Use of natural language processing to extract uterine weight from pathology reports.
TL;DR
Natural language processing followed by simple CART can accurately extract total uterine weight from narrative pathology reports, and a small number of reports with data quality issues specific to low-weight specimens should be excluded or subjected to manual adjudication.
Abstract
INTRODUCTION For patients requiring hysterectomy, uterine weight is an important factor affecting surgical complexity and surgical outcomes. However, uterine weight is not readily available for analysis in administrative databases. In the present study, we use natural language processing to extract uterine weight from narrative pathology text.
Methods
Pathology text was obtained for 5000 patients sampled ±3 days of hysterectomy from hospital administrative datasets. The gross pathology subsections that describe the submitted tissues were retained for analysis. Manual annotation was performed to create a gold standard dataset. The texts were pre-processed, tokenized (split into words and punctuation), and tagged with parts-of-speech (e.g., noun, verb, number, adverb). Numeric tokens were retained for analysis. Feature engineering (e.g., variable construction and selection) included creating subspecimen flags, three tokens before and after the numeric tokens, weight units, and the prefix "weigh." Models included classification and regression trees (CART), extreme gradient boosting (XGB), and CatBoost, trained on 75%, validated on 15%, and tested in a 10% hold-out dataset. Performance on a real-world application was assessed for all hysterectomies performed over a 4-year period.
Results
Trained on all numeric tokens using features determined a priori, CART had 99.1% precision and 100% recall, with eight false positives. XGB performed similarly (99.3% precision; 99.9% recall). Performance was best using CatBoost's ordered target encoding and all values of the previous and subsequent three tokens (99.8% precision; 100% recall). Applied to the test dataset, CART slightly outperformed CatBoost with two fewer false positives. Summing all weights in the test set reports (n = 474), the extracted total specimen weight was a mean 9.4 g higher than the truth, owing predominantly to duplicate weights (22/37 discrepant reports). Application in the real-world dataset (n = 57 546) identified a small number of reports with data quality issues specific to low-weight specimens that should be excluded (≤10 g) or subjected to manual adjudication (>10 to ≤15 g).
Conclusion
Natural language processing followed by simple CART can accurately extract total uterine weight from narrative pathology reports. Further work is needed to extract specific anatomic weights.
