Cue-Based Narrative Defect Detection for Medical Record Quality Control: A Monte Carlo Simulation Study Comparing Cue Rules, TF-IDF Baselines and a Learned Classifier (Preprint)

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

BACKGROUND Medical record quality control relies on manual review of narrative fields where defects such as negated or temporally confused findings are common. Automated cue-based rules can flag these defects, but their precision depends on how well cues capture real clinical language. We compare a cue-based narrative pipeline against TF-IDF baselines and a learned classifier on synthetic defect-bearing records. OBJECTIVE To quantify how cue-based rule induction, TF-IDF features and a learned classifier perform on narrative defect detection in a simulated medical-record quality-control setting, and to test whether adding narrative cues improves on coded-only rules. METHODS We simulated 8,000 records with 12 defect types distributed across coded and narrative fields. Three detectors were compared: coded rules alone, a cue-based narrative pipeline, and a learned logistic-regression classifier with cue features. McNemar tests and bootstrap 95% confidence intervals compared paired F1 scores. All data were synthetic; no real patient records were used. RESULTS Coded rules alone achieved F1 0.17. The cue-based narrative pipeline reached F1 0.60, and the learned classifier reached F1 0.84. The improvement from coded rules to the cue pipeline was statistically significant (McNemar χ²=42.3, p<0.001). The cue-based pipeline was most effective for negation and temporal defects. CONCLUSIONS Adding narrative cue features substantially improves defect detection over coded-only rules in simulation. The approach is transparent and requires no labelled training corpus. Validation on real clinical text from an established hospital is the essential next step.