Distinguishing Extraction Failures and Clinical Discordance with LLM Uncertainty on Radiology Reports
Organizations: Halıcıoğlu Data Science Institute, University of California, San Diego · Department of Radiology, University of California, San Diego · Department of Radiology and Halıcıoğlu Data Science Institute, University of California, San Diego
Abstract
Large language models (LLMs) can structure clinical datasets, but extraction error must be distinguished from discordance in the underlying data. From a real-world cohort of 1{,}045 patients undergoing coronary CT angiography (CCTA), we evaluated whether LLM uncertainty during vessel-level coronary artery disease (CAD) stenosis extraction could identify extraction failure and clinical discordance. References were expert CCTA annotations (253 patients) and subsequent invasive coronary angiography (ICA; 729 patients). We quantify extraction failure as disagreement between stenosis severity in LLM extracted versus expert annotations, and clinical discordance as disagreement in stenosis severity between CCTA and subsequent ICA. Stenosis Margin, selected from ten candidate measures, was used to quantify LLM uncertainty, and generalized estimating equation logistic regression and AUC evaluated its association and incremental discrimination for each failure mode. LLM uncertainty was strongly associated with extraction failure (adjusted OR 4.15 [95% CI 2.57--6.69], ) and substantially improved discrimination beyond LLM-extracted CCTA stenosis severity alone (AUC 0.81 to 0.86; AUC [, ], ). In contrast, LLM uncertainty weakly associated with clinical discordance (adjusted OR 1.22 [1.04--1.43], Holm-adjusted ) but provided no incremental discrimination beyond LLM-extracted CCTA stenosis severity (AUC [, ], ). Moreover, extraction errors and clinical discordance showed little overlap, with 38 of 42 clinically discordant vessels (90%) having been extracted correctly. LLM uncertainty may identify extraction failures without identifying underlying clinical discordance, supporting failure attribution when evaluating LLM-derived clinical datasets.
Figures & tables
| Abbreviation | Population | Analysis | Reference | Rows | Patients |
|---|---|---|---|---|---|
| P1 | All extracted vessels | — | — | 5,225 | 1,045 |
| P2 | CCTA external radiologist | Extraction Failure | External radiologist | 1,077 | 253 |
| P3 | CCTA ICA | Clinical Discordance | ICA | 1,641 | 729 |
| P4 | CCTA external radiologist ICA | Both | Both | 138 | 57 |
| Method | Quadratic weighted | Regex method, | Holm-adjusted |
|---|---|---|---|
| Regex | 0.946 (0.925–0.964) | — | — |
| Llama 3.3 70B | 0.974 (0.966–0.981) | ( to ) | 0.003 |
| MedGemma 27B | 0.963 (0.949–0.974) | ( to 0.004) | 0.21 |
| Qwen 3.5 9B | 0.957 (0.940–0.971) | ( to 0.012) | 0.36 |
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
| Abbreviation | Full name |
| Coronary vessels | |
| LM | Left main coronary artery |
| LAD | Left anterior descending artery |
| LCX | Left circumflex artery |
| RCA | Right coronary artery |
| RI | Ramus intermedius |
| Vessel | Segments | ICA vessel |
|---|---|---|
| Left main (LM) | LM | LM |
| Left anterior descending (LAD) | proximal, mid, distal LAD; D1; D2 | LAD |
| Ramus intermedius (RI) | RI | RI |
| Left circumflex (LCX) | proximal, mid LCX; OM1; OM2; L-PLB; L-PDA | LCX |
| Right coronary (RCA) | proximal, mid, distal RCA; R-PLB; R-PDA | RCA |
| Cohort | Step | Patients | Vessels |
|---|---|---|---|
| SOURCE | index studies | 7,847 | |
| SOURCE | excluded prior ica | 255 | |
| SOURCE | excluded stent or cabg | 302 | |
| SOURCE | excluded no ica in window | 6,443 | |
| SOURCE | ica eligible | 847 | |
| P1 | Vessel rows built from the cohort export | 1,045 | 5,225 |
| Analysis | Model | Measure | OR (95% CI) | P | Holm P |
|---|---|---|---|---|---|
| Extraction Failure | Univariable | Stenosis Margin | 1.10 (0.96–1.27) | 0.169 | 0.508 |
| Extraction Failure | Univariable | Stenosis Perplexity | 1.21 (0.99–1.47) | 0.060 | 0.240 |
| Extraction Failure | Univariable | Self-Confidence | 2.16 (1.77–2.64) | < 0.001 | < 0.001 |
| Extraction Failure | Univariable | Self-Ambiguity | 1.75 (1.48–2.06) | < 0.001 | < 0.001 |
| Extraction Failure | Univariable | Findings Margin | 1.12 (0.87–1.44) | 0.367 | 0.508 |
| Extraction Failure | Univariable | Findings Perplexity | 1.14 (0.94–1.38) | 0.193 | 0.508 |
| Model | AUC (95% CI) | AUC vs baseline (95% CI) | Bootstrap P | Brier score | Optimism- corrected AUC |
| P2 — Extraction Failure | |||||
| Baseline | 0.805 (0.763–0.842) | — | — | 0.0788 | 0.783 |
| Stenosis Margin alone | 0.580 (0.530–0.628) | — | — | 0.0917 | — |
| Baseline + Stenosis Margin | 0.862 (0.828–0.892) | +0.0571 (+0.0290, +0.0872) | 0.001 | 0.0702 | 0.845 |
| Stenosis Perplexity alone | 0.784 (0.738–0.824) | — | — | 0.0914 | — |
| Baseline + Stenosis Perplexity | 0.819 (0.779–0.854) | +0.0139 (+0.0028, +0.0268) | 0.014 | 0.0771 | 0.799 |
| Analysis | Measure | Vessels | Present (%) | Sparsest vessel |
|---|---|---|---|---|
| Extraction Failure | Stenosis Margin | 1,077 | 100.0 | — |
| Extraction Failure | Stenosis Perplexity | 1,077 | 100.0 | — |
| Extraction Failure | Self-Confidence | 1,077 | 100.0 | — |
| Extraction Failure | Self-Ambiguity | 1,077 | 100.0 | — |
| Extraction Failure | Findings Margin | 1,077 | 46.0 | LAD |
| Extraction Failure | Findings Perplexity | 1,077 | 46.0 | LAD |
| Measure A | Measure B | Spearman | Pearson r |
|---|---|---|---|
| Self-Confidence | Self-Ambiguity | 0.99 | 0.92 |
| Confidence Margin | Confidence Perplexity | 0.88 | 0.53 |
| Ambiguity Margin | Ambiguity Perplexity | 0.86 | 0.37 |
| Self-Ambiguity | Ambiguity Perplexity | 0.84 | 0.34 |
| Confidence Margin | Ambiguity Margin | 0.82 | 0.83 |
| Self-Confidence | Ambiguity Perplexity | 0.82 | 0.26 |
| Method | Assigned radiologist | ICA reference |
|---|---|---|
| Llama 3.3 70B | 0.969 (0.956–0.980) | 0.453 (0.410–0.494) |
| MedGemma 27B | 0.963 (0.949–0.974) | 0.450 (0.409–0.493) |
| Qwen 3.5 9B | 0.957 (0.940–0.971) | 0.411 (0.370–0.453) |
| Regex | 0.946 (0.925–0.964) | 0.335 (0.294–0.378) |
| Method | Quadratic weighted | Regex method, | Holm-adjusted |
|---|---|---|---|
| Regex | 0.335 (0.294–0.378) | — | — |
| Llama 3.3 70B | 0.454 (0.412–0.494) | ( to ) | 0.003 |
| MedGemma 27B | 0.450 (0.409–0.493) | ( to ) | 0.003 |
| Qwen 3.5 9B | 0.411 (0.370–0.453) | ( to ) | 0.003 |
LLM Reliability
Large language model reliability studies whether outputs remain correct, calibrated, robust, and controllable across prompts, contexts, and deployment conditions.
See all 858 papers