Using Natural Language Processing to Predict Fatal Drug Overdose From Autopsy Narrative Text: Algorithm Development...

Area of Science:

• Computational toxicology
• Public health surveillance
• Medical informatics

Background:

• Fatal drug overdose surveillance is crucial for prevention but often delayed due to autopsy and death certificate processing.
• Autopsy reports contain valuable narrative text that can serve as an early data source for overdose identification.

Purpose of the Study:

• To develop and validate a natural language processing (NLP)-based model for predicting accidental or undetermined fatal drug overdoses from autopsy report narratives.

Main Methods:

• Extracted text from 17,342 autopsy reports (2019-2021) using optical character recognition.
• Preprocessed text and applied machine learning classifiers (logistic regression, SVM, random forest, gradient boosted trees).
• Evaluated model performance using metrics like area under the receiver operating characteristic, precision, recall, F1-score, and F2-score, with calibration and subgroup analyses.

Main Results:

• All models demonstrated excellent performance (Area under the ROC curve ≥0.95).
• Random forest and SVM classifiers achieved the highest F2-scores (0.948 and 0.947).
• Key terms like "fentanyl" and "accident" had high importance (Shapley additive explanations).

Conclusions:

• The random forest classifier is a promising tool for identifying potential accidental and undetermined fatal overdose autopsies.
• Further validation is needed to ensure early detection across diverse subgroups.