MM Curator summary
[MM Curator Summary]: A new study suggests that we can predict OD using 284 variables used in a new ML algo.
The article below has been highlighted and summarized by our research team. It is provided here for member convenience as part of our Curator service.
A new study shows that a machine-learning model using state Medicaid data may accurately predict opioid overdose in beneficiaries.
June 01, 2022 – Researchers have developed and validated a machine-learning (ML) model that can accurately predict opioid overdose risk in Medicaid beneficiaries in Pennsylvania and Arizona, indicating that the model may be valuable for making such predictions using beneficiary data from other states.
According to the study published in The Lancet Digital Health, the US opioid crisis continues to be a significant threat to public health, with an estimated 75,673 opioid overdose deaths in the 12 months ending in April 2021. Policymakers, health systems, and payers have implemented programs and policies to mitigate the crisis, but the shortcomings of current opioid risk prediction tools hamper progress.
The researchers note that despite these shortcomings, some progress is being made to develop more advanced models to improve t identification of individuals at risk of opioid overdose. Their previous work found that machine-learning approaches can improve risk prediction and stratification for opioid use disorder and subsequent overdose in Medicare beneficiaries.
However, there is little available information about whether ML algorithms developed to predict opioid overdose risk using Medicaid data from earlier years in a single state can be used for predictions in other states’ populations in later years. The researchers aimed to develop a model to predict the three-month risk of opioid overdose using Pennsylvania Medicaid data and then externally validate the model on two other datasets.
To develop their model, the researchers gathered data from Pennsylvania Medicaid beneficiaries with one or more opioid prescriptions from 2013 to 2016. A total of 284 potential predictors were pulled from pharmaceutical and healthcare encounter claims found in these data. Predictors were measured in three-month periods, starting three months before the first opioid prescription and continuing until the study’s end. This information was then used to predict the risk of hospital or emergency department (ED) visits for overdose in the subsequent three months.
The researchers externally validated their model using data with the same parameters from Pennsylvania between 2017 and 2018 and Arizona between 2015 and 2017.
Overall, the model achieved high performance with all three datasets. A total of 1.35 percent of 2013-2016 Pennsylvania Medicaid beneficiaries, 0.85 percent of 2017-2018 Pennsylvania Medicaid beneficiaries, and 0.61 percent of 2015-2017 Arizona beneficiaries had at least one overdose during the study period.
In external validation datasets, 22.4 percent of 2017-2018 Pennsylvania beneficiaries and 10 percent of 2015-2017 Arizona beneficiaries were in high-risk subgroups for opioid overdose. Lower risk subgroups in both external validation datasets showed 0.2 percent of beneficiaries or fewer with overdose risk.
These findings indicate that an ML model trained on 2013-2016 Pennsylvania Medicaid data successfully predicted opioid overdose in data from the same state in different years and a different state in different years. This may indicate that the model could be valuable for opioid overdose risk prediction and stratification in Medicare beneficiaries across states and time periods, according to the researchers.
Other risk models have also been recently developed to address opioid misuse and overdose.
In a study published by the University of Michigan earlier this year, researchers found that their risk prediction model helped identify and address potential misuse of opioids.
The model was developed using data from the Michigan Genomics Initiative and consisted of three different versions. The researchers found that all three versions performed better at predicting continued opioid use than existing models and were more successful at predicting opioid use among preoperative opioid users than inexperienced users.