Machine learning tool may determine opioid use disorder risk10/17/2018
By Janet Miller
Using data found on health care claims, researchers developed a machine learning tool that resulted in a “high level of predictive accuracy” to determine who is at risk for future opioid use disorder in the next 12 months, according to findings presented at PainWeek.
“Manual review of health care records by pain practitioners for [opioid use disorder] risk factors is resource-intensive due to the diversity of factors and the complexity of interactions found in health care data,” David M. Simon, of axialHealthcare in Nashville, Tennessee, and colleagues wrote. “Predictive models using machine learning offer a mechanism for distilling this high dimensional data into forms that are suitable for both risk stratification of patients and supporting health care decisions by practitioners.”
These models also provide a way to tailor treatment to patients’ individual needs and reduce health care costs, Simon and colleagues added.
Using government and commercial health care claims databases, researchers selected 23,371 individuals who were newly diagnosed with opioid use disorderand randomly chose 1,051,695 individuals without opioid use disorder.
For each patient, 983 features describing demographics, health care use and medical diagnoses were gathered and entered into the machine learning algorithm known as extreme gradient boosting (xGBoost) to predict whether a patient would be diagnosed with opioid use disorder.
Simon and colleagues found that the model’s predictive accuracy on the validation data (area under the receiver operating characteristic curve = 0.895; 95% CI, 0.890.9, chance = 0.5). Area under the precision-recall curve ranged from 0.071 to 0.123 depending patient subset (chance = 0.002-0.02).
In addition, uncommon medical diagnoses such as unspecified amphetamine abuse (ICD-9 305.70) and unspecified drug abuse (ICD-9 305.90) which researchers wrote have been previously been described in medical literature as major risk factors for opioid use disorder, were correctly retained by the model even though their prevalence was low in the sample.
“Crucially, the model was able to predict future [opioid use disorder] outcomes with a surprising level of fidelity in individuals who did not have a prescription for opioids documented in administrative claims during the predictive period,” Simon and colleagues wrote.
According to Greg Rudolf, MD, a staff physician at Swedish Pain Center in Seattle, this is the first time health care claims data entered into an electronic health record was used to develop a tool for ascertaining who is at risk for opioid use disorder.
Rudolf, who was not affiliated with Simon et al’s research, said the results are encouraging, but he cautioned that every model has its limitations, and pointed to the fact that much of the data recorded on the EHR, such as BP and weight, are just a snapshot of the patient at a moment in time.
“Even the validated screening tool we use, the Opioid Risk Tool, or ORT, which is a questionnaire that takes one to two minutes and has five questions, is subjective and relies on patient recall and honesty and thus has limitations. Other simple, commonly used tools, such a urine toxicology test, have inherent limitations, and some patients can find a way to beat,” Rudolf said. “Every screening model will have some positives and some negatives. But more data would likely be a good thing with regard to predictability and accuracy.”
Reference: Simon DM, et al. Identifying individuals at risk for a future opioid use disorder diagnosis with machine learning. Presented at PainWeek 2018; Sept. 4-8, Las Vegas.
Disclosure: Rudolf reports no relevant financial disclosures. Healio Family Medicine was unable to determine the other authors’ relevant financial disclosures prior to publication.