Machine learning for phenotyping opioid overdose events.

Area of Science:

• Clinical informatics
• Machine learning in healthcare
• Opioid overdose research

Background:

• Opioid overdose is a critical public health issue.
• Accurate classification of overdose severity is essential for appropriate clinical response.
• Existing methods for severity classification may be limited.

Purpose of the Study:

• To develop and evaluate machine learning models for classifying opioid overdose severity.
• To compare the performance of different feature sets and modeling approaches.

Main Methods:

• Identified opioid overdoses using diagnosis codes from the Marshfield Clinic population.
• Assigned severity scores via chart review to create a gold standard dataset.
• Constructed feature sets from disparate data sources and a common data model.
• Trained and evaluated random forest and penalized logistic regression models.

Main Results:

• Random forest models achieved high performance (AUC 0.893).
• Features from a common data model outperformed those from disparate sources (0.893 vs. 0.837).
• Incorporating free text features significantly improved model accuracy (AUC increased from 0.827 to 0.893).
• Key NLP-derived features like 'Narcan' and 'Endotracheal Tube' were crucial.

Conclusions:

• Machine learning models, particularly random forest, are effective for classifying opioid overdose severity.
• Utilizing a common data model and incorporating free text analysis enhances classification accuracy.
• NLP-derived features play a vital role in improving model performance for overdose event severity.