Machine-learning Algorithm Predicts Future Mortality Following Non-fatal Opioid Overdose
This retrospective cohort study used 2014–2016 Pennsylvania Medicaid data to develop a predictive model for all-cause mortality following non-fatal opioid overdose through applied machine learning. The algorithm used 348 predictors for 9686 individuals, including variables at the individual level (socio-demographics, health status, and health service utilization) and community level (e.g., poverty level, suicide rate) from the 180 days that preceded an index overdose. The main outcome was all-cause mortality within 180 days after the index overdose.
- Overall, 346 (3.6%) individuals died within 180 days after an index overdose.
- Those in the highest-risk group (≥98th percentile of risk) had a 180-day mortality rate of 20%; in the lowest-risk group (<25th percentile), the mortality rate was 1.5%.*
- When sensitivity and specificity were balanced, the algorithm’s negative and positive predictive values were 98% and 6.5%, respectively.
- Receiving medications for opioid use disorder or risk-mitigation interventions (naloxone, urine drug testing, substance use disorder counseling) after overdose were associated with lower mortality.
- Several community-level variables, such as county-level poverty or suicide rates, were important predictors of mortality.
* Individuals were stratified into 6 subgroups “at similar risk according to the risk scores (i.e., the individual’s estimated probability of death) generated by the validated machine learning algorithm.”
Comments: Having a score at the time of non-fatal opioid overdose to estimate future mortality risk could be useful for clinicians, health insurers, or government agencies. However, this algorithm used all-cause mortality as an outcome, which may have placed more emphasis on age and disability as predictors than if overdose mortality was used as the outcome. Even without a risk score, this study reinforces the importance of prescribing naloxone as well as identifying and treating opioid use disorder at the time of an overdose.
Aaron D. Fox, MD
Reference: Guo J, Lo-Ciganic WH, Yang Q, et al. Predicting mortality risk after a hospital or emergency department visit for nonfatal opioid overdose. J Gen Intern Med. 2021;36(4):908–915.