Large Language Model Versus Manual Review for Clinical Data Curation in Breast Cancer: Retrospective Comparative Study
Bibliographic record
Abstract
Background: Manual review of electronic health records for clinical research is labor-intensive and prone to reviewer-dependent variations. Large language models (LLMs) offer potential for automated clinical data extraction; however, their feasibility in surgical oncology remains underexplored. Objective: This study aimed to evaluate the feasibility and accuracy of LLM-based processing compared with manual physician review for extracting clinical data from breast cancer records. Methods: We conducted a retrospective comparative study analyzing breast cancer records from 5 academic hospitals (January 2019-December 2019). Two data extraction pathways were compared: (1) manual physician review with direct electronic health record access (group 1: 1366/3100, 44.06%) and (2) LLM-based processing using Claude 3.5 Sonnet (Anthropic) on deidentified data automatically extracted through a clinical data warehouse platform (group 2: 1734/3100, 55.94%). The automated extraction system provided prestructured, deidentified data sheets organized by clinical domains, which were then processed by the LLM. The LLM prompt was developed through a 3-phase iterative process over 2 days. Primary outcomes included missing value rates, extraction accuracy, and concordance between groups. Secondary outcomes included comparison with the Korean Breast Cancer Society national registry data, processing time, and resource use. Validation involved 50 stratified random samples per group (900 data points each), assessed by 4 breast surgical oncologists. Statistical analysis included chi-square tests, 2-tailed t tests, Cohen κ, and intraclass correlation coefficients. The accuracy threshold was set at 90%. Results: The LLM achieved 90.8% (817) accuracy in validation analysis. Missing data patterns differed between groups: group 2 showed better lymph node documentation (missing: 152/1734, 8.76% vs 294/1366, 21.52%) but higher missing rates for cancer staging (211/1734, 12.17% vs 43/1366, 3.15%). Both groups demonstrated similar breast-conserving surgery rates (1107/1734, 63.84% vs 868/1366, 63.54%). Processing efficiency differed substantially: LLM processing required 12 days with 2 physicians versus 7 months with 5 physicians for manual review, representing a 91% reduction in physician hours (96 h vs 1025 h). The LLM group captured significantly more survival events (41 vs 11; P=.002). Stage distribution in the LLM group aligned better with national registry data (Cramér V=0.03 vs 0.07). Application programming interface costs totaled US $260 for 1734 cases (US $0.15 per case). Conclusions: LLM-based curation of automatically extracted, deidentified clinical data demonstrated comparable effectiveness to manual physician review while reducing processing time by 95% and physician hours by 91%. This 2-step approach-automated data extraction followed by LLM curation-addresses both privacy concerns and efficiency needs. Despite limitations in integrating multiple clinical events, this methodology offers a scalable solution for clinical data extraction in oncology research. The 90.8% accuracy rate and superior capture of survival events suggest that combining automated data extraction systems with LLM processing can accelerate retrospective clinical research while maintaining data quality and patient privacy.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.034 | 0.085 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.002 |
| Bibliometrics | 0.004 | 0.003 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.002 | 0.003 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".