resulting probability of the true disease. The table shows the probability assigned to each action at each decision, with the action taken outlined in orange and the circled steps in bold. a , A work-up of meningitis. b , The work-up of pulmonary embolism shown in Figure 1 . c , A work-up of acute cholecystitis. d , The mean change in the probability of the true disease per action, by disease, over 5 runs on the 546 paired admissions.
Figure 3: Audit of the 4 high-confidence diagnostic errors. a – d , The 4 admissions out of 546 in which Decision2Care retained an incorrect diagnosis at a confidence of at least 0.70 in the baseline run. Each table details the history and the condensed findings from each action, along with the diagnosis the agent would have coded at that point. The bars show the confidence in that diagnosis (navy), the summed probability of the ICD-10-CM codes that match the recorded disease under the benchmark’s rule (teal) and the probability of the code the agent finally chose (gold). The pink band indicates the arrival of the implicated imaging report and delta ( Δ ) indicates the change from the previous state. The dial with the red cross marks the first state in which confidence was at least 0.70 in a diagnosis different from the recorded one. The red text highlights the evidence that led to the change in assessment. The quotation marks denote verbatim extracts. The agent’s final diagnosis and the recorded principal diagnosis are both shown at the bottom of each work-up. a , Gallstone pancreatitis coded as acute cholecystitis. b , Cholecystitis coded as gallstones without cholecystitis. c , Cholecystitis coded as bile duct obstruction. d , Cholecystitis coded as haemoperitoneum with the single recorded code shown in teal. The CT report describes a postoperative haematoma. The benchmark uses the recorded principal diagnosis as the reference.
Figure 4: Treatment selection and its dependence on patient information. a , The medication decisions for one appendicitis admission with a listed cefaclor allergy. The agent selects medication categories, a regimen, and doses (2). It then checks each drug for a listed or cross-reacting allergy and for an excessive dose or a drug to avoid at the recorded kidney function (3). Here, the allergy check flagged ceftriaxone, then cefazolin; metronidazole with ciprofloxacin was kept (4). Categories were selected and concerns flagged at scores of at least 0.50, and regimen and dose probabilities are shared among the options offered. b , Treatment selections by disease in the 475 admissions that the agent diagnosed correctly. Filled circles give the share of admissions in which the agent selected each medication category or procedure, and rings the share in which the hospital recorded it in first-day orders or admission procedures. Circle area is proportional to the share. c , The probability that each treatment was indicated, decided again on all 546 admissions with information withheld. Rows are the treatments that the agent selected, with the number of admissions (n); procedure groups carry ICD-9-CM category names. The left heat map shows the mean probability for each set of information, with the diagnosis as the decision made. The right heat map gives the mean change from that decision when one kind of information is withheld and the diagnosis is kept. Bold marks cells in which at least 25% of decisions fell below 0.50, and dots mark cells with fewer than 3 admissions to withhold from.
Figure 5: Mortality predictions in intracerebral haemorrhage (ICH). a – d , Predictions after each retrieval in 4 patients with ICH. a , b , The probability of death within 30 days, predicted 8 h after arrival in the emergency department (ED), for a patient who died within 30 days ( a ) and for one who was alive at 30 days ( b ). c , d , The probability of death in hospital, predicted 12 h after admission to the intensive care unit (ICU), for a patient who died in hospital ( c ) and for one who was discharged alive ( d ). The curve gives the prediction with the disease alone (step 0) and after each retrieval. Shown are the mean and the range of 3 predictions made on the records read up to that step; the diamond is the prediction that the agent issued when it stopped. The heatmap shows the probability the agent gave to each class of records at each step, summed over the actions for that class. The class that it read is outlined in orange, and a dash marks a class that was not offered. The cards show the records returned at 4 steps, with the prediction after each.
Figure 6: Accuracy of the mortality predictions in the emergency department (ED) and intensive care unit (ICU). a , b , Receiver operating characteristic (ROC) curves of the prediction made 8 h after arrival in the ED for death within 30 days ( a ) and 90 days ( b ), from the mean of 5 runs on 8 disease cohorts (8,565 and 8,570 entries; 1,673 and 2,062 deaths). Crosses show the predictions of 3 intensivists, and the dot is their mean, on 72 cases in which wrong agent predictions were over-represented. The stacked bars show their confidence. c , AUROC for predicting in-hospital death from the first 1 to 12 h of ICU records, by disease. Bands are 95% bootstrap intervals. Grey lines show the Acute Physiology Score III (APS III), the Simplified Acute Physiology Score II (SAPS II) and the Oxford Acute Severity of Illness Score (OASIS) at 24 h in the same patients. The dot marks the first hour at or above the highest score.
Code
ICD-10-CM title
Answers
Share of errors (%)
K80.20
Calculus of gallbladder without cholecystitis without obstruction
98
68.1
K85.10
Biliary acute pancreatitis without necrosis or infection
20
13.9
K83.1
Obstruction of bile duct
10
6.9
K71.2
Toxic liver disease with acute hepatitis
5
3.5
K66.1
Hemoperitoneum
5
3.5
K80.71
Calculus of gallbladder and bile duct without cholecystitis with obstruction
5
3.5
Supplementary Table 1: ICD-10-CM codes of Decision2Care’s incorrect cholecystitis predictions. The 144 incorrect predictions over 5 runs on the 546 paired admissions, by the code given. Every reference code specifies acute cholecystitis. In 88 of the 98 predictions coded K80.20, the reference code lay in the same category, K80. Decision2Care’s accuracy in cholecystitis was 77.7% (95% confidence interval 70.2–84.3%; 129 admissions), and of the comparisons by disease with the 3 MIRA-v2 configurations, only that with Claude Sonnet 5 in cholecystitis remained significant after Holm’s correction over the 12 diseases ( p<0.001 ).
Disease
Physicians, correct
Physicians, accuracy, %
Decision2Care, admissions
Decision2Care, accuracy, %
Difference, points
Appendicitis
38 of 43
88.4
148
99.9
+11.5
Cholecystitis
36 of 45
80.0
129
77.7
−2.3
Diverticulitis
36 of 44
81.8
54
88.1
+6.3
Pancreatitis
33 of 42
78.6
52
91.5
+13.0
Pneumonia
15 of 26
57.7
21
45.7
−12.0
Pulmonary embolism
40 of 45
88.9
15
92.0
+3.1
Supplementary Table 2: Diagnostic accuracy of Decision2Care and of board-certified physicians on the MIRA benchmark. The physicians’ results are those published for the 4 board-certified physicians of the MIRA study [ 8 ] , who each worked up a random, non-overlapping share of 311 benchmark admissions. The 21 admissions with pancreatic cancer, which MIRA-v2 does not include, are left out. The admissions that the physicians read have not been published, so the 2 results were not compared admission by admission. In the 4 abdominal diseases (appendicitis, cholecystitis, diverticulitis and pancreatitis), the physicians’ admissions are a random sample of those on which Decision2Care was evaluated. Decision2Care’s accuracy is the mean of 5 runs. The pooled rows weight each disease by the number of admissions the physicians read. 95% confidence intervals (in parentheses) are Wilson intervals for the physicians and percentile intervals from 20,000 resamples of admissions within each disease for Decision2Care. For the difference, the physicians’ correct answers were resampled as well, from a binomial distribution. The physicians took the history from MIRA’s patient agent, whose instructions restricted disclosure of the emergency department course. The last row keeps only the admissions whose recorded history holds no such course (Methods).
Variant
Accuracy (%)
Change (points)
95% half- width
p
p (Holm)
ICD options in a seeded random order (seed 1)
87.7
+1.25
1.39
0.080
0.798
ICD options in a seeded random order (seed 2)
86.8
+0.33
1.17
0.579
1.000
ICD options in a seeded random order (seed 3)
86.6
+0.15
1.45
0.843
1.000
Every ICD Choice asked under 5 random orders, probabilities averaged
Supplementary Table 3: Decision2Care’s accuracy under variants of option order, wording and record format. Each variant run once on the 546 paired admissions and compared case by case with the mean of the 5 runs as recorded (86.48%). Change: the mean per-case difference in strict accuracy, in points, with the half-width of its 95% interval and the two-sided p value, uncorrected and after Holm’s correction over the 14 variants. JSON, JavaScript Object Notation.
Disease
n
Jev 1.13.0 (5 runs)
PPLX decider v1.2 (5 runs)
Qwen3.8-Flash-Next readout (5 runs)
Appendicitis
148
99.9
98.6
98.8
Cholecystitis
129
77.7
80.6
87.8
Diverticulitis
54
88.1
90.7
92.6
Pancreatitis
52
91.5
85.8
82.7
Pneumonia
21
45.7
41.9
52.4
Pulmonary embolism
15
92.0
86.7
70.7
Supplementary Table 4: Diagnostic accuracy of Decision2Care with 3 decision models on the 546 paired admissions. Each model drove the same controller, tools and typed questions over 5 runs: Jev 1.13.0, the primary decision model; local PPLX decider v1.2; and Qwen3.8-Flash-Next, a general large language model read through its answer-token probabilities (temperature 1, thinking disabled). Accuracy is the share of admissions whose top-ranked ICD-10-CM code title matched the recorded disease under the strict MIMIC-CDM rule, the mean over runs; 95% confidence intervals (CIs) from 5,000 admission-bootstrap resamples with all runs of an admission resampled together, and the paired differences from Jev 1.13.0 likewise. Intervals condition on the saved runs. Qwen minus PPLX, +1.3 points ( −1.1 to +3.6 ). Jev 1.14.0 reached 85.6% over 5 runs ( −0.8 from 1.13.0, −2.5 to +0.8 ) and PPLX’s hosted service 86.3% in 1 run. PPLX used its checkpoint’s decision readout and question splitting; Qwen used independent answer-token readouts (Methods).
Disease
Guideline
Tests counted
Appendicitis
WSES Jerusalem guidelines 2020 (Di Saverio et al., World J Emerg Surg 15:27) [ 21 ]
ultrasound, CT or MRI of the abdomen
Cholecystitis
WSES 2020 (Pisano et al., World J Emerg Surg 15:61) [ 26 ]
ultrasound, MRI or HIDA scan of the abdomen
Bowel obstruction
WSES position paper on small bowel obstruction 2021 (Amara et al., World J Emerg Surg, doi:10.1186/s13017-021-00379-8) [ 23 ]
WSES 2019 (Leppäniemi et al., World J Emerg Surg 14:27) [ 29 ]
ultrasound or CT of the abdomen; lipase or amylase
Pneumonia
ATS/IDSA 2019 (Metlay et al., Am J Respir Crit Care Med 200:e45) [ 60 ]
CT or radiograph of the chest
Supplementary Table 5: The guideline test of each disease. The disease-specific test categories used in the trajectory analyses and their source guidelines: MIMIC-IV-Ext Clinical Decision Making (MIMIC-CDM) for its 4 diseases [ 7 ] , Medical Intelligence for Reasoning and Action (MIRA) for the 3 it added [ 8 ] , and for the 5 added here a guideline of the clinical-guidelines corpus from the École polytechnique fédérale de Lausanne (EPFL) [ 65 ] or ESCMID. Examination, blood count and inflammatory markers were excluded from these mappings; for cholecystitis, CT was excluded. The bowel-obstruction source addresses a virgin abdomen and the meningitis source addresses bacterial disease; these documents define the operational mappings rather than universal diagnostic requirements.
Disease
Case- runs
Decisive step, n (%)
At a guideline test, n (%)
Received (%)
Change at a guideline test
Change at other actions
Appendicitis
740
179 (24)
104 (58)
98
+0.11 (155)
+0.03 (238)
Cholecystitis
645
279 (43)
171 (61)
85
+0.15 (119)
+0.04 (376)
Bowel obstruction
335
119 (36)
114 (96)
99
+0.15 (84)
-0.01 (221)
Diverticulitis
270
176 (65)
135 (77)
96
+0.43 (53)
+0.04 (105)
Pancreatitis
260
102 (39)
97 (95)
71
+0.24 (59)
+0.01 (148)
Pneumonia
105
37 (35)
28 (76)
79
+0.12 (19)
+0.00 (123)
Supplementary Table 6: The decisive step and the guideline test by disease. Decision2Care’s 5 runs on the 546 paired admissions (2,730 case-runs). Decisive step: the action with the largest rise in the probability of the true disease when that rise is at least 0.20, as a share of the disease’s case-runs; at a guideline test, as a share of the decisive steps. Received: case-runs with at least 1 guideline test of the disease. Change: the mean rise in the probability of the true disease per step, with steps per run in parentheses.
Domain
Disease
Decision2Care (this work)
MIRA-v2 + GPT-OSS-120B
MIRA-v2 + GLM-5
MIRA-v2 + Claude Sonnet 5
Physical examination
Overall
96.4 (94.5, 97.7)
74.8 (71.0, 78.3)
100.0 (99.3, 100.0)
100.0 (99.3, 100.0)
Physical examination
Appendicitis
98.6 (95.2, 99.6)
90.4 (84.6, 94.2)
100.0 (97.5, 100.0)
100.0 (97.5, 100.0)
Physical examination
Cholecystitis
95.8 (90.8, 98.1)
49.9 (41.4, 58.4)
100.0 (97.1, 100.0)
100.0 (97.1, 100.0)
Physical examination
Diverticulitis
98.1 (90.2, 99.7)
85.9 (74.3, 92.8)
100.0 (93.4, 100.0)
100.0 (93.4, 100.0)
Physical examination
Pancreatitis
90.0 (78.9, 95.6)
51.2 (38.0, 64.2)
100.0 (93.1, 100.0)
100.0 (93.1, 100.0)
Physical examination
Pneumonia
100.0 (84.5, 100.0)
81.0 (60.0, 92.3)
100.0 (84.5, 100.0)
100.0 (84.5, 100.0)
Supplementary Table 7: Test selection against the record. The share of each admission’s recorded blood tests, microbiology tests and imaging studies that each system requested on the 546 paired cases, pooled over cases and averaged over 5 runs, and the share of cases in which the physical examination was requested, with Wilson 95% intervals. Tversky distance per case: 1−both/(both+2×extra+missed) , 0 when the requests equal the record.
System
Rule
Retained, n (%)
Retained accuracy (%)
Retained errors / all
Decision2Care
confidence ≥ 0.70 (Supplementary Fig. 6 b)
269 (49.3)
98.5 [96.2, 99.4]
4 / 71
Decision2Care
confidence ≥ 0.70, each of the 5 runs
269–273 (49.3–50.0)
97.8–98.5
4–6 / 71–77
Decision2Care
confidence ≥ 0.80, set on the 325 other admissions
148 (27.1)
99.3 [96.3, 99.9]
1 / 71
MIRA-v2 + GPT-OSS-120B
ConsistencyDx ≥ 0.90, the published threshold (Supplementary Fig. 6 d, f, h)
300 (54.9)
96.0 [93.1, 97.7]
12 / 69
MIRA-v2 + GPT-OSS-120B
ConsistencyDx ≥ 0.95
205 (37.5)
97.6 [94.4, 99.0]
5 / 69
MIRA-v2 + GLM-5
ConsistencyDx ≥ 0.90, the published threshold (Supplementary Fig. 6 d, f, h)
343 (62.8)
98.0 [95.8, 99.0]
7 / 53
Supplementary Table 8: Operating points of the 4 systems on the 546 paired admissions. Retained: cases not deferred for review. Coverage is their share of all cases; retained accuracy is the share diagnosed correctly, with a Wilson 95% interval, and retained errors are reported over all errors. Decision2Care: its confidence in its diagnosis, read once after the work-up, baseline run unless the row says otherwise. MIRA-v2: ConsistencyDx, the agreement of the diagnoses of its 5 runs [ 9 ] , a case counted as correct when the majority of its runs are correct. At the published threshold of 0.90, MIRA-v2 retained more cases than Decision2Care did at 0.70 (exact McNemar tests, p=0.016 with GPT-OSS-120B and p<0.001 with GLM-5 and with Claude Sonnet 5). Below the rules, the highest-confidence cases for each system at 2 coverage levels, the largest set each signal retains at a given number of errors, and Decision2Care’s retained cases by disease. ConsistencyDx computed on Decision2Care’s own 5 runs is at chance (area under the receiver operating characteristic curve [AUROC], 0.51): 451 of the 546 cases had the same diagnosis in all 5 runs.
Condition
Accuracy, all cases (%)
Coverage (%)
Retained accuracy (%)
Retained errors
AUROC
Mean confidence
As recorded
86.5 (85.9–87.0)
49.7 (49.3–50.0)
98.4 (97.8–98.5)
4.4 (4.0–6.0)
0.834 (0.831–0.839)
0.650 (0.649–0.652)
History withheld
59.4 (58.6–60.3)
35.1 (34.2–36.4)
97.5 (97.3–98.0)
4.8 (4.0–5.0)
0.910 (0.908–0.912)
0.538 (0.535–0.542)
Decisive results made normal
22.2 (22.0–22.5)
4.2 (3.7–4.6)
84.3 (80.0–87.0)
3.6 (3.0–5.0)
0.569 (0.555–0.580)
0.379 (0.376–0.383)
Supplementary Table 9: Decision2Care’s confidence after work-ups with the history withheld or the decisive results made normal. Results from 5 runs of each condition on the 546 paired admissions (Methods), the mean with the lowest and highest run; as recorded, the 5 runs of the Results. Accuracy over all cases by the strict rule; the share retained at a confidence of at least 0.70 and accuracy among retained cases, with the errors among them; the AUROC of the confidence for a correct diagnosis; and its mean.
Disease group
Treatment
n
Decision2Care
MIRA-v2 + GPT-OSS-120B
MIRA-v2 + GLM-5
MIRA-v2 + Claude Sonnet 5
Documented care
Appendicitis
Antibiotics
145
99.3 [97.2, 100.0]
89.2 [87.1, 91.3]
93.5 [91.0, 95.9]
100.0 [100.0, 100.0]
84.7*** [79.0, 90.3]
Appendicitis
Supportive care
145
100.0 [100.0, 100.0]
89.1 [87.0, 91.2]
69.1*** [65.3, 73.1]
90.9 [88.6, 93.1]
100.0 [100.0, 100.0]
Appendicitis
Appendectomy
134
100.0 [100.0, 100.0]
97.5 [96.0, 98.8]
100.0 [100.0, 100.0]
99.9 [99.6, 100.0]
–
Cholecystitis
Antibiotics
91
97.8 [94.5, 100.0]
89.5 [86.2, 92.6]
97.4 [95.1, 99.2]
98.4 [95.9, 100.0]
93.3 [87.8, 97.8]
Cholecystitis
Supportive care
91
100.0 [100.0, 100.0]
89.2 [85.8, 92.5]
72.8*** [66.8, 78.4]
99.1 [98.0, 100.0]
98.9 [96.6, 100.0]
Cholecystitis
Cholecystectomy
83
100.0 [100.0, 100.0]
66.5*** [61.2, 71.2]
100.0 [100.0, 100.0]
99.5 [98.8, 100.0]
–
Supplementary Table 10: Treatments ordered when indicated. Per disease group and treatment, the number of admissions whose rubric indicated it (Methods) and the share in which each system ordered it, a model’s runs averaged per admission, with 95% bootstrap intervals: above the rule, the 437 of the 546 paired admissions that every system diagnosed correctly, correctly diagnosed runs only (Supplementary Fig. 7 a); below it, all 546 admissions and every run. Documented care: the first day’s prescriptions and the emergency department’s dispensing, for the drug items and for cholecystectomy in biliary pancreatitis. Stars: a model, or the documented care, against Decision2Care by an exact McNemar test, Benjamini–Hochberg corrected; * q<0.05 , ** q<0.01 , *** q<0.001 .
Disease
Category
n
Decision2Care
MIRA-v2 + GPT-OSS-120B
MIRA-v2 + GLM-5
MIRA-v2 + Claude Sonnet 5
Documented care
All, per patient
mean over the categories
355
88.2 [86.0, 90.3]
65.6*** [63.9, 67.3]
58.7*** [57.0, 60.4]
79.7*** [77.6, 81.7]
81.4*** [78.8, 83.9]
Urinary tract infection
antibiotic
10
70.0 [40.0, 100.0]
69.0 [47.0, 89.0]
90.0 [70.0, 100.0]
80.0 [58.0, 100.0]
90.0 [70.0, 100.0]
Cholecystitis
analgesic
91
100.0 [100.0, 100.0]
89.2 [85.9, 92.5]
70.0*** [64.1, 75.6]
98.7 [97.4, 99.8]
98.9 [96.6, 100.0]
Cholecystitis
antibiotic
91
91.2 [84.6, 96.7]
94.6* [92.6, 96.5]
91.8 [88.2, 95.2]
98.1 [95.4, 100.0]
93.3 [87.8, 97.8]
Cholecystitis
intravenous fluids
91
100.0 [100.0, 100.0]
13.7*** [10.1, 17.6]
4.5*** [2.3, 7.0]
71.8*** [65.7, 77.7]
97.8 [94.4, 100.0]
Diverticulitis
analgesic
49
28.6 [16.3, 40.8]
7.8** [2.4, 13.9]
7.8*** [4.1, 12.2]
5.3** [1.2, 10.2]
36.7 [24.5, 51.0]
Supplementary Table 11: Guideline adherence of the drug orders by disease and category. The share of admissions whose orders the evaluator of Ferber et al. [ 8 ] judged adherent (Methods), on the 355 of the 437 admissions every system diagnosed correctly that have one of its 7 diseases, a model’s correctly diagnosed runs averaged per admission, with 95% bootstrap intervals (Supplementary Fig. 7 b); the first row averages each patient’s categories. Documented care: the first day’s prescriptions and the emergency department’s dispensing, judged the same way, on 353 admissions. Stars: against Decision2Care, an exact McNemar test for the categories, Benjamini–Hochberg corrected (* q<0.05 , ** q<0.01 , *** q<0.001 ), and a Wilcoxon signed-rank test for the first row, uncorrected (the same thresholds for p ).
Measure
Decision2Care
MIRA-v2 + GPT-OSS-120B
MIRA-v2 + GLM-5
MIRA-v2 + Claude Sonnet 5
Drug classes against those given on the first day or in the emergency department (435 admissions with care recorded; mean per admission)
F1
40.3 [39.1, 41.6]
34.7*** [33.4, 36.1]
35.2*** [33.8, 36.7]
40.2 [39.0, 41.5]
Precision: classes ordered that were given (%)
46.8 [45.2, 48.4]
49.5** [47.6, 51.5]
57.0*** [54.7, 59.4]
54.2*** [52.3, 56.1]
Recall: classes given that were ordered (%)
37.2 [35.9, 38.5]
30.1*** [28.8, 31.3]
27.4*** [26.2, 28.7]
33.6*** [32.4, 34.8]
Classes ordered per admission (given: 8.9)
6.90 [6.70, 7.11]
5.61*** [5.37, 5.85]
4.37*** [4.15, 4.59]
5.56*** [5.35, 5.78]
Antibiotic decision as given (%)
81.8 [78.0, 85.3]
79.3 [76.4, 82.2]
83.5* [80.3, 86.6]
84.7 [81.6, 88.0]
Supplementary Table 12: Drug classes, anticoagulation and procedures against the admission’s record. The 437 of the 546 paired admissions that every system diagnosed correctly; a model’s correctly diagnosed runs averaged per admission; 95% bootstrap intervals. Drug classes: a system’s orders against the classes given on the first day or in the emergency department (Methods), and with the 2 heparins counted as 1 class (Supplementary Fig. 7 c). Anticoagulation: against the first day’s orders at treatment dose and at prophylactic dose. Procedures: the measures of Ferber et al. [ 8 ] , pooled and then by disease; last, Decision2Care’s requests as the probability required for a procedure group rises above 0.5. Stars: a model against Decision2Care, Wilcoxon signed-rank test on the per-admission values, the antibiotic decision and anticoagulation by an exact McNemar test, uncorrected; * p<0.05 , ** p<0.01 , *** p<0.001 . F1 is the harmonic mean of precision and recall.
Disease
Database
Patients (30/90 d)
Deaths (30/90 d)
AUROC, 30 d, mean (range)
AUROC, 90 d, mean (range)
Intracerebral haemorrhage
MIMIC-III
316 / 317
131 / 140
0.864 (0.862–0.865)
0.854 (0.851–0.857)
Intracerebral haemorrhage
MIMIC-IV
652 / 653
220 / 242
0.895 (0.893–0.899)
0.879 (0.877–0.880)
Traumatic brain injury
MIMIC-III
1,528
271 / 319
0.871 (0.870–0.873)
0.855 (0.855–0.856)
Traumatic brain injury
MIMIC-IV
2,481
334 / 421
0.891 (0.889–0.893)
0.856 (0.854–0.858)
Sepsis
MIMIC-III
614 / 615
164 / 225
0.779 (0.774–0.783)
0.762 (0.759–0.766)
Sepsis
MIMIC-IV
335 / 336
81 / 119
0.768 (0.757–0.776)
0.717 (0.709–0.726)
Supplementary Table 13: The prediction of death at 8 h after arrival in the emergency department, by cohort. Results from 5 runs on each fixed cohort; the AUROC of the prediction for all-cause death within 30 and 90 days of the prediction, the mean with the lowest and highest run. Patients and deaths are those with a known outcome at each horizon. The pooled curves of Fig. 6 a,b concatenate the 8 disease cohorts from MIMIC-III and MIMIC-IV within each run. Each disease-task membership is retained when an admission belongs to multiple cohorts (Methods). Clinical screening defined the acute respiratory distress syndrome (ARDS) candidate cohorts. CAP, community-acquired pneumonia. The polytrauma cohorts are matched between the 2 databases. The last 2 rows are from INSPECT, a data set of Stanford Health Care (Methods; Supplementary Fig. 14 ), and are not part of the pooled curves; all other rows are from Beth Israel Deaconess Medical Center. Its pneumonia group consists of candidates whose community acquisition could not be verified. The 95% bootstrap intervals over patients at 30 and 90 days were 0.746–0.873 and 0.743–0.851 in pulmonary embolism, and 0.658–0.883 and 0.646–0.871 in the pneumonia candidates.
Sens., 30 d
Spec., 30 d
AUROC, 30 d (agent minus intensivist)
Unsure, 30 d
Sens., 90 d
Spec., 90 d
AUROC, 90 d (agent minus intensivist)
Unsure, 90 d
Intensivist 1
0.63
0.60
0.720 (-0.15 to +0.13)
51 (71%)
0.65
0.53
0.655 (-0.17 to +0.12)
55 (76%)
Intensivist 2
0.70
0.76
0.755 (-0.15 to +0.05)
37 (51%)
0.68
0.78
0.741 (-0.23 to -0.01)
37 (51%)
Intensivist 3
0.63
0.49
0.670 (-0.06 to +0.15)
15 (21%)
0.62
0.50
0.637 (-0.13 to +0.11)
15 (21%)
Mean of the 3
0.65
0.61
0.727 (-0.12 to +0.08)
103 (48%)
0.65
0.60
0.696 (-0.19 to +0.05)
107 (50%)
Agent, 1 run
0.707 (0.565–0.836)
0.628 (0.496–0.757)
Weighted back to the cohorts
Supplementary Table 14: Three intensivists’ predictions and the agent’s on the same 72 cases. The 72 review cases of Fig. 6 a,b, 7–10 from each of 8 diseases (35 from MIMIC-III and 37 from MIMIC-IV), drawn within disease and database from strata of 90-day mortality crossed with the correctness of the agent’s 90-day prediction (21 true positive, 19 false negative, 18 true negative and 14 false positive); 27 and 40 ended in death by 30 and 90 days of the prediction made 8 h after arrival. Wrong agent predictions of 90-day mortality were thus over-represented (33 of 72, 46%, compared with 2,247 of the 8,570 cohort entries, 26%), so the AUROCs of the review cases are lower than those of the cohorts (0.871 and 0.849). The intensivists were given every record offered to the agent (a median of 124 per case), of which the agent had read a median of 54%. Each intensivist predicted the interval of death (within 7 days, days 8–30, days 31–90, alive at 90 days) and the probability of death by 30 and by 90 days and rated their certainty (1, sure, to 4, unsure) before seeing the agent’s work-up (Methods); sensitivity (Sens.) and specificity (Spec.) are those of the interval prediction for death by each horizon, the crosses of Fig. 6 a,b; the AUROC is that of the probability, with the 95% interval of the agent’s AUROC minus the intensivist’s from 2,000 paired case-bootstrap resamples; unsure, the predictions rated rather unsure or unsure. The agent’s AUROC comes from 1 run on the same cases, with its 95% bootstrap interval. In the weighted rows each case is weighted by the number of cohort entries in its stratum divided by the number of review cases drawn from that stratum (effective sample size 14.6; 77 cohort entries lay in 2 strata without a review case), and the intervals come from 2,000 case-bootstrap resamples with the weights fixed (Methods). Of the 33 cases in which the agent’s 90-day prediction was wrong, the interval predictions of at least 2 of the 3 intensivists were also wrong in 20 and those of all 3 in 10.
Reviews
Essential record unread
Step unnecessary
Prediction at least defensible
Prediction plausible
Plausibility (mean)
Interval agreed
Intensivist 1
72
5 (7%)
55 (76%)
67 (93%)
44 (61%)
3.90
50 (69%)
Intensivist 2
72
12 (17%)
27 (38%)
50 (69%)
30 (42%)
3.21
40 (56%)
Intensivist 3
72
3 (4%)
40 (56%)
65 (90%)
51 (71%)
4.10
53 (74%)
Intracerebral haemorrhage
30
0 (0%)
16 (53%)
27 (90%)
19 (63%)
3.97
21 (70%)
Traumatic brain injury
30
0 (0%)
8 (27%)
25 (83%)
18 (60%)
3.87
18 (60%)
Sepsis
30
8 (27%)
16 (53%)
26 (87%)
15 (50%)
3.60
18 (60%)
Supplementary Table 15: The intensivists’ review of the agent’s work-ups and predictions. After making their own prediction for each of the 72 cases, the 3 intensivists were shown the records the agent retrieved, in order, and its prediction, and recorded whether an essential record had been left unread, whether a step had been unnecessary, how plausible the prediction was on a scale of 1 to 5 and whether they agreed with the interval it predicted (Methods). On the scale, 1 was labelled ‘implausible given the data’, 3 ‘defensible, I would differ’ and 5 ‘fully consistent with the data’; at least defensible, a rating of 3 or more; plausible, a rating of 4 or 5. Counts of reviews by intensivist (72 each), by disease (21 to 30) and in all (216).
Disease
Patients (deaths)
1 h
2 h
3 h
4 h
5 h
6 h
7 h
8 h
9 h
10 h
11 h
12 h
First hour
Intracerebral haemorrhage
774 (196)
0.678
0.787
0.825
0.839
0.845
0.854
0.842
0.862
0.863
0.871
0.875
0.874
2
Traumatic brain injury
2,790 (358)
0.724
0.790
0.819
0.835
0.847
0.854
0.857
0.865
0.868
0.866
0.871
0.872
2
Sepsis
851 (170)
0.611
0.659
0.698
0.688
0.698
0.720
0.701
0.724
0.726
0.744
0.734
0.750
none
Acute kidney injury
531 (159)
0.692
0.766
0.763
0.737
0.736
0.738
0.742
0.735
0.702
0.711
0.733
0.721
2
Pulmonary embolism
326 (29)
0.702
0.707
0.751
0.710
0.784
0.814
0.796
0.820
0.814
0.818
0.829
0.801
none
ARDS candidates
92 (24)
0.624
0.704
0.722
0.706
0.748
0.742
0.760
0.723
0.724
0.723
0.762
0.737
3
Supplementary Table 16: The prediction of death in hospital from the records of the first 1 to 12 h in the intensive care unit (ICU). AUROC by disease, patients from MIMIC-III and MIMIC-IV pooled, each hour an independent prediction from the records of the first h hours after admission to the ICU, in patients alive and in hospital at 24 h (Methods); 95% bootstrap intervals are drawn in Fig. 6 c and given for 12 h in Supplementary Table 17 . First hour, the earliest hour at which the AUROC reached the highest of the 3 24-h scores (a descriptive mark, not a test).
Disease
Score
Score at 24 h, AUROC (95% CI)
Prediction at 2 h
Difference at 2 h (95% CI)
Prediction at 12 h (95% CI)
Difference at 12 h (95% CI)
Intracerebral haemorrhage
APS III
0.631 (0.582–0.679)
0.787
+0.156 (+0.097, +0.214)
0.874 (0.846–0.899)
+0.244 (+0.190, +0.296)
SAPS II
0.713 (0.672–0.751)
+0.074 (+0.022, +0.127)
+0.161 (+0.120, +0.203)
OASIS
0.728 (0.687–0.766)
+0.059 (+0.011, +0.109)
+0.146 (+0.105, +0.188)
Traumatic brain injury
APS III
0.701 (0.670–0.733)
0.790
+0.089 (+0.053, +0.126)
0.872 (0.854–0.891)
+0.171 (+0.139, +0.205)
SAPS II
0.756 (0.730–0.781)
+0.034 (+0.004, +0.066)
+0.116 (+0.087, +0.147)
OASIS
0.763 (0.738–0.789)
+0.027 (-0.001, +0.058)
+0.109 (+0.084, +0.135)
Supplementary Table 17: The ICU prediction against APS III, SAPS II and OASIS. The AUROC of each score computed from the first 24 h in the same patients as Fig. 6 c, the AUROC of the prediction with 2 and 12 h of records, and the paired difference, prediction minus score, with 95% intervals from 2,000 patient bootstrap replicates stratified by database and outcome and shared by all hours and scores (Methods). APS III, Acute Physiology Score III. CI, confidence interval. OASIS, Oxford Acute Severity of Illness Score. SAPS II, Simplified Acute Physiology Score II.
Supplementary Fig. 1: Top- n accuracy of the final differential by disease. a – l , Strict disease-level accuracy for n=1 –10 on the 546 paired admissions, with the number of admissions per disease; lines, mean of 5 runs; shading, range of runs. Panels are ordered by Decision2Care’s mean lead over the strongest comparator across n . In cholecystitis ( f ), the disease was among the first 5 entries of the differential in 139 of Decision2Care’s 144 incorrect predictions (96.5%).
Supplementary Fig. 2: Work-up paths by disease. a – l , Mean transitions between information sources over Decision2Care’s 5 runs on the 546 paired admissions, with the number of admissions per disease. Black arrows follow the most frequent next unvisited source; grey arrows mark steps taken in fewer than 5% of cases; repeat requests are counted inside the boxes.
Supplementary Fig. 3: The differential over the work-up. Decision2Care’s 5 runs on the 546 paired admissions; lines, means; bands, range of runs. a , Share of case-runs with the true disease first in the differential (circles) and with a correct coded diagnosis (squares) after each action; numbers, cases still acting. Where the history alone did not put the true disease first, the differential favoured none of the 12 diseases in 17.3% of case-runs, another disease of the same organ system in 10.5% and a disease of another organ system in 0.7%. b , Cases per run whose top-ranked disease changed from incorrect to correct (blue) and from correct to incorrect (orange). c , By disease, the share of case-runs with a decisive step, a rise of at least 0.20 in the probability of the true disease (light), and the share whose decisive step was a guideline test (dark).
Supplementary Fig. 4: The diagnosis made again with one source left out of the chart. Decision2Care’s baseline run on the 546 paired admissions. a , Change in strict accuracy, in points, by disease (columns) and withheld source (rows), among cases containing that information source; cells require at least 5 cases, with the number printed below each value. b , Over all diseases, the cases correct only with the source (blue) and only without it (orange), the change in points, the number of cases and the exact McNemar test (* p<0.05 , ** p<0.01 , *** p<0.001 ).
Figure 29Figure 30Figure 31
Supplementary Fig. 5: Receiver operating characteristic (ROC) curves of the deferral signals. Each system’s signal against its own verdicts on the 546 paired admissions, with the area under the ROC curve (AUROC): Decision2Care’s confidence in its diagnosis, read once after the work-up in the baseline run, and MIRA-v2’s ConsistencyDx, the agreement of the diagnoses of its 5 runs, a case counted as correct when the majority of its runs are correct.
Supplementary Fig. 6: Deferral from 1 Decision2Care work-up and 5 MIRA-v2 work-ups. The 546 paired admissions: Decision2Care’s baseline run ( a , b ) and MIRA-v2 with GPT-OSS-120B ( c , d ), GLM-5 ( e , f ) and Claude Sonnet 5 ( g , h ). a , c , e , g , Accuracy among retained cases (solid) and coverage, the share retained without review (hatched) as the threshold on Decision2Care’s confidence or on ConsistencyDx [ 9 ] rises; a MIRA-v2 case is correct when most of its runs are. b , d , f , h , Simulated deferral at a confidence of at least 0.70 and at the published ConsistencyDx threshold of 0.90; counts in Supplementary Table 8 . Correct and error describe the agent’s diagnosis before review.
Supplementary Fig. 7: Treatment orders against the rubrics, the guidelines and the documented care. The 437 of the 546 paired admissions that every system diagnosed correctly; whiskers, 95% bootstrap intervals over the admissions. a , Share of admissions in which each system ordered a treatment that the disease’s rubric indicated, by disease group and item; grey, the care documented on the first day and in the emergency department (ED). b , Share of admissions whose drug orders the evaluator of Ferber et al. [ 8 ] judged guideline-adherent, by disease and category and, in the first row, averaged over each patient’s categories, on the 355 admissions with one of its 7 diseases; diamonds, documented care. c , Drug classes ordered against those given on the first day or in the ED, precision against recall per admission, with the F1 score (the harmonic mean of precision and recall) in the legend. Stars, a model against Decision2Care (exact McNemar test, Benjamini–Hochberg corrected; * q<0.05 , ** q<0.01 , *** q<0.001 ; the first row of b , Wilcoxon signed-rank test, uncorrected, with the same thresholds for p ).
Supplementary Fig. 8: Auditing emergency department (ED) mortality predictions along the work-up. ICH, TBI, sepsis and AKI cases at ED +8 h, with higher scores on the left and lower scores on the right. Curves show mean 30-day death scores and ranges of 3 queries along fixed retrieval paths; cards show returned evidence; heatmaps show pooled original action scores.
Supplementary Fig. 9: Mortality prediction audits in four additional disease cohorts. Examples from PE, the ARDS candidate cohort, pneumonia and polytrauma at ED +8 h; conventions as in Supplementary Fig. 8 .
Supplementary Fig. 10: Auditing intensive care unit (ICU) mortality predictions. ICH, TBI, sepsis and AKI at ICU +12 h: higher-score cases (left), lower-score cases (right). Curves show mean in-hospital death scores and ranges of 3 queries along fixed retrieval paths; cards show returned evidence; heatmaps show pooled original action scores.
Supplementary Fig. 11: ICU mortality audits in four additional disease cohorts. PE, the ARDS candidate cohort, pneumonia and polytrauma at ICU +12 h; conventions as in Supplementary Fig. 10 .
Supplementary Fig. 12: 30-day mortality predictions by disease. At 8 h after emergency department (ED) arrival, receiver operating characteristic (ROC) curves and areas under those curves (AUROCs) average 5 agent runs in the full MIMIC-III and MIMIC-IV disease cohorts. Crosses show the interval predictions of 3 intensivists (Readers 1–3) on 72 selected cases (7–10 per disease); dots, their mean. Stacked bars show their certainty ratings, pooled (All) and by intensivist.
Supplementary Fig. 13: 90-day mortality predictions by disease. At 8 h after emergency department (ED) arrival, receiver operating characteristic (ROC) curves and areas under those curves (AUROCs) average 5 agent runs in the full MIMIC-III and MIMIC-IV disease cohorts. Crosses show the interval predictions of 3 intensivists (Readers 1–3) on 72 selected cases (7–10 per disease); dots, their mean. Stacked bars show their certainty ratings, pooled (All) and by intensivist.
Supplementary Fig. 14: Mortality predictions at Stanford Health Care. a – d , Receiver operating characteristic (ROC) curves of the prediction made 8 h after arrival in the emergency department for death within 30 days ( a , c ) and 90 days ( b , d ), in pulmonary embolism (PE; a , b ) and pneumonia ( c , d ). Green, the patients of INSPECT, a data set of Stanford Health Care, whose number and deaths are given above each panel; grey, the cohorts of MIMIC-III and MIMIC-IV, from Beth Israel Deaconess Medical Center. Each curve is the mean of 5 runs, and the band is the range of the 5 runs in INSPECT. The key gives the area under each curve. In INSPECT the pneumonia group consists of candidates whose community acquisition could not be verified.
Supplementary Fig. 15: The records the intensivists marked essential and the records the agent read. a , Share of the 216 assessments of the 72 review cases (3 intensivists, before seeing the agent’s work-up) that marked each kind of record essential, by disease; dash, not on that disease’s checklist. b , Share of the cases with an available record of each kind at 8 h after arrival in which the agent read one, in the full MIMIC-III and MIMIC-IV cohorts of the 8 diseases (8,578 cohort entries, n above each column; 1 run); NA, no case had an available record of that kind.
Clinical decision-making in emergency medicine demands rapid, accurate diagnoses under uncertainty. Despite benchmark progress, evidence for LLMs as interactive aids in live physician workflows remains sparse. MedSyn lets physicians iteratively query an LLM provided with the full clinical record while initially viewing only the chief complaint. Seven physicians (three seniors, four residents) completed baseline and AI-assisted sessions across 52 MIMIC-IV cases stratified by difficulty. Blinded evaluation showed residents' Hard-case correctness rose from 0.589 to 0.734; difficulty-standardised completely-correct rates confirmed a medium effect (Δ = 0.092; p = 0.071; d = 0.47). Automated metrics corroborated these gains: standardised any-match accuracy improved by 0.156 (p < 0.0001), and residents showed the largest F1 gain (Δ = 0.138; p < 0.0001). Dialogue analysis revealed expertise-dependent strategies (seniors asked targeted, hypothesis-driven questions; residents relied on broader queries) and cross-expertise concordance increased (Δ = 0.145; p < 0.0001). Interactive LLM support meaningfully enhances diagnostic reasoning.
Burcu Sayin, Ngoc Vo Hong, Ipek Baris Schlicht +8
Department of Information Engineering and Computer Science, University of Trento, Italy · Department of Medicine, Azienda Sanitaria Universitaria Integrata del Trentino (ASUIT), Trento, Italy · Universitat Politècnica de València, Spain +2
We present PULSE, a medical reasoning agent that combines a domain-tuned large language model with scientific literature retrieval to support diagnostic decision-making in complex real-world cases. To evaluate its capabilities, we curated a benchmark of 82 authentic endocrinology case reports encompassing a broad spectrum of disease types and incidence levels. In controlled experiments, we compared PULSE's performance against physicians with varying levels of expertise-from residents to senior specialists-and examined how AI assistance influenced human diagnostic reasoning. PULSE attained expert-competitive accuracy, outperforming residents and junior specialists while matching senior specialist performance at both Top@1 and Top@4 thresholds. Unlike physicians, whose accuracy declined with disease rarity, PULSE maintained stable performance across incidence tiers. The agent also exhibited adaptive reasoning, increasing output length with case difficulty in a manner analogous to the longer deliberation observed among expert clinicians. When used collaboratively, PULSE enabled physicians to correct initial errors and broaden diagnostic hypotheses, but also introduced risks of automation bias. The study explores both serial and concurrent collaboration workflows, revealing that PULSE offers robust support across common and rare presentations. These findings underscore both the promise and the limitations of language model-based agents in clinical diagnosis, and offer a framework for evaluating their role in real-world decision-making.
Zhongzhen Huang, Yan Ling, Hong Chen +7
Shanghai Jiao Tong University, Shanghai, China · Department of Endocrinology, Zhongshan Hospital, Fudan University, Shanghai, China · Department of Endocrinology, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China +3
Large language models applied to clinical prediction exhibit case-level heterogeneity: simple cases yield consistent outputs, while complex cases produce divergent predictions under minor prompt changes. Existing single-agent strategies sample from one role-conditioned distribution, and multi-agent frameworks use fixed roles with flat majority voting, discarding the diagnostic signal in disagreement. We propose CAMP (Case-Adaptive Multi-agent Panel), where an attending-physician agent dynamically assembles a specialist panel tailored to each case's diagnostic uncertainty. Each specialist evaluates candidates via three-valued voting (KEEP/REFUSE/NEUTRAL), enabling principled abstention outside one's expertise. A hybrid router directs each diagnosis through strong consensus, fallback to the attending physician's judgment, or evidence-based arbitration that weighs argument quality over vote counts. On diagnostic prediction and brief hospital course generation from MIMIC-IV across four LLM backbones, CAMP consistently outperforms strong baselines while consuming fewer tokens than most competing multi-agent methods, with voting records and arbitration traces offering transparent decision audits.
Yuxing Lu, Yushuhong Lin, Jason Zhang
Georgia Institute of Technology · Peking University