Natural Language Processing Algorithm Accurately Classifies Diverticulitis-Related Complications and Predicts...

Area of Science:

• Medical Informatics
• Computational Pathology
• Gastroenterology

Background:

• Diagnostic codes lack precision for diverticulitis complications in real-world data.
• Computed tomography (CT) reports contain detailed information on diverticulitis.
• Natural language processing (NLP) can extract features from unstructured text.

Purpose of the Study:

• Develop and validate an NLP algorithm to classify diverticulitis and associated features from CT reports.
• Assess the association between NLP-defined initial diverticulitis severity and severe recurrence risk.
• Evaluate the predictive value of NLP-detected features for severe diverticulitis recurrence.

Main Methods:

• Utilized Mass General Brigham Research Patient Data Registry (1979-2024).
• Developed and validated an NLP algorithm on abdominopelvic CT reports for patients with diverticular disease codes.
• Employed Cox proportional hazards regression and random forest models to analyze recurrence risk and feature prediction.

Main Results:

• The NLP algorithm demonstrated high positive (82.8%) and negative (99.9%) predictive values, outperforming ICD codes and a general large language model.
• Among 16,349 patients, initial diverticulitis severity (mild, severe, chronic) significantly increased the hazard ratio for severe recurrence (1.39 to 5.41).
• NLP-detected features significantly improved prediction of severe diverticulitis recurrence compared to codified variables.

Conclusions:

• An NLP algorithm accurately classifies diverticulitis features from CT reports, enabling large, high-quality electronic health record (EHR) cohorts.
• Initial diverticulitis severity is a strong predictor of severe recurrence risk.
• AI-driven risk stratification using NLP holds promise for long-term diverticulitis management.