Machine Learning Algorithm May Diagnose Opioid Use Disorder Months Before Physicians

A prescription bottle of pills spilled over
A prescription bottle of pills spilled over
As the diagnosis of opioid use disorder is often delayed or missed altogether, the researchers sought to test the utility of machine learning in creating a prediction model and algorithm for early diagnosis of OUD.

A machine learning algorithm from researchers at Ariel University in Israel allowed for an earlier diagnosis of opioid use disorder (OUD). These results were published in Pharmacological Research and Perspectives.

The medical claims database collected in the United States between 2006 and 2018 contained health care claims from 20 million patients was used for this study. Patients (N=130,120) who purchased at least 1 opioid were included. OUD was defined by presenting with at least 2 of the 11 defining problems (n=3239). The machine learning algorithm was built from 436 predictor candidates. Patients were randomly split into training (70%) and validation (30%) cohorts and a Gradient Boosting tree algorithm was implemented.

Control and OUD patients were aged mean 47.4 (standard deviation [SD], 15.5) and 53.6 (SD, 16.7) years and 40.3% were women. All medication features, diagnoses, and interactions with the health care system differed significantly (all P <.0001) between patients with and without OUD.

Patients with OUD had significantly more annual opioid prescriptions (3.66 ± 5.75 vs 0.64 ± 1.81; P <.0001), days of opioid treatment (87.63 ± 144.58 vs 11.96 ± 47.52 days; P <.0001), and longest consecutive opioid prescription (126.27 ± 319.15 vs 28.84 ± 156.13 days; P <.0001).

Patients with OUD also had costlier annual health care ($31,242.8 ± $78,000.7 vs $16,266 ± $30,471.7; P <.0001).

Beyond medication features, diagnoses, and interactions with the health care system, the machine learning model identified that hypertension, hyperlipidemia, number of hypertensive crisis events, and age were significant predictors for OUD.

The c-statistic for the OUD-predicting machine learning model was 0.959 (sensitivity, 0.85; specificity, 0.882; positive predictive value, 0.362; negative predictive value, 0.998). The positive predictive value was increased among the top 1% of patients (0.80).

The investigators divided medical records into 3-month increments and predicted future OUD, they observed that their algorithm could robustly predict OUD an average of 14.4 months before clinicians made a formal diagnosis.

This study may have been limited by using billable claims because much of the patient clinical details were not available.

This newly developed algorithm may allow for earlier diagnosis of OUD. By diagnosing this condition earlier, a substantial decrease in the cost of medical care coupled with earlier interventions may ultimately prevent death and suffering.

Disclosure: Multiple authors declared affiliations with industry. Please refer to the original article for a full list of disclosures.


Segal Z, Radinsky K, Elad G et al. Development of a machine learning algorithm for early detection of opioid use disorder. Pharmacol Res Perspect. 2020;8(6):e00669. doi:10.1002/prp2.669