A Risk Prediction Model for Long-term Prescription Opioid Use

Abstract

Background: Tools are needed to aid clinicians in estimating their patients' risk of transitioning to long-term opioid use and to inform prescribing decisions.

Objective: The objective of this study was to develop and validate a model that predicts previously opioid-naive patients' risk of transitioning to long-term use.

Research design: This was a statewide population-based prognostic study.

Subjects: Opioid-naive (no prescriptions in previous 2 y) patients aged 12 years old and above who received a pill-form opioid analgesic in 2016-2018 and whose prescriptions were registered in the California Prescription Drug Monitoring Program (PDMP).

Measures: A multiple logistic regression approach was used to construct a prediction model with long-term (ie, >90 d) opioid use as the outcome. Models were developed using 2016-2017 data and validated using 2018 data. Discrimination (c-statistic), calibration (calibration slope, intercept, and visual inspection of calibration plots), and clinical utility (decision curve analysis) were evaluated to assess performance.

Results: Development and validation cohorts included 7,175,885 and 2,788,837 opioid-naive patients with outcome rates of 5.0% and 4.7%, respectively. The model showed high discrimination (c-statistic: 0.904 for development, 0.913 for validation), was well-calibrated after intercept adjustment (intercept, -0.006; 95% confidence interval, -0.016 to 0.004; slope, 1.049; 95% confidence interval, 1.045-1.053), and had a net benefit over a wide range of probability thresholds.

Conclusions: A model for the transition from opioid-naive status to long-term use had high discrimination and was well-calibrated. Given its high predictive performance, this model shows promise for future integration into PDMPs to aid clinicians in formulating opioid prescribing decisions at the point of care.

Figure 1:

Participant Flow Chart through Study. PDMP: Prescription drug monitoring program ^a With complete patient records ^b Defined as a patient receiving a new opioid analgesic prescription with no other opioid prescribed in the previous two years (730 days) and no opioid indicated for treatment of opioid use disorder prescribed in the index prescription or follow-up period ^c Excluded patients may meet multiple exclusion criteria ^d Some patients (n = 94,335) qualify as opioid-naïve in both 2016 and 2018 due to at least 730 days between their opioid use episodes and therefore appear in both the development and validation datasets

Figure 2:

Calibration Curve for a Model Predicting the Transition of Opioid-naïve Patients to Long-term Use among California Residents, 2016-2018. The plot shows mean observed outcome (transition of previously opioid-naïve patients to long-term opioid use), by deciles of predicted probability (each circle represents a decile). Predicted probabilities are from application of the prediction model, developed on the 2016-2017 cohort, to the validation (2018) cohort. The grey line reflects perfect agreement between observed and predicted probabilities. Error bars represent 95% confidence intervals.

Figure 3:

Decision curve analysis comparing clinical utility, as measured by net benefit, for various prescribing strategies by decision threshold The plot shows decision curve analysis comparing the net benefit for four different prediction strategies: the thick solid line represents the validated risk prediction (full) model (A), the thick dashed line a prediction model without the dose trajectory (set 2) variables (B), the thin solid line when transition to long-term use is predicted for all patients (C), and the thin dotted line when transition to long-term use is predicted for no patients (D). The optimal threshold (4.89%) was calculated by maximizing the Youden’s J statistic (= Sensitivity + Specificity − 1). Appendix 5 (Supplementary Digital Content 2) contains details on how net benefit is calculated.