Predicting benzodiazepine prescriptions: A proof-of-concept machine learning approach

Area of Science:

• Pharmacovigilance
• Health Informatics
• Machine Learning in Healthcare

Background:

• Benzodiazepines are widely prescribed psychotropic medications with potential for serious adverse effects.
• Predictive modeling for benzodiazepine prescriptions is crucial for developing targeted prevention strategies.

Purpose of the Study:

• To develop and evaluate machine learning algorithms for predicting benzodiazepine prescription receipt and quantity.
• To utilize de-identified electronic health record (EHR) data for predictive model development.

Main Methods:

• Support-vector machine (SVM) and random forest (RF) algorithms were applied to EHR data from outpatient psychiatry, family medicine, and geriatric medicine.
• Models were trained on data from January 2020 to December 2021 (N = 204,723 encounters) and tested on data from January to March 2022 (N = 28,631 encounters).
• Features included diagnoses (anxiety, sleep disorders), demographics, concurrent medications (opioids, antidepressants, antipsychotics), clinical variables, and insurance status.

Main Results:

• Both SVM and RF models demonstrated high accuracy and Area Under the ROC Curve (AUC) for predicting benzodiazepine prescription receipt (yes/no).
• SVM achieved accuracies of 0.868-0.883 and AUCs of 0.864-0.924; RF achieved accuracies of 0.860-0.887 and AUCs of 0.877-0.953.
• High accuracy was also observed for predicting the number of benzodiazepine prescriptions (0, 1, 2+), with SVM (0.861-0.877) and RF (0.846-0.878).

Conclusions:

• Machine learning models, specifically SVM and RF, can accurately predict benzodiazepine prescription status and quantity.
• These predictive models hold potential for informing system-level interventions to mitigate the public health impact of benzodiazepine use.
• Further replication is recommended to validate these findings for clinical implementation.