Toward Better Assessment of LLMs' Performance in Clinical Error Detection

Yifan Zhang; Rahmatollah Beheshti

cs.CLAug 17, 2026

Toward Better Assessment of LLMs' Performance in Clinical Error Detection

Authors: Yifan Zhang, Rahmatollah Beheshti

Organizations: University of Delaware, Newark, DE, USA

Abstract

Automated detection of errors in clinical documentation is a promising application of large language models (LLMs), yet decisions to deploy such models rest on benchmarks that evaluate each clinical note in isolation. Error-detection benchmarks are typically constructed by injecting errors into notes, such that each erroneous note has a natural counterpart. Aggregate discriminative metrics (e.g., balanced accuracy or F1) do not exploit this structure. We show that this omission is consequential. In particular, evaluating 15 diverse LLMs on 4 standardized clinical error-detection test sets across 3 languages, we find that 13 of 15 models fall below the level of random pairwise discrimination, even while achieving F1 scores that standard practice would read as moderate. We also observe that the underlying bias patterns differ across languages: the same model can default to "no error" on one language and over-flag errors on another. To diagnose where discrimination breaks down, we further introduce a procedure to score the evidence models cite in their outputs. We find that while models consistently locate error-relevant content, they fail to produce the corresponding correct verdict on the clean counterpart. Finally, we show that F1 and pairwise accuracy are driven in opposite directions by the same underlying bias, so that ranking models by F1 may systematically promote the weakest discriminators. For safety-critical clinical NLP applications, we advocate for supplementing aggregate metrics with paired evaluations in benchmark reporting. Code and analysis scripts are available at https://github.com/healthylaife/paired-clinical-eval.

Figures & tables

Figure 1: Paired clinical error detection. (a) Each clinical scenario produces an error-injected note and a correct note differing by one sentence. (b) A model’s predictions on the pair fall into four mutually exclusive outcomes.

Figure 2: Our framework for evaluating LLMs in clinical error detection tasks. Four paired clinical error-detection test sets across three languages, LLMs across three size tiers, and four prompt-decoding configurations are used. Each model’s structured per-note output is then assessed through three parallel evaluation layers: traditional pointwise metrics, pairwise Both-Correct Rate (BCR), and Evidence Contrastive Analysis (ECA).

Dataset	Samples	Error	Clean	Pairs	Language	Source
MS-Test	597	311	286	286	English	Ben Abacha et al. (2025)
MEB-EN	208	104	104	104	English	Ma et al. (2026)
MEB-CN	200	100	100	100	Chinese	Ma et al. (2026)
MRT-JA	295	190	105	190	Japanese	Iwase et al. (2025)

Table 1: Datasets used for inference and evaluation. Error and Clean are the sub-counts of notes with and without errors, respectively; Pairs is the number of matched (error, correct) tuples used for BCR evaluation, which can differ from

\min(\text{Error},\text{Clean})

when notes lack a counterpart or share counterparts across pairs.

Figure 3: Evidence Contrastive Analysis ( ECA ). (a) The two indicators, illustrated on a shared clinical scenario: TP localization asks whether the cited evidence on the error note (Streptococcus pneumoniae) overlaps the error sentence; FP evidence-hit asks the same question on the correct note (Haemophilus influenzae). (b) The five outcomes arising from combining the two indicators across a pair. Both-Hit (A) isolates a judgment failure: the model attends to the correct location in both notes, yet still cannot determine which version is incorrect. Neither-Hit (D) reflects full attention failure. TP-Only (B) and FP-Only (C) are partial-attention failures. Extraction-Fail (E) captures pairs with unparseable Evidence fields.

Model	MS-Test	MEB-EN	MEB-CN	MRT-JA	Mean	Sens	Spec
	BCR (%)					Marginal (%)
Qwen 3-32B	$22.6{\scriptstyle\pm 3.7}$	$30.8{\scriptstyle\pm 11.1}$	$32.0{\scriptstyle\pm 5.5}$	$26.6{\scriptstyle\pm 1.9}$	28.0	69.4	51.4
UltraMed-70B $\dagger$	$24.3{\scriptstyle\pm 9.8}$	$43.7{\scriptstyle\pm 2.2}$	$16.8{\scriptstyle\pm 10.0}$	$18.0{\scriptstyle\pm 5.7}$	25.7	47.0	73.5
MedGemma 27B $\dagger$	$20.3{\scriptstyle\pm 4.9}$	$29.1{\scriptstyle\pm 8.6}$	$14.6{\scriptstyle\pm 5.7}$	$28.5{\scriptstyle\pm 4.2}$	23.1	61.2	56.0
Llama 3.1-70B	$20.2{\scriptstyle\pm 3.6}$	$27.1{\scriptstyle\pm 15.6}$	$15.3{\scriptstyle\pm 2.8}$	$17.4{\scriptstyle\pm 7.0}$	20.0	63.6	51.4
UltraMed-8B $\dagger$	$19.6{\scriptstyle\pm 5.8}$	$28.5{\scriptstyle\pm 2.5}$	$18.4{\scriptstyle\pm 5.5}$	$13.5{\scriptstyle\pm 1.8}$	20.0	43.4	65.0

Table 2: Four-configuration mean BCR (%)

\pm

standard deviation by model and dataset. Sens and Spec are the model’s mean pointwise sensitivity and specificity, averaged over all 16 runs; the highest- BCR models pair moderate sensitivity with moderate specificity, while the lowest- BCR models collapse one marginal toward its extreme.

\dagger

Medical-domain model. Sorted by cross-dataset mean BCR .

Model	MS-Test	MEB-EN	MEB-CN	MRT-JA	Shift?
Llama 3.2-3B	.94	.88	.63	.89
Gemma 3-4B	.94	.98	.89	.93
Qwen 3-4B	.81	.91	.61	.80
MedGemma 4B $\dagger$	.60	.66	.50	.80	✓
MediPhi 4B $\dagger$	.84	.89	.85	.86
Phi-4-mini	.79	.89	.88	.81

Table 3: Mean error-flag rate per model–dataset combination. Pink cells indicate yes-bias (rate

\geq 0.6

); blue cells indicate no-bias (rate

<0.4

); unshaded cells are balanced (

0.4\leq

rate

<0.6

). A check in the Shift column marks models whose bias category changes across datasets.

\dagger

Medical-domain model.

Figure 4: The localization-judgment gap on MS-Test (four-configuration mean across 15 models). (a) TP localization per model; dashed line is the random baseline (12.6%). (b) ECA category composition per model, as percentage of Pred1 pairs; category definitions in Figure 3 .

Figure 5: Metric deception across 60 model–dataset entries (four-configuration means). (a) Shaded region indicates F1

\geq

0.6 with BCR

<

25%. (b) Actual BCR vs. expected BCR under statistical independence (Equation 2 ); the dashed line is the independence baseline (

R_{\text{independence}}=1

), below which all 60 entries fall.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Model (Reference)	HuggingFace ID	Family	Domain	Prec.
Llama 3.2-3B ( Grattafiori et al., 2024 )	meta-llama/Llama-3.2-3B-Instruct	Llama	General	bf16
Gemma 3-4B ( Gemma Team et al., 2025 )	google/gemma-3-4b-it	Gemma	General	bf16
Qwen 3-4B ( Yang et al., 2025 )	Qwen/Qwen3-4B-Instruct-2507	Qwen	General	bf16
MedGemma 4B ( Sellergren et al., 2025 )	google/medgemma-4b-it	Gemma	Medical	bf16
MediPhi 4B ( Corbeil et al., 2025 )	microsoft/MediPhi-Instruct	Phi	Medical	bf16
Phi-4-mini ( Microsoft et al., 2025 )	microsoft/Phi-4-mini-instruct	Phi	General	bf16

Appendix

Table 4: References, HuggingFace identifiers, family, domain, and loading precision for all 15 evaluated models. All models are instruction-tuned.

Model	Dataset	Config	parse-failure rate (%)
UltraMed-8B	MEB-CN	neut-gre	24.5
UltraMed-70B	MRT-JA	cons-sam	22.0
UltraMed-70B	MRT-JA	neut-sam	14.9
Llama 3.2-3B	MEB-CN	cons-sam	9.5
UltraMed-8B	MEB-CN	neut-sam	9.5
Llama 3.2-3B	MEB-CN	neut-sam	9.0

Appendix

Table 5: Runs exceeding 5% parse failure rate (9 of 240).

Model	Skip (%)	Random (%)	$\Delta$
Qwen 3-32B	28.0	28.0	0.0
UltraMed-70B	25.7	26.1	$-$ 0.4
MedGemma 27B	23.1	23.1	0.0
Llama 3.1-70B	20.0	20.1	$-$ 0.1
UltraMed-8B	20.0	20.3	$-$ 0.3
Qwen 3-4B	17.0	17.0	0.0

Appendix

Table 6: Skip vs. random-assignment BCR comparison (four-configuration mean, cross-dataset mean). All deltas

<

0.5 pp.

Encoder	Recall@1	Recall@3	MRR	Pearson $r$	Above baseline
S-PubMedBert (biomedical)	65.3%	73.8%	0.73	0.99	15/15
BGE-large (general)	67.1%	82.5%	0.77	0.98	15/15

Appendix

Table 7: Embedding-based re-scoring of TP localization on MS-Test Pred1 pairs (

N=9{,}428

judgments pooled over 15 models

\times

4 configurations). Recall@1 counts a hit only when the gold error sentence is the single most similar sentence to the cited evidence. Random baseline

=12.6\%

. Pearson

r

is the per-model correlation with the substring TP-localization rate.

Figure 6: BCR (%) by configuration. Rank ordering is largely stable; absolute values vary up to 36.2 pp.

Dataset	Error-flag $\to$ F1	Error-flag $\to$ BCR	F1 $\to$ BCR
MS-Test	$+$ 0.946	$-$ 0.855	$-$ 0.659
MEB-EN	$+$ 0.406	$-$ 0.737	$+$ 0.310
MEB-CN	$+$ 0.911	$-$ 0.530	$-$ 0.142
MRT-JA	$+$ 0.963	$-$ 0.480	$-$ 0.233
All ( $n=60$ )	$+$ 0.853	$-$ 0.488	$-$ 0.058

Appendix

Table 8: Per-dataset error-flag-rate mediation correlations (Pearson

r

n=15

models per dataset). Error-flag rate suppresses BCR across all datasets and strongly inflates F1 on three of four (weaker on MEB-EN); the direct F1– BCR correlation varies in sign across datasets and is near zero only when pooled.

Dataset	Top-3 by F1	Top-3 by BCR	Overlap
MS-Test	Gemma 3-4B, Llama 3.2-3B, Mistral 7B	UltraMed-70B, Qwen 3-32B, MedGemma 27B	0/3
MEB-EN	Qwen 3-32B, UltraMed-70B, Llama 3.1-70B	UltraMed-70B, Qwen 3-32B, MedGemma 27B	2/3
MEB-CN	Mistral 7B, Gemma 3-4B, MediPhi 4B	Qwen 3-32B, Qwen 3-4B, UltraMed-8B	0/3
MRT-JA	Gemma 3-27B, Gemma 3-4B, Llama 3.2-3B	MedGemma 27B, Qwen 3-32B, Qwen 3-4B	0/3

Appendix

Table 9: Top-3 models by F1 vs. top-3 by BCR (4-config mean). On 3 of 4 datasets, no model appears in both top-3 lists.

Dataset	Expected BCR	Actual BCR	Ratio	Below
MS-Test	21.9%	14.2%	0.685 $\times$	15/15
MEB-EN	21.7%	17.8%	0.844 $\times$	15/15
MEB-CN	23.5%	14.4%	0.621 $\times$	15/15
MRT-JA	20.3%	15.2%	0.765 $\times$	15/15
All	21.9%	15.4%	0.729 $\times$	60/60

Appendix

Table 10: Per-dataset independence ratios. Ratio is the mean of per-model independence ratios

R_{\text{independence}}

(Equation 3 ), not the quotient of the two column means. Actual BCR falls below the expected value (sensitivity

\times

specificity) for all 60 model–dataset entries.

Dataset	Model	BalAcc	F1	Prec	Rec	Spec	MCC	Flag%
MS-Test	Qwen 3-32B	.584	.564	.616	.521	.647	.169	44.1
	Llama 3.1-70B	.566	.375	.690	.257	.874	.166	19.4
	MedGemma 27B	.555	.300	.720	.190	.920	.159	13.7
	UltraMed-70B	.554	.262	.778	.158	.951	.177	10.6
	Llama 3.1-8B	.550	.539	.575	.506	.593	.100	45.9
	MedGemma 4B	.549	.482	.588	.408	.689	.101	36.2

Appendix

Table 11: Traditional metrics — Conservative-Greedy configuration (all 15 models

\times

4 datasets = 60 runs). Sorted by balanced accuracy within each dataset. Flag%(error-flag rate) is the fraction of notes flagged as containing an error.

Dataset	Model	BalAcc	F1	Prec	Rec	Spec	MCC	Flag%
MS-Test	Llama 3.1-70B	.588	.462	.688	.347	.829	.200	26.3
	UltraMed-70B	.570	.355	.730	.235	.906	.188	16.8
	MedGemma 27B	.569	.345	.737	.225	.913	.188	15.9
	Qwen 3-32B	.566	.543	.596	.498	.633	.132	43.6
	Gemma 3-27B	.555	.518	.588	.463	.647	.112	41.0
	Phi-4-mini	.542	.606	.555	.669	.416	.088	62.8

Appendix

Table 12: Traditional metrics — Conservative-Sampling configuration (all 15 models

\times

4 datasets = 60 runs). Sorted by balanced accuracy within each dataset. Flag%(error-flag rate) is the fraction of notes flagged as containing an error.

Dataset	Model	BalAcc	F1	Prec	Rec	Spec	MCC	Flag%
MS-Test	UltraMed-70B	.642	.650	.661	.640	.643	.283	50.4
	MedGemma 27B	.596	.690	.585	.839	.353	.221	74.7
	Llama 3.1-70B	.586	.669	.584	.784	.389	.188	70.2
	Qwen 3-32B	.561	.676	.559	.855	.266	.150	79.7
	UltraMed-8B	.541	.618	.552	.701	.381	.087	66.2
	Qwen 3-8B	.533	.672	.540	.891	.175	.094	85.9

Appendix

Table 13: Traditional metrics — Neutral-Greedy configuration (all 15 models

\times

4 datasets = 60 runs). Sorted by balanced accuracy within each dataset. Flag%(error-flag rate) is the fraction of notes flagged as containing an error.

Dataset	Model	BalAcc	F1	Prec	Rec	Spec	MCC	Flag%
MS-Test	UltraMed-70B	.621	.688	.611	.788	.455	.258	67.2
	MedGemma 27B	.568	.666	.567	.807	.329	.155	74.2
	UltraMed-8B	.557	.586	.577	.597	.517	.114	54.2
	Qwen 3-32B	.545	.666	.549	.846	.245	.113	80.2
	Llama 3.1-70B	.543	.675	.548	.878	.208	.116	83.7
	Qwen 3-8B	.538	.674	.543	.887	.189	.107	85.1

Appendix

Table 14: Traditional metrics — Neutral-Sampling configuration (all 15 models

\times

4 datasets = 60 runs). Sorted by balanced accuracy within each dataset. Flag%(error-flag rate) is the fraction of notes flagged as containing an error.

Dataset	Model	$n$	TP loc	FP hit	A	B	C	D	E
MS-Test	Qwen 3-32B	521	86.9	48.8	240	213	11	46	11
	Qwen 3-4B	818	81.4	49.9	383	286	24	101	24
	Qwen 3-8B	452	78.8	41.5	202	182	19	43	6
	Gemma 3-27B	668	75.7	43.7	273	221	23	133	18
	Gemma 3-4B	1033	75.2	58.0	507	269	90	107	60
	MedGemma 27B	311	73.3	42.9	145	92	11	53	10

Appendix

Table 15: Full ECA breakdown per model–dataset combination (totals pooled across the four configurations).

n

: Pred1 pairs categorized into A–E. TP loc and FP hit: four-configuration mean overlap rates (%). A–E: pooled category counts (Both-Hit / TP-Only / FP-Only / Neither-Hit / Extraction-Fail). Within each dataset, rows are sorted by TP loc. Every model–dataset combination had at least one configuration with

\geq 5

Pred1 pairs, so all 60 combinations are listed. Because TP loc is an unweighted four-configuration mean while A–E are pooled counts, TP loc and

(A+B)/n

can differ by several points when Pred1 counts are uneven across configurations.

Clinical Language Model Evaluation

Clinical language model evaluation measures whether AI systems are accurate, safe, fair, and useful across medical tasks and care settings.

See all 419 papers

Beyond Scalar Scores: Exploring LLM-based Metrics for Clinical Significance Evaluation in Radiology Reports
Jun 17, 2026Qingyu Lu, Ruochen Li, Liang Ding +3Clinical Language Model Evaluation Language Model Robustness
Nanyang Technological University · Technical University of Munich · Alibaba +2
MedRECT: A Bilingual Medical Reasoning Benchmark for Error Correction in Clinical Texts
Nov 1, 2025Naoto Iwase, Hiroki Okuyama, Junichiro IwasawaMultilingual Language Model Evaluation Healthcare
Preferred Networks, Inc., Tokyo, Japan · School of Medicine, Nagoya University, Nagoya, Japan
Importance of Prompt Optimisation for Error Detection in Medical Notes Using Language Models
Feb 25, 2026Craig Myles, Patrick Schrempf, David Harris-BirtillLanguage Model Error Detection Prompt Optimization
University of St Andrews, St Andrews, United Kingdom · Canon Medical Research Europe Ltd., Edinburgh, United Kingdom

Jun 17, 2026J/K move · Enter open · S save

Beyond Scalar Scores: Exploring LLM-based Metrics for Clinical Significance Evaluation in Radiology Reports

Qingyu Lu, Ruochen Li, Liang Ding, Yufei Xia+2

Nanyang Technological University · Technical University of Munich · Alibaba+2

Clinical Language Model Evaluation Language Model Robustness Medical LLMs Radiology Report Evaluation

Open paper AskPDF arXiv

Reliable evaluation of generated radiology reports requires strict clinical accuracy, as omitted critical findings or mischaracterized radiographic observations can directly affect patient care. Existing metrics obscure this requirement by reducing report quality to a medically ungrounded scalar. Although Large Language Models (LLMs) possess rich medical knowledge, they likewise struggle to draw a reliable boundary between clinically significant errors and harmless variation. We study this boundary using ReEvalMed benchmark as testbed and evaluate metric-level clinical significance from detecting true clinical errors ("Discrimination") and tolerating insignificant variations ("Robustness"). Across 8 LLM evaluators under one-pass and two-pass settings, we identify a widespread discrimination bias: models effectively detect errors but also over-penalize harmless rephrasings. To mitigate this, we synthesize 4k report pairs and train lightweight interpretable metrics on Qwen3-8B and MedGemma-4B. Our trained metric sharpens the clinical significance boundary, surpassing 32B-scale medical LLMs and remaining competitive with proprietary models. Crucially, the more costly two-pass setting fails to consistently improve overall performance and mainly trades discrimination for robustness. These findings suggest one-pass trained metrics as the practical choice for cost-sensitive deployment, with two-pass inference reserved for settings where D-R balance is critical. We will release the dataset and metric.

Abstract

Clinical Language Model Evaluation

Explore similar work

Beyond Scalar Scores: Exploring LLM-based Metrics for Clinical Significance Evaluation in Radiology Reports

MedRECT: A Bilingual Medical Reasoning Benchmark for Error Correction in Clinical Texts

Importance of Prompt Optimisation for Error Detection in Medical Notes Using Language Models