Toward Better Assessment of LLMs' Performance in Clinical Error Detection

1. Introduction

• Clinical documentation errors can propagate into care, while judging notes requires clinical reasoning and distinguishing incorrect decisions from unconventional but appropriate ones.Existing benchmarks and systems leave how and where LLMs fail largely unexamined; manual review is time-consuming, inconsistent, and difficult to scale.Source: §1 · pp. 1–2
• Benchmark pairs differ by one sentence; correctly flagging the erroneous note and clearing its counterpart distinguishes discrimination from assigning both the same label.The introduction notes that clinical error detection had not adopted pairwise evaluation.Source: §1 · p. 2, Figure 1 · p. 2
• The framework defines BCR for paired-note discrimination and uses ECA to compare cited evidence with ground-truth error and correction sentences.Experiments cover 15 LLMs, four paired datasets, three languages, and four prompt configurations, focusing on substitution errors available in public benchmarks.Source: §1 · pp. 2–3
• Most tested models fell below 25% random pairwise discrimination despite reasonable aggregate scores, with bias varying by language.The same model could show yes-bias in one language and no-bias in another.Source: §1 · p. 3
• Evidence Contrastive Analysis found that models locate error-relevant text but often give both notes the same verdict.In most yes-bias failures, models located the error sentence in the erroneous note yet labeled both notes erroneous.Source: §1 · pp. 2–3
• F1 and Both-Correct Rate can move in opposite directions under the same bias, so F1-highest models are typically BCR-lowest.BCR requires correct classification of both notes in a pair.Source: §1 · pp. 2–3

2.1. Error Detection in Clinical Texts

• Clinical error benchmarks have expanded from MEDIQA-CORR and MEDEC to multilingual tests, yet still emphasize aggregate comparisons rather than paired discrimination.This paper addresses that gap by adapting contrast consistency evaluation to the paired structure of MEDEC-style datasets.Source: §2.1 · p. 3

2.2. Prediction Bias in Healthcare

• Prediction bias concerns the output class a model defaults to, unlike demographic bias concerning which patients it treats differently; stability across languages and prompt configurations has not been investigated.Sycophancy can produce yes-bias in clinical error detection, defaulting to “error present” (Sharma et al., 2024).Source: §2.2 · p. 4

2.3. Pairwise Evaluation and Representation Gaps

• Signal detection theory distinguishes class discrimination from response bias, while minimal-pair and contrast-set evaluations test whether models learn relevant distinctions (Warstadt et al., 2020; Gardner et al., 2020).Although clinical error datasets have paired structure, clinical NLP benchmarks have not adopted contrastive evaluation (Ben Abacha et al., 2024; Iwase et al., 2025; Ma et al., 2026).Source: §2.3 · p. 4
• Prior studies find that models can encode truth internally yet generate wrong answers, and that chain-of-thought can misstate prediction drivers; whether correct evidence appears alongside wrong verdicts remains unstudied.Probing requires weight access, limiting its applicability to closed models prevalent in clinical deployment (Burns et al., 2023; Orgad et al., 2025; Turpin et al., 2023).Source: §2.3 · p. 4

3. Methods

• The framework’s organizing principle is pairing each error-injected note with its correct counterpart, then assessing performance through pointwise metrics, BCR, and ECA.The framework is applied across four test sets, three languages, and four prompt–decoding configurations per model–dataset combination.Source: §3 · pp. 4–5, Figure 2 · p. 5

3.1. Datasets and Task

• Four test sets cover English, Chinese, and Japanese, pairing injected-error notes with correct counterparts from the same scenarios; only the binary error-flag task is evaluated, with notes judged independently.Usable matched pairs are identified using dataset-specific procedures; MedRECT-JA has 190 error notes but only 105 unique clean notes, potentially inflating within-pair error correlation.Source: §3.1 · p. 5

3.2. Models

• The study evaluates 15 instruction-tuned LLMs across three size tiers and five model families, comparing scale within families and medical specialization at matched scale in a zero-shot setting.All models are evaluated without task-specific examples or fine-tuning.Source: §3.2 · pp. 5–6

3.3. Prompt Design and Decoding

• A 2×2 prompt-and-decoding matrix tests instruction-driven shifts in prediction bias and sampling variability, with cross-configuration means and standard deviations used to assess robustness.Neutral and conservative prompts require structured Evidence, Analysis, Confidence, and Error:Yes/No fields; greedy decoding is primary for reproducibility, alongside stochastic sampling.Source: §3.3 · p. 6

3.4.1. Traditional Metrics

• Traditional reporting includes balanced accuracy, F1, precision, recall, specificity, MCC, and error-flag rate, with balanced accuracy and MCC as primary anchors.F1 and recall are included to expose discrepancies between bias-sensitive and bias-resistant metrics.Source: §3.4.1 · p. 7

3.4.2. Both-Correct Rate (BCR)

• BCR measures the fraction of matched pairs correctly classified on both notes, while its independence ratio diagnoses within-pair dependence and its upper bound exposes how class bias constrains paired success.Under conditional independence, expected BCR equals sensitivity × specificity; on balanced pair sets, the random-classifier baseline is 25%, while the prevalence-calibrated MedRECT-JA baseline is approximately 22.9%.Source: §3.4.2 · pp. 7–8

3.4.3. Evidence Contrastive Analysis (ECA)

• ECA scores whether cited evidence overlaps relevant sentences in each note, distinguishing failures to locate evidence from failures to judge it correctly.It combines error-note true-positive localization and clean-note false-positive evidence-hit using substring containment or at least 60% word coverage, with character-level counting for Chinese and Japanese; Figure 3 shows five resulting categories.Source: §3.4.3 · p. 8, Figure 3 · p. 9

3.4.4. Parse Failure Handling

• Pairs with either member lacking a parseable verdict are excluded from BCR, checked against random-assignment imputation, and counted as Extraction-Fail in ECA.Some outputs add text, omit the Error field, or are malformed despite the required four-line format.Source: §3.4.4 · pp. 8–9

4.1. Traditional Metrics Suggest Moderate Performance

• 178 of 240 runs (74%) have F1 above 0.5 while balanced accuracy is at or below 0.6, despite balanced accuracy ranging from 0.45 to 0.71 and MCC being near zero for most models.F1 reaches 0.80, a pattern that standard evaluation practice would suggest is moderate performance, but the pairwise analysis does not support that impression.Source: §4.1 · p. 9

4.2. Pairwise BCR Reveals Low Discrimination

• 13 of 15 models have mean BCR below the 25% balanced-data random baseline; only Qwen 3-32B (28.0%) and UltraMedical 70B (25.7%) exceed it.Both exceptions still fail to correctly classify both members of a pair more than 70% of the time. The lowest-BCR models can collapse one marginal, as with Gemma 3-4B’s 94.2% sensitivity and 7.2% specificity.Source: §4.2 · pp. 9–10, Table 2 · p. 10
• Qwen’s mean BCR rises from 17.0% (4B) to 28.0% (32B), while medical-domain pretraining raises Gemma 3-27B’s 10.2% to MedGemma 27B’s 23.1%.Scaling gains are concentrated on English datasets, and the medical-domain comparison holds for four of five matched general–medical pairs. Cross-configuration standard deviations range from 1.1 to 15.6 pp; Llama 3.1-70B’s MedErrBench-EN BCR spans 9.0–45.2% across configurations.Source: §4.2 · p. 10

4.3. Bias Is Bidirectional and Language-Dependent

• Eight of 15 models change bias category across datasets, and seven have mean error-flag rates on opposite sides of parity across datasets.Qwen 3-8B reverses from no-bias on Chinese (0.36) to yes-bias on Japanese (0.69); seven models also switch from yes-bias under neutral prompting to no-bias under conservative prompting on at least one dataset.Source: §4.3 · p. 11, Table 3 · p. 11

4.4. The Localization-Judgment Gap

• 70% of categorized Pred1 pairs on MS-Test show models attending to the error sentence on the error note, yet still misjudging the pair.Both-Hit cases, where models attend to the relevant sentence on both notes but judge both erroneous, average 38%; full attention failure averages 18%. Mean TP localization across models is 87% on MEB-EN, 70% on MEB-CN, 69% on MS-Test, and 42% on MRT-JA.Source: §4.4 · p. 12, Figure 4 · p. 12

4.5. Prediction Bias Mediates the Relationship Between F1 and BCR

• 32 of 60 model–dataset entries (53%) have F1 ≥0.6 but BCR <25%, while error-flag rate correlates with higher F1 (r = +0.85) and lower BCR (r = −0.49).The direct F1–BCR correlation is near zero (r = −0.06), and the top-three F1 and BCR models have zero overlap on three of four datasets; all 60 entries fall below the independence line.Source: §4.5 · p. 13, Figure 5 · p. 13
• GPT-5 mini’s MS-Test BCR is 42.3% (F1 = 0.66, balanced accuracy = 0.69), exceeding every open-weight model on that dataset, whose best BCR is 24.3%.This illustrative two-run result remains below GPT-5 mini’s independence baseline of 47.1%; it is not a full proprietary evaluation.Source: §4.5 · pp. 13–14

5.1. Implications for Clinical Model Selection

• Thirteen of 15 models fell below 25% random BCR despite many scoring above 0.6 F1; on MS-Test, F1 ranked Gemma 3-4B, Llama 3.2-3B, and Mistral 7B highest, although their BCRs were just 4.5–5.9%.The paper reports that high F1 can be driven by bias in the ≈50–60% balanced-accuracy regime of clinical error detection, unlike at ≈90% balanced accuracy. Across tested models, the high-F1/high-BCR quadrant was almost empty, supporting BCR alongside traditional metrics.Source: §5.1 · p. 14
• Benchmarkers, healthcare governance teams, and vendors can use BCR and the independence ratio alongside F1 and MCC to screen candidates, but paired screening cannot replace prospective clinician-led workflow evaluation.The screen tests whether a model can distinguish an erroneous note from its clean counterpart; most tested models cannot.Source: §5.1 · pp. 14–15

5.2. Evidence Production and Clinical Judgment

• On MS-Test, Both-Hit pairs averaged 38% of Pred1 failures: models cited corresponding error-relevant sentences in both notes but predicted error for both, a localization diagnostic rather than proof of comprehension.The finding is detectable from generated outputs without internal representations, and embedding-based Recall@1 reproduced the magnitude and per-model ranking of TP localization.Source: §5.2 · p. 15
• Contrastive fine-tuning or localization-then-judgment pipelines could target weak judgment; high-recall yes-biased models may support pre-filtering only with human review, workable precision, and monitored alert fatigue.The paper cautions that current evidence may not support standalone clinical error detection deployment.Source: §5.2 · p. 15

5.3. Limitations

• Under zero-shot prompting, conservative prompts shifted BCR by up to 12.8 pp, yet at least 13 of 15 models remained below the 25% random level; evaluation otherwise covered open-weight models up to 70B and one proprietary reference.Sampling-based configurations used one run, while the primary four-configuration mean included two reproducible greedy runs. The authors frame the work as diagnostic, not a deployment ranking, and note possible benchmark-data leakage and limited proprietary coverage.Source: §5.3 · pp. 15–16
• The empirical results cover substitution errors; insertion and omission pairs remain future work, while ECA’s sentence-level matching and MedRECT-JA’s multisentence notes limit localization granularity.TP localization may partly reflect entity salience, and internal mechanisms behind the localization-judgment gap were not investigated; the gap was observed across all four benchmarks.Source: §5.3 · p. 16

Appendix A. Pair Construction Details

• BCR requires matched error–correct notes, constructed per dataset: MEDEC pairs used Jaccard overlap, while MedErrBench-EN and -CN paired consecutive alternating error and clean samples.For MEDEC (Ben Abacha et al., 2025), one-to-one matching at ≥0.6 paired 286 of 311 error notes; unmatched notes were excluded from BCR but retained for pointwise metrics. MedErrBench (Ma et al., 2026) yielded 104 EN pairs and 100 CN pairs with no unmatched samples.Source: Appendix A · p. 20
• MedRECT-JA pairs matched error notes to same-scenario clean notes above 0.85 character-level overlap, yielding 190 pairs with some clean notes reused across pairs.MedRECT (Iwase et al., 2025) had 190 error notes but only 105 unique clean notes, so systematic errors on shared clean notes could inflate within-pair error correlation.Source: Appendix A · p. 21

Appendix B. An Upper Bound on BCR

• BCR is bounded above by min(sensitivity, specificity), making class-bias-driven F1/BCR divergence structural rather than specific to the tested models.Both-Correct pairs are the intersection of correctly flagged error notes and correctly cleared correct notes, so their rate cannot exceed either marginal rate.Source: Appendix B · pp. 21–22
• An always-error predictor has sensitivity 1, specificity 0, and BCR 0, yet F1 is 2/3 ≈ 0.67 on a class-balanced set.Its precision is 1/2 and recall is 1; more generally, bias toward one class caps BCR at the weaker marginal even as F1 rises through recall on the favored class.Source: Appendix B · pp. 21–22

Appendix C. Evaluated Models: References and Identifiers

• Table 4 catalogs the 15 evaluated instruction-tuned models by citation, HuggingFace identifier, family, medical-domain status, and loading precision, with separators for three size tiers.The tiers are Small (3–8B), Large (27–32B), and Very Large (70B).Source: Table 4 · p. 22, Appendix C · p. 22

Appendix D. Prompt Templates and Decoding Settings

• Greedy decoding used T = 0, while stochastic sampling used T = 0.7 and top-p = 0.9; each model–dataset pair had four runs across prompt-by-decoding configurations.Each model was run once per prompt × decoding configuration.Source: Appendix D · p. 22
• Prompts required four lines—Evidence, Analysis, Confidence, and Error—and neutral prompts required an exact quote of explicitly incorrect or unsafe text to flag an error.Evidence could be a quoted span or NA; Analysis was 3–5 sentences, Confidence ranged from 0–100, and without an exact quotable span the instruction was Error: No.Source: Appendix D · pp. 22–23
• The conservative variant asked reviewers to flag only clearly supported significant errors, avoid assumptions or invented facts, and prefer no error for incomplete or ambiguous notes.Its user prompt otherwise followed the neutral structure, replacing “error” with “significant medical error.”Source: Appendix D · p. 23
• Dataset-specific prompts expanded MedErrBench-EN to 10 error categories and translated the prompt structure for MedErrBench-CN and MedRECT-JA, retaining English output keys.MedRECT-JA used nine Japanese-language error types derived from its annotation taxonomy.Source: Appendix D · p. 23

Parse failure handling.

• Nine of 240 runs exceeded a 5% parse failure rate, concentrated in UltraMedical 8B and 70B on non-English datasets and Llama 3.2-3B on MedErrBench-CN.The listed runs are summarized in Table 5.Source: Parse failure handling · p. 23, Table 5 · p. 24
• Skip versus random assignment changed BCR by less than 0.5 pp for all 15 models, and the same 13 of 15 remained below 25% BCR under either strategy.All reported BCR values use skip; random assignment is only a robustness check.Source: Parse failure handling · pp. 23–24, Table 6 · p. 24
• Changing ECA word-coverage and MS-Test Jaccard pairing thresholds from 0.5 to 0.7 shifted per-model TP-localization rates by less than 3 pp while preserving model rankings.For MRT-JA, lowering the character-overlap threshold from 0.85 to 0.75 added no pairs; the 0.85 threshold yields 190 pairs.Source: Parse failure handling · p. 24
• The embedding robustness test pooled 9,428 localization judgments across 15 models, four configurations, and biomedical and general encoders, and required the gold error sentence to be the single most similar sentence.The relative-ranking Recall@1 test is stricter than an absolute threshold and is not tied to a word-overlap cutoff.Source: Parse failure handling · pp. 24–25
• Per-model embedding Recall@1 correlated with substring TP-localization at Pearson r = 0.99 / 0.98, reproducing the 69% localization finding without claiming semantic abstraction from word overlap.The agreement reflects that models predominantly cited the error sentence verbatim, even when labeling both erroneous and corrected notes as containing an error.Source: Parse failure handling · p. 25
• Embedding-based scoring yielded pooled Recall@1 of 65.3% / 67.1% against a 12.6% random baseline, and all 15 models exceeded baseline under both encoders.This re-scored the same outputs using the cited evidence span’s similarity to each error-note sentence; hits required the gold error sentence to rank first.Source: Table 7 · p. 25, Parse failure handling · p. 25
• Model rankings remained largely stable across four configurations, although Llama 3.1-70B’s absolute BCR varied by up to 36.2 pp.Figure 6 reports BCR for all 15 models under each configuration.Source: Parse failure handling · p. 25, Figure 6 · p. 26

• Error-flag rate suppresses BCR across all datasets and strongly inflates F1 on three of four, with a weaker effect on MEB-EN.Source: Table 8 · p. 26, Appendix F · p. 26
• The direct F1–BCR correlation varies in sign across datasets and is near zero only when pooled.Source: Table 8 · p. 26
• On three of four datasets, no model appears in both the top three by F1 and the top three by BCR.The lists use four-configuration means.Source: Table 9 · p. 27
• Actual BCR falls below expected sensitivity × specificity for all 60 model–dataset entries.The independence ratio is the mean of per-model R_independence values, not the quotient of the two column means.Source: Table 10 · p. 27

G.1. Full Traditional Metrics (All 240 Runs)

• Tables 11–14 each cover 15 models across four datasets, with 60 runs per configuration, and sort entries by balanced accuracy within each dataset.The appendix defines Flag% as the fraction of notes flagged as containing an error.Source: §G.1 · p. 27, Table 11 · p. 27, Table 12 · p. 29
• In Conservative-Greedy, the highest reported balanced accuracies are .692 for UltraMed-70B on MedErrBench-EN, .590 for Qwen 3-32B on MedErrBench-CN, and .585 for Qwen 3-32B on MedRECT-JA.These are the top balanced-accuracy entries in the supplied rows for each named dataset.Source: Table 11 · p. 27, §G.1 · p. 28
• The supplied Conservative-Sampling rows include UltraMed-70B at .558 balanced accuracy and Llama 3.2-3B at .767 F1; the dataset name is not included.These are distinct model-metric pairs from the table’s first dataset listing.Source: Table 12 · p. 29, §G.1 · p. 30
• In Neutral-Greedy, Qwen 3-4B and MedGemma 27B tie at .590 balanced accuracy in the first dataset listing, and Qwen 3-4B reaches .600 on MedRECT-JA.The first listing’s dataset name is not included in the supplied passage.Source: Table 13 · p. 30, §G.1 · p. 31
• In Neutral-Sampling, Qwen 3-32B has the highest reported balanced accuracy on MedErrBench-CN (.655), while Qwen 3-32B leads MedRECT-JA (.595).The supplied rows also show UltraMed-8B at .615 balanced accuracy in the first dataset listing, whose name is not included in the passage.Source: Table 14 · p. 31, §G.1 · p. 32

G.2. Full ECA Category Breakdown

• Table 15 lists all 60 model–dataset combinations, with category counts pooled across four prompt–decoding configurations.Each combination had at least one configuration with five or more Pred1 pairs.Source: §G.2 · p. 32, Table 15 · p. 33
• TP localization and FP evidence-hit are four-configuration mean overlap rates, while categories A–E are pooled counts; the two TP measures can differ by several points.TP localization is an unweighted mean, whereas (A + B)/n is based on pooled counts; uneven Pred1 counts across configurations can produce the difference.Source: §G.2 · p. 32, Table 15 · p. 33