Machine-Learning Algorithms Can Predict Opioid Overdose Risk

Machine-learning algorithms indicate that 90% of Medicare beneficiaries with opioid prescriptions have a low risk of overdose, according to a study published in JAMA.

In this prognostic study conducted between September 2017 and December 2018, researchers evaluated the administrative claims data of 560,057 (mean age, 68, 63% female, 82% white) Medicare beneficiaries without cancer who filled one or more opioid prescriptions from January 2011 to December 2015. The beneficiaries were arbitrarily and equally partitioned into training (n=186,686), testing (n=186,685), and validation (n=186,686) samples. The researchers measured potential predictors (n=268), including socio-demographics, health status, patterns of opioid use, and practitioner-level and regional-level factors in three-month windows, and began the assessment three months prior to initiating opioids, until loss of follow-up or until the conclusion of the observation.

Opioid overdose occurrences from inpatient and emergency department claims were identified before researchers used multivariate logistic regression (MLR), least absolute shrinkage and selection operator-type regression (LASSO), random forest (RF), gradient boosting machine (GBM), and deep neural network (DNN) to project overdose risk in the subsequent three months after the start of treatment with prescription opioids. Additionally, researchers gauged prediction performance using the C statistic and other metrics such as sensitivity, number needed to evaluate (NNE) to identify one overdose, and specificity.

Minimal Overdose Episodes in Low-Risk Group

In the validation sample, the study’s results suggest the DNN (C statistic=0.91; 95% CI, 0.88 to 0.93) and GBM (C statistic=0.90; 95% CI, 0.87 to 0.94) algorithms outperformed the LASSO (C statistics=0.84; 95% CI, 0.80 to 0.89), RF (C statistic=0.80; 95% CI, 0.75 to 0.84), and MLR (C statistic=0.75; 95% CI, 0.69 to 0.80) methods for predicting opioid overdose. Moreover, in optimized sensitivity and specificity, DNN had a sensitivity of 92.3%, specificity of 75.7%, NNE of 542, positive predictive value of 0.18%, and negative predictive value of 99.9%.

Furthermore, the DNN categorized patients into low-risk (76.2% [142,180] of the cohort), medium-risk (18.6% [34,579] of the cohort), and high-risk (5.2% [9,747] of the cohort) subgroups, with only one in 10,000 in the low-risk subgroup incurring an overdose episode. More than 90% of overdose episodes occurred in the high-risk and medium-risk subgroups, although positive predictive values were low, given the rare overdose outcome.

“For those in the high-and medium-risk groups, although most will be false-positives for overdose given the overall low prevalence, additional screening and assessment may be warranted,” researchers wrote about the study’s results. “Although certainly not perfect, these machine-learning models allow interventions to be targeted to the small number of individuals who are at greater risk, and these models are more useful than other prediction criteria that have considerably more false-positives.”