NLP

NLP detects incidental durotomy in free-text thoraco-lumbar spine operative notes to enable automated surveillance of intraoperative adverse events.

Key Features:

• Automated Surveillance: Performs automated screening of operative notes to identify incidental durotomy without manual chart review.
• High Accuracy: Demonstrates AUC-ROC performance ranging from 0.74 to 0.99 across evaluated datasets.
• Geographical Validation: Validated across independent cohorts from Massachusetts and Maryland in the United States and Australia.
• Algorithmic Flexibility: Employs extreme gradient boosting as a tree-based algorithm to distinguish durotomy outcomes.
• Comprehensive Metrics: Evaluated using sensitivity, specificity, positive predictive value, negative predictive value, F1-score, likelihood ratios, Brier score, and calibration metrics.

Scientific Applications:

• Clinical Utility: Enables automated detection of incidental durotomy for monitoring surgical outcomes in spine surgery cohorts.
• Quality Improvement: Supports integration with registries and safety programs to evaluate and prevent adverse intraoperative events.
• Research Facilitation: Enables large-scale analyses of operative note corpora for studies of surgical outcomes and patient safety.

Methodology:

Retrospective analysis of free-text operative notes from thoraco-lumbar spine surgery across three cohorts; datasets were split 80:20 into training and test sets; separate NLP models were trained per cohort; model performance was assessed using discrimination (AUC-ROC), calibration metrics, sensitivity, specificity, positive predictive value, negative predictive value, F1-score, likelihood ratios, and Brier score.