OBJECTIVES: Medication non-adherence contributes to significant health care costs and detriments overall quality of life. Non-adherence may cause underlying conditions to spiral out of control until a disease is no longer manageable. Historically, only around 60% of prescribed medications are taken properly. Blue Cross and Blue Shield of Louisiana introduces the risk of medication non-adherence (RoMNA) models which leverage machine learning to predict whether someone is at risk of not taking their medication as prescribed. The models predict three classes of drugs: diabetes medications (non-insulin), anti-hypertensives and statins.

METHODS: Medication non-adherence is defined in terms of proportion of days covered (PDC). This is found via {PDC = (days of medication supply) / (total days of coverage) * 100}. If PDC is less than 80%, the individual is non-adherent. The dataset uses the past year of medical claims, social vulnerability data and previous adherence rates to predict the probability that the individual will be non-adherent by the end of the year. Researchers randomly sampled Medicare members who are managing one or more of the three drug types. After training the model, they refer who is most at risk of non-adherence to pharmacists to provide intervention.

RESULTS: The test area under the curve is 0.82 for each medication model. Based on member utilization, Blue Cross can accurately predict likelihood of medication adherence. The individual medications prove more challenging for precision and recall: RoMNA-Diabetes reached 0.61, RoMNA-Statins are 0.58 and RoMNA-Hypertension with 0.46. With these results, Blue Cross can concentrate outreach efforts on high-risk individuals.

CONCLUSIONS: Medication non-adherence remains an immense challenge in health care. The risk of medication non-adherence predictive models identifies member patterns before a problem begins, allowing for intervention that can improve an individual’s quality of life and lower their overall health care costs.