Developing and validating a machine-learning algorithm to predict opioid overdose in Medicaid beneficiaries in two US...

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Public Health Surveillance

Background:

• Limited understanding of machine learning (ML) model generalizability for opioid overdose prediction across different populations and timeframes.
• Need for robust predictive models to identify individuals at high risk of opioid overdose.

Purpose of the Study:

• To develop and externally validate an ML algorithm for predicting the 3-month risk of opioid overdose.
• To assess the algorithm's performance using data from Pennsylvania and Arizona Medicaid beneficiaries.

Main Methods:

• A gradient-boosting machine algorithm was developed using Pennsylvania Medicaid data (2013-2016).
• The model utilized 284 predictors from pharmaceutical and healthcare claims data.
• External validation was performed on later Pennsylvania (2017-2018) and Arizona (2015-2017) Medicaid data.

Main Results:

• The algorithm demonstrated strong predictive performance with C-statistics ranging from 0.817 to 0.841 across validation datasets.
• High-risk subgroups identified by the model captured a significant proportion of overdose events (55-73%) with low overdose rates in lower-risk groups.

Conclusions:

• A Pennsylvania-derived ML algorithm for opioid overdose prediction performed well in external validation, including with data from a different state.
• The algorithm shows potential for clinical utility in predicting and stratifying overdose risk among Medicaid beneficiaries.