New risk prediction model for opioid misuse after surgery surpasses accuracy of previous models

Feature Stories

New risk prediction model for opioid misuse after surgery surpasses accuracy of previous models

More than 90% of the annual 100 million surgical patients in the U.S. currently receive a prescription for opioids. A significant number of these patients—9 to 13% of those who have not used opioids before, and an even higher percentage of those already taking them before surgery—will continue to use opioids long-term, which exacerbates an already serious national opioid epidemic.

Using data from the Michigan Genomics Initiative (MGI), a risk prediction model developed by U-M researchers—the subject of a recent article in Surgery—has the potential to pinpoint those individuals most at risk of persistent opioid use (POU) after surgery and allow health care practitioners to take POU prevention measures before surgery.

“Opioid-based pain medications have an important therapeutic role in surgical recovery, but they also introduce risks of long-term use, physical dependence, and addiction. Many people use them for acute pain after surgery and stop without incident, but some do not, said Anne Fernandez, PhD, an Assistant Professor of Psychiatry and corresponding author on the study. “This research could help identify surgical patients who are at risk of persistent opioid use and trigger prevention efforts to mitigate this risk before it happens.”

One way to mitigate post-surgery opioid misuse is through opioid-sparing pathways (i.e., not prescribing opioids at all, and instead using other means and medications to manage pain). Another is preoperative opioid counseling. Offering counseling to those at highest risk targets patients most likely to benefit from these measures, and contains the costs and resources involved in counseling to those who most need it.

Because intervention strategies for patients who had previously taken opioids would include different information than opioid-naïve patient counseling, researchers tested model performance in each group separately. This was a first; previous POU prediction models had not been tested separately on each population.

Also novel to this model’s development was incorporating more diverse sources of information, instead of merely relying on claims data. In developing the model, U-M researchers considered patient-reported measures of pain and stress, prescription drug monitoring data, and data from the electronic health record.

“By integrating diverse data resources at the University of Michigan, including patient-reported outcome data from MGI, EHR data and prescription drug monitoring data, we were able to develop the most accurate model for predicting POU published in the literature,” Fernandez said. She added, “I would like to credit the work of MGI and Precision Health, and the Data Office for Clinical & Translational Research, for collecting, preparing, and making this data available to researchers like me to make this work possible and successful.”

The team assembled derivation and validation cohorts from the MGI, a longitudinal cohort of Michigan Medicine patients. Three versions of the model were developed: a full model, including 216 predictors, or variables; a restricted model consisting of 10 predictors, and minimal model using 5 predictors. All three performed better than existing POU prediction models, likely due to the inclusion of patient-reported measures and comprehensive opioid fill data.

“The development of this model was the cumulative goal of my funded Precision Health Investigators Award” —Anne Fernandez

Overall, the three variations were more accurate at predicting risk among preoperative opioid users than among opioid-naïve patients. Of the three, the minimal model, which did not include any patient-reported measures, performed the worst in the opioid-naïve population.

With only 10 predictors, the restricted model would be far easier to implement in a real-world clinical setting than the full model, while having comparable results, and the inclusion of patient-reported measures in the restricted model means it performs better among opioid-naïve patients than the more limited minimal model.

“We chose the restricted, 10 variable model, based on two things: parsimony and accuracy,” explained Fernandez. “To be useful in a real-world setting, the model has to be simple enough to implement in clinical care. A model with 100s of variables is not very easy to implement….The 10-variable model was the simplest model that didn’t sacrifice accuracy.”

One predictor not included in the restricted model, however, is race; the researchers carefully considered the inclusion or exclusion of race in the model. After three different analyses of how race as a predictor impacted model performance, they recommended the restricted model, which does not include race, over the full model.

“As this paper indicates, MGI data and Michigan Medicine data in general is fairly homogenous, with a very high proportion of White, Non-Hispanic individuals. We took several steps to evaluate model performance in diverse groups. It performed very well in White and Black individuals, however, the size of other subgroups was too small for analysis. For this reason, Dr. [Karandeep] Singh and I are working with Dr. Rahul Ladhania on a recently funded MIDAS grant to 1) get more data to increase the sample size for subgroups, and 2) evaluate the equitability of this model and several other persistent opioid use prediction models across diverse groups.”

The researchers plan to validate the restricted model at other health centers, using prospectively collected patient-reported data. If successful in a clinical setting, the model has the potential to reduce persistent opioid use by identifying those patients at highest risk and implementing opioid-sparing pathways or preoperative counseling.

Fernandez credits her 2018 Precision Health Investigators Award with enabling the research. “The development of this model was the cumulative goal of my funded Precision Health Investigators Award,” she said. “That grant was integral in this work, as was the connection it facilitated with the Precision Health data team and resources. I am very grateful for their investment in this line of research.”