Development and internal validation of a prediction model for long-term opioid use-an analysis of insurance claims data

In the United States, a public-health crisis of opioid overuse has been observed, and in Europe, prescriptions of opioids are strongly increasing over time. The objective was to develop and validate a multivariable prognostic model to be used at the beginning of an opioid prescription episode, aiming to identify individual patients at high risk for long-term opioid use based on routinely collected data. Predictors including demographics, comorbid diseases, comedication, morphine dose at episode initiation, and prescription practice were collected. The primary outcome was long-term opioid use, defined as opioid use of either >90 days duration and >= 10 claims or >120 days, independent of the number of claims. Traditional generalized linear statistical regression models and machine learning approaches were applied. The area under the curve, calibration plots, and the scaled Brier score assessed model performance. More than four hundred thousand opioid episodes were included. The final risk prediction model had an area under the curve of 0.927 (95% confidence interval 0.924-0.931) in the validation set, and this model had a scaled Brier score of 48.5%. Using a threshold of 10% predicted probability to identify patients at high risk, the overall accuracy of this risk prediction model was 81.6% (95% confidence interval 81.2% to 82.0%). Our study demonstrated that long-term opioid use can be predicted at the initiation of an opioid prescription episode, with satisfactory accuracy using data routinely collected at a large health insurance company. Traditional statistical methods resulted in higher discriminative ability and similarly good calibration as compared with machine learning approaches.