Documenting Trofinetide Treatment Phenotypes in Rett Syndrome with a Dual-LLM Consensus Pipeline
Authors: Daniel Palacios, Guantong Qi, Mikael Guzman Karlsson, Shawntel K. David, Mohammed Sameeruddin, Bernhard Suter, Zhandong Liu, Davut Pehlivan, Hu Chen
Organizations: Quantitative and Computational Biosciences, Baylor College of Medicine, Houston, Texas, 77030, USA · Department of Pediatrics, Baylor College of Medicine, Houston, Texas, 77030, USA · Jan and Dan Duncan Neurological Research Institute, Texas Children’s Hospital, Houston, Texas, 77030, USA · Data Science Center, Texas Children’s Hospital, Houston, Texas, 77030, USA · Genetics & Genomics Graduate Program, Baylor College of Medicine, Houston, Texas, 77030, USA · Department of Pathology and Immunology, Baylor College of Medicine, Houston, Texas, 77030, USA · Blue Bird Circle Rett Center, Texas Children’s Hospital, Houston, Texas, 77030, USA · Section of Pediatric Neurology and Developmental Neuroscience, Department of Pediatrics, Baylor College of Medicine, Houston, Texas, 77030, USA
Objective: Trofinetide, approved in 2023 as the first Rett syndrome treatment, lacks scalable real-world monitoring because chart review is labor-intensive. We developed a dual large language model (LLM) consensus pipeline to extract 18 treatment-response phenotypes, comprising 4 adverse effects, 13 Rett clinical features, and medication refusal, from clinical notes. Materials and Methods: We analyzed 26,763 notes from 108 patients with confirmed sustained trofinetide exposure at Texas Children's Hospital. The pipeline filters notes by keyword, classifies each through two LLMs requiring consensus, and aggregates a patient-by-timepoint matrix. Architecture and prompt were selected on a diarrhea development set, then frozen and applied to the other 17 phenotypes with the phenotype definition substituted, as an out-of-sample test. The reported prompt comparison scores 183 expert-labeled cells. Results: A prompt requiring diarrhea to be active and severe performed best (84.2 ± 0.0% accuracy; F1 = 0.674), 47 percentage points above regular-expression matching (p<0.0001). Expert reviewers judged six of the pipeline's calls per phenotype, three positive and three negative, and agreed with 90.2% (92/102). The 13 Rett clinical features are manifestations of the disorder itself rather than adverse effects of trofinetide. Gross and fine motor impairment, language/communication, and stereotypy were documented present in 92-100% of patients at every visit. Discussion: In-context LLM extraction captured multi-domain treatment responses. The outcome is documented presence at each visit, not a direction-of-change judgment. Conclusion: Coupling in-context LLM reasoning with a consensus architecture offers a scalable alternative to keyword and regular-expression extraction for real-world evidence, adapting to new phenotypes and medications through configuration alone.
Figures & tables
Characteristic
Full cohort
Validation subset
p
( n=108 )
( n=56 )
Age at trofinetide start, years
12.7 (8.2–20.3)
12.6 (8.3–18.5)
0.72
Female, n (%)
106 (98.1)
56 (100.0)
0.23
Follow-up duration, months
29.4 (20.0–34.1)
32.4 (25.3–34.1)
0.06
Notes analyzed per patient, median (IQR)
148 (49–349)
171 (61–397)
0.07
Notes analyzed per patient, mean
270
271
Table 1: Cohort characteristics for the full QC-confirmed cohort and the diarrhea-validation subset. Values are median (IQR) or n (%). The p -column compares the validation subset against the remaining QC-confirmed patients, Mann–Whitney U for continuous variables and Fisher exact for sex. Variable definitions, and the patient count each row rests on, are in Supplementary Note S8.
Method
Type
What counts as diarrhea
Accuracy (%)
Precision
Recall
F1
Regular expression
Regex
Any note containing a diarrhea keyword
37.2
0.282
0.936
0.434
Negation-aware NLP
NLP
The same keywords, with negated and historical mentions filtered out
42.1
0.288
0.851
0.430
Any diarrhea, with rules
LLM
Any documented diarrhea, unless the note reports it as resolved or attributes it to another cause
75.0 ± 0.3
0.507
0.766
0.610
Any diarrhea
LLM
Any documented diarrhea, including diarrhea that treatment has brought under control
75.2 ± 0.3
0.514
0.787
0.622
Uncontrolled diarrhea
LLM
Diarrhea unless the note describes it as mild or controlled without needing intervention
79.6 ± 0.3
0.596
0.660
0.626
Severe diarrhea, with rules
LLM
Active diarrhea causing concern or prompting a management change, unless the note reports it as resolved or attributes it to another cause
83.1 ± 0.5
0.700
0.596
0.644
Table 2: Method comparison on the diarrhea phenotype. All seven methods are scored on the identical 183-cell expert ground truth from the 56 labeled patients. Type is the kind of system doing the reading, either lexical pattern matching, rule-based NLP, or a large language model. The five LLM rows share the same two agents, the same consensus gate, and the same model (Claude Opus 4.6), and differ only in what they count as diarrhea. All five also exclude resolved, negated, templated, anticipatory and other-cause mentions, so the two with rules variants restate that guidance rather than introduce it. Accuracy is the mean ± SD over three independent runs, and the two non-LLM methods are deterministic. The other metrics are computed from the modal prediction across the three runs. The bolded row is the configuration reported throughout the paper. Full prompt wording, and the identifiers these names carry in the released code, are in Supplementary Notes S3 and S7. Best value in each column in bold.
Anxiety, Bruxism, General Mood, Other Side Effects, Sleep, Weight Loss
4/6 (67%)
2
Dysautonomia, Refusing Medication
Table 3: Expert spot-check agreement for the 17 phenotypes without a ground truth. For each phenotype the Rett Center clinical team judged six model calls (three model-positive and three model-negative) as agreeing or disagreeing with the note (102 cases in total); overall model–reviewer agreement was 90.2% (92/102; 95% CI 82.9–94.6%). Diarrhea’s six cases were omitted from this spot-check because diarrhea is validated against a quantitative ground truth and characterized by a dedicated error analysis.
Mechanism
What happens
N (%)
Model reading defensible?
Threshold disagreement
Note documents active or medication-managed diarrhea. The reviewer scores No under a stricter severity threshold, or the converse, where the reviewer counts mild or managed changes the model discounts
14 (48%)
Arguable
Temporal displacement
The supporting positive note exists but falls in an adjacent six-month window, so the model and the reviewer are describing the same episode at different visits
12 (41%)
Arguable
Evidence absent
Diarrhea recorded in the ground truth but not present in any keyword-passing note in the window, having been reported by phone, through the patient portal, or only retrospectively
3 (10%)
N/A (no evidence)
Table 4: Error taxonomy for the diarrhea phenotype under the reported configuration (29 total errors on 183 cells, 17 false negatives and 12 false positives). “Model reading defensible?” indicates whether the model’s call is defensible against the notes available in the window it scored. Assignment is rule-based, so it separates the window a piece of evidence falls in but not whether a note was misread within its own window.
Figure 1: Study design, cohort, corpus, and phenotype panel. (A) The five-stage extraction pipeline, from keyword filtering through independent classification by two LLMs, the consensus gate, aggregation into a patient-by-timepoint matrix, and evaluation against ground truth. (B) The cohort as an icon array, one square per patient, showing the 115 patients screened, the 108 with QC-confirmed trofinetide exposure, and the 56 of those who formed the expert-annotated diarrhea ground truth. (C) Composition of the 91,453 retrieved notes by note type, with the top seven shown and 60 further types collapsed into “Other”. (D) The 18 extracted phenotypes, grouped into adverse effects, Rett clinical features, and medication refusal. Definitions are in Methods and the keyword lists in Supplementary Table S2.
Figure 2: Diarrhea validation against the expert ground truth, 183 evaluable patient-timepoint cells from the 56 labeled patients, all panels on Claude Opus 4.6. (A) Accuracy (bars, error bars = SD over three runs) and F1 (line) for the seven methods of Table 2 , ordered by accuracy, with the two non-LLM methods shaded apart from the five prompts. (B) Confusion structure at each timepoint, stacked true and false positives and negatives, with overall accuracy as a dashed line. A patient contributes at most one cell per timepoint, so each bar counts patients. (C) Expert and predicted diarrhea prevalence at each timepoint, the shaded band showing the difference and n the evaluable patients. (D) The 29 errors of the reported configuration grouped by mechanism, with the mechanisms defined in Table 4 . Panels B to D use the reported configuration, whose three runs are identical on every cell. Statistical tests for panel A are in Supplementary Table S4.
Figure 3: Multi-phenotype documented presence across treatment. (A) The percentage of patients with each phenotype documented present at each visit, from Claude Opus 4.6, on one shared 0–100% scale. Denominators are the anchor-confirmed patients with at least one keyword-matching note in that window, so they vary by phenotype and by visit and are listed in Supplementary Table S8. The baseline medication-refusal cell rests on a single patient and is suppressed ( † ). Dashed lines separate adverse effects, medication refusal, and Rett clinical features, and that last block is ordered by mean documented presence. Each visit is a separate cell and not a point on a connected line, so the panel shows documentation rates rather than a trajectory of improvement or worsening, and the Rett clinical features are manifestations of the disorder rather than drug-related harm. Diarrhea is validated separately in Figure 2 and is not plotted here. (B) Expert spot-check agreement for the same 17 phenotypes, six cases each, three model-positive and three model-negative. Row labels in panel A are bold for the nine phenotypes reaching 6/6. Full per-visit rates are in Supplementary Table S8.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Item
TRIPOD-LLM description (abbrev.)
Where addressed
1
Title identifies the study as evaluating an LLM, the task, population, and outcome
Title
2
Abstract
Abstract
3a
Healthcare context/use case and rationale
Introduction
3b
Target population and intended use in the care pathway
Introduction; Methods (Study design)
4
Study objectives (development/validation stage)
Introduction
5a
Sources of data and rationale
Methods (Data sources)
Appendix
Table 8
Note Type
Notes
% Corpus
Keyword-passing
Classified
Telephone Encounter
20,760
22.7%
425
164
Progress Notes
20,683
22.6%
3,125
879
(Unspecified)
15,899
17.4%
381
177
Letter
8,584
9.4%
1,399
536
Nsg/Anc Progress Note
7,753
8.5%
86
36
Plan of Care
2,965
3.2%
11
2
Appendix
Table S1: Note-type distribution of the 91,453 notes retrieved for the 108 QC-confirmed patients, giving the top 10 types by volume with 57 further types collapsed. Corresponds to main-text Figure 1C. The last two columns give the diarrhea funnel, first the notes containing at least one term from the diarrhea keyword list, then the subset of those falling inside a patient’s evaluation window and therefore classified by both agents. Retrieval is broader than analysis, and the 26,763 notes falling inside an evaluation window are the analyzed subset counted in main-text Table 1. The 1,650 MR AVS Snapshot notes illustrate the distinction, contributing 157 keyword-passing notes but no classifications, because none falls inside a window.
Phenotype
Representative Keywords
Diarrhea
diarrhea, loose stool, watery stool, pasty stool, GI side effect
Vomiting
vomiting, emesis, retching, gagging, nausea
Weight Loss
weight loss, poor appetite, failure to thrive, eating less
hand wringing, hand mouthing, midline hand movements, stereotyped
Appendix
Table S2: Representative keywords for each phenotype category. The lists are abbreviated here, and the full lists contain 14–22 terms per phenotype and 24 for diarrhea.
Phenotype
N instances
Raw agreement
Cohen’s κ
Overall (17 phenotypes)
60,530
96.6%
0.930
Vomiting
2,959
98.3%
0.943
Aggressive Behavior
4,297
98.9%
0.932
Gross Motor
6,022
97.8%
0.931
Dysautonomia
2,699
96.6%
0.917
Weight Loss
1,905
95.6%
0.910
Appendix
Table S3: Inter-LLM agreement between LLM 1 (clinical NLP expert) and LLM 2 (GI/neurology specialist) for each phenotype (Claude Opus 4.6), given as Cohen’s κ and raw agreement over the note–phenotype instances that returned a valid response from both models. Table S9 gives the raw contingency counts, and Figure S1 plots these values. Sorted by κ . The diarrhea row uses the severe-diarrhea prompt V2 on its own validated 2,260-note corpus, a different prompt from the other 17, so it is listed separately and excluded from the overall row.
Configuration A
Configuration B
b
c
p
pHolm
Prompts (family of 7)
V2
V0 (as deployed)
9
1
0.021
0.149
V2
V3
2
0
0.500
0.500
V2
Agent 1 alone (V0)
13
3
0.021
0.149
V2
Agent 2 alone (V0)
10
2
0.039
0.193
V0
V1
14
6
0.115
0.346
Appendix
Table S4: Pairwise McNemar exact tests among the comparator configurations on the 183-cell diarrhea ground truth, computed on the modal prediction across three runs. b counts the cells where the first configuration is correct and the second wrong, c the reverse, and p is the two-sided exact p -value. pHolm is Holm-adjusted within the pre-declared family the comparison belongs to, which is of size 7 for the prompts, 1 for the alignment contrast and 2 for the rule-based floors, rather than across all comparisons. Significant results ( p<0.05 ) in bold. Only the two floors are separated from the reported prompt. No contrast within the prompt family clears Holm correction, which is why V2 is described as the top of a sweep rather than as a demonstrated improvement. Every test is cell-comparable, meaning both configurations are scored against the same annotation on the same windows.
Timepoint
N
Accuracy
TP
TN
FP
FN
Baseline
54
96.3%
0
52
2
0
6 months
55
67.3%
21
16
4
14
12 months
33
90.9%
5
25
2
1
18 months
22
81.8%
1
17
2
2
24 months
13
84.6%
2
9
2
0
30 months
6
100.0%
1
5
0
0
Appendix
Table S5: Per-timepoint performance of the reported diarrhea configuration (Claude Opus 4.6, prompt V2, severe diarrhea, under the dual-LLM consensus gate). All three runs returned identical predictions on all 183 cells, so these are the predictions themselves rather than a representative run, although the note-level calls behind them are not fully stable (Note S1). Baseline is not negative by construction, since two of the 54 baseline cells are false positives.
Timepoint
Acc
BalAcc
PPV
Sens
Spec
F1
NPV
MCC
Prev
Overall
.842
.775
.714
.638
.912
.674
.879
.571
.257
Baseline
.963
–
.000
–
.963
.000
1.00
–
.000
6 months
.673
.700
.840
.600
.800
.700
.533
.386
.636
12 months
.909
.880
.714
.833
.926
.769
.962
.716
.182
18 months
.818
.614
.333
.333
.895
.333
.895
.228
.136
24 months
.846
.909
.500
1.00
.818
.667
1.00
.640
.154
Appendix
Table S6: Additional per-timepoint metrics for the reported diarrhea configuration (Claude Opus 4.6, prompt V2 under the consensus gate), computed from the same confusion counts as Table S5 , namely 30 TP, 124 TN, 12 FP and 17 FN over 183 evaluable cells. BalAcc is balanced accuracy, PPV precision, Sens recall, Spec specificity, NPV negative predictive value, and MCC the Matthews correlation coefficient. Dashes mark metrics undefined at a timepoint, which at baseline is every metric requiring a positive gold case. The rows at 24 and 30 months rest on 13 and 6 cells and should be read as descriptive.
Phenotype
Base
6m
12m
18m
24m
30m
n range
Weight Loss
63.2
64.2
61.7
69.2
62.7
71.4
42–67
Vomiting
26.5
37.3
37.5
31.6
22.2
21.7
46–75
Other Side Effects
11.0
57.8
49.4
37.7
39.3
41.2
51–90
Refusing Medication
†
18.2
30.3
37.5
16.1
14.8
24–33
Appendix
Table S7: Adverse effects, given as the percentage of patients with the finding documented as present at each visit (Claude Opus 4.6). Base is the pre-treatment baseline. A patient contributes a cell at a visit only where a keyword-passing note falls inside that window, so denominators vary by phenotype and visit, and the last column gives the range of per-visit denominators across the six visits. † marks a cell suppressed because it rests on a single patient. Diarrhea is not listed here because it is scored by the severe-diarrhea prompt V2 against its own ground truth, with per-visit performance in Table S5 .
Phenotype
Base
6m
12m
18m
24m
30m
n range
Attention/Alertness
73.3
58.1
34.4
56.5
33.3
47.4
15–32
General Mood
32.1
31.5
35.0
30.3
38.3
40.7
54–89
Gross Motor
97.6
93.5
97.6
95.7
98.4
98.1
53–93
Fine Motor
98.8
95.5
98.7
94.0
96.6
95.9
49–89
Dysautonomia
74.7
81.2
83.3
76.1
75.0
75.0
48–85
Language/Communication
96.6
95.7
100.0
91.9
96.9
94.5
55–93
Appendix
Table S8: Rett clinical features, given as the percentage of patients with the finding documented as present at each visit (Claude Opus 4.6). These features are manifestations of Rett syndrome itself rather than adverse effects of trofinetide. Values reflect documented presence of the finding regardless of whether it is improving or worsening, and Base is the pre-treatment baseline. A patient contributes a cell at a visit only where a keyword-passing note falls inside that window, so denominators vary by phenotype and visit, and the last column gives the range of per-visit denominators across the six visits. Documented presence is lower at baseline than at 6 months for 5 of the 13 features, so baseline-to-follow-up differences in this table should not be interpreted as clinical change.
Phenotype
Both YES
LLM 1 only
LLM 2 only
Both NO
N
Presence prompt, 17 phenotypes
23,251
1,123
923
35,233
60,530
Diarrhea, V0 (deployed)
276
97.7
12.7
1,873.3
2,260
Diarrhea, V2 (reported)
183
25.3
20.3
2,031.3
2,260
Appendix
Table S9: Contingency table of LLM 1 (clinical NLP expert) and LLM 2 (domain specialist) calls (Claude Opus 4.6). The first row covers the 17 phenotypes scored by the presence prompt, restricted to the 60,530 of 60,537 note–phenotype matches that returned a valid response from both models, which is the same base Table S3 uses. The two diarrhea rows use the diarrhea-specific prompts on the identical 2,260-note corpus, so the difference between them isolates the effect of prompt wording on how often the two agents differ. Their cells are means over the three runs and are therefore not integers.
Variant
Change from the deployed prompt
Acc. (%)
Δ
Prec.
Rec.
pHolm
A. Prompt wording and model count
V0
Uncontrolled diarrhea. Base case, as deployed. Both agents exclude adequately controlled diarrhea.
79.6 ± 0.3
—
0.596
0.660
—
V1
Any diarrhea. Controlled diarrhea reclassified as present. “Being treated means it is present.”
75.2 ± 0.3
−4.4
0.514
0.787
0.35
V2
Severe diarrhea. Same exclusion, restated explicitly as a severity bar. “YES only if active AND clinically significant.”
84.2 ± 0.0
+4.6
0.714
0.638
0.15
V3
Severe diarrhea with rules. V2 plus two reasoning rules. A later resolution governs an earlier onset in the same note, and a finding attributed to another cause is absent.
83.1 ± 0.5
+3.5
0.700
0.596
0.35
V4
Any diarrhea with rules. V1 plus the same two reasoning rules as V3.
75.0 ± 0.3
−4.6
0.507
0.766
0.25
Appendix
Table S10: Prompt-level, variable-decomposition and temporal-alignment sensitivity analyses (Note S7). Every variant is scored on the same 183 annotated diarrhea cells, with the architecture, consensus gate, keyword pre-filter and output schema held fixed, so every Δ is a like-for-like comparison. V0 is the deployed prompt and serves as the base case, and V2 is the configuration reported in the main text. Accuracy in panels A and B is the mean ± SD of three independent runs, as in main-text Table 2, and the other metrics are computed from the modal prediction across those runs. Panel C is a deterministic transformation of the V0 modal predictions, so it was run once and its Δ is taken against those predictions. The p values are McNemar exact against the base case, Holm-adjusted within the family the comparison belongs to (Table S4 ), and are therefore not comparable across panels. Three readings follow. The severity threshold in panel A is the only element that moves accuracy materially. No decomposition rule in panel B beats the deployed prompt, and the best comes within 0.2 points of it. In panel C, S1 shifts only the model output and is therefore scored against the same annotation on the same cells, whereas S3 shifts the annotation as well and is positional, so its 85.2% is an upper bound and never a metric of the pipeline. S1 is not deployable either, since the shift presumes onset at the first on-treatment visit, and it is reported as a diagnostic of where the error sits. The panel B labels are the identifiers used in the released code, and the family varies the combination rule, the direction of the consensus gate and the within-window aggregation order, all as functions of the same two extracted variables and therefore at no additional inference cost. C1 and C2b, and C2 and C2_strictB, return identical predictions on all 183 cells.
A. Six methods on the 189-cell grid
Method
Accuracy (%)
Precision
Recall
F1
Regex-only (floor)
42.3
0.293
0.936
0.447
Rule-based NLP
70.9
0.423
0.468
0.444
Hybrid NLP
70.9
0.436
0.575
0.495
Single-LLM
72.7 ± 0.6
0.458 ± 0.010
0.532 ± 0.000
0.492 ± 0.006
Dual-LLM (permissive)
73.0 ± 0.5
0.464 ± 0.008
0.553 ± 0.000
0.505 ± 0.005
Appendix
Table S11: Architecture selection, reported on the numbers it was decided with (Note S2). Six methods of increasing sophistication were evaluated on the identical 189-cell diarrhea ground truth under the original note corpus and the order-date exposure anchor, and this comparison is what selected the two-agent consensus architecture that the reported pipeline holds fixed. These are not the reported metrics, and they are not comparable to main-text Table 2 or to Table S10 , both of which run on the refreshed corpus, the corrected anchor and a 183-cell grid. Panel A gives accuracy as the mean ± SD over three runs for the LLM methods, with the three deterministic methods exact, and the best value in each column in bold. Panel B gives all 15 pairwise McNemar exact tests, Holm-adjusted across the 15 comparisons, where b counts the cells on which the first method is correct and the second wrong. Only the contrasts against the regex floor are significant, so the architecture was adopted as the best-performing configuration rather than as a proven necessity. The final rerun confirms that reading by placing the gate within 0.4 points of the stronger single agent under the deployed prompt, differing on two of 183 cells (Note S1). The consensus accuracy here, 76.4%, is not the 77.1% that Note S8 reports for the deployed prompt on a 189-cell grid, because this comparison is an earlier stage of development that ran on a different note pull and a different set of evaluation windows.
List of suppfigures S1 Inter-LLM agreement supports the consensus-gate design (Claude Opus 4.6). (A) Cohen’s κ between the two agents for each phenotype, sorted descending. Every phenotype reaches the substantial band, so the bars share one color, and the dashed line is the overall κ=0.93 across the 60,530 instances both agents scored. Agreement is highest for discrete, observable events such as vomiting, aggressive behavior and gross-motor findings, and lowest for attention/alertness, the phenotype with the fewest instances (313), and for the open-ended other-side-effects category. The hatched bar is diarrhea, scored by the diarrhea-specific prompt on its own validated corpus rather than by the presence prompt, and shown for reference. (B) Consensus outcome composition over the same 60,530 matches. The gate retains only “both YES” (38.4%) as positive, and the 3.4% one-sided disagreements are conservatively classified negative.
Accurate extraction of Alzheimer's Disease and Related Dementias (ADRD) phenotypes from electronic health records (EHR) is critical for early-stage detection and disease staging. However, this information is usually embedded in unstructured textual data rather than tabular data, making it difficult to be extracted accurately. We therefore propose LLM-MINE, a Large Language Model-based phenotype mining framework for automatic extraction of ADRD phenotypes from clinical notes. Using two expert-defined phenotype lists, we evaluate the extracted phenotypes by examining their statistical significance across cohorts and their utility for unsupervised disease staging. Chi-square analyses confirm statistically significant phenotype differences across cohorts, with memory impairment being the strongest discriminator. Few-shot prompting with the combined phenotype lists achieves the best clustering performance (ARI=0.290, NMI=0.232), substantially outperforming biomedical NER and dictionary-based baselines. Our results demonstrate that LLM-based phenotype extraction is a promising tool for discovering clinically meaningful ADRD signals from unstructured notes.
Mingchen Shao, Yuzhang Xie, Carl Yang +1
Northwestern University, Evanston, IL · Emory University, Atlanta, GA
Hallucinations in medical large language models (LLMs) remain a safety-critical issue, particularly when available evidence is insufficient or conflicting. We study this problem in diabetic retinopathy (DR) decision settings and introduce RETINA-SAFE, an evidence-grounded benchmark aligned with retinal grading records, comprising 12,522 samples. RETINA-SAFE is organized into three evidence-relation tasks: E-Align (evidence-consistent), E-Conflict (evidence-conflicting), and E-Gap (evidence-insufficient). We further propose ECRT (Evidence-Conditioned Risk Triage), a two-stage white-box detection framework: Stage 1 performs Safe/Unsafe risk triage, and Stage 2 refines unsafe cases into contradiction-driven versus evidence-gap risks. ECRT leverages internal representation and logit shifts under CTX/NOCTX conditions, with class-balanced training for robust learning. Under evidence-grouped (not patient-disjoint) splits across multiple backbones, ECRT provides strong Stage-1 risk triage and explicit subtype attribution, improves Stage-1 balanced accuracy by +0.15 to +0.19 over external uncertainty and self-consistency baselines and by +0.02 to +0.07 over the strongest adapted supervised baseline, and consistently exceeds a single-stage white-box ablation on Stage-1 balanced accuracy. These findings support white-box internal signals grounded in retinal evidence as a practical route to interpretable medical LLM risk triage.
Zhe Yu, Wenpeng Xing, Meng Han
Binjiang Institute of Zhejiang University, Hangzhou, China · Communication University of Zhejiang, Hangzhou, China · Zhejiang University, Hangzhou, China
As large language models (LLMs) are increasingly used in clinical settings, it is critical to evaluate their reliability under realistic variation in clinical text. We study this question in clinical triage, comparing LLMs to practicing physicians under text perturbations that preserve the underlying clinical setting. We introduce a benchmark of over 6,000 clinical scenarios, 7,000 physician annotations, and 225,000 model responses. Using this benchmark, we make two key observations. First, LLMs are more likely than physicians to recommend unnecessary care at baseline, and this tendency increases under perturbed inputs. Further, we find that LLM recommendations are more sensitive to gender and tone perturbations than human recommendations. Together, these results demonstrate that LLMs can vary under clinically irrelevant textual changes, highlighting the need for deployment-oriented evaluations grounded in expert physician behavior.
Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science · Worcester Polytechnic Institute, Department of Computer Science