Predicting self-administered biologic nonadherence in Medicare fee-for-service beneficiaries with inflammatory bowel disease using machine learning models.

BACKGROUND

Nonadherence to self-administered biologic therapies increases risk for disease flares and adverse outcomes in inflammatory bowel disease (IBD).

OBJECTIVE

To use machine learning and Medicare fee-for-service claims to create models predictive of medication nonadherence.

METHODS

This study included Medicare fee-for-service beneficiaries with at least 2 IBD claims separated by 30 days between 2017 and 2021 and analyzed claims data from 2021. Beneficiaries with missing data or fewer than 3 self-administered biologic dispenses (adalimumab, certolizumab, golimumab, ustekinumab) on different dates in 2021 were excluded. Beneficiary-level nonadherence was defined as a proportion of days covered less than 0.8, and dispense-level nonadherence as dispenses occurring more than 5 days after previous supply elapsed. Sixteen machine learning models were trained, and model performance was evaluated based on area under the receiver operating characteristic curve (AUC) scores, accuracy, F1 score, positive and negative predictive value (P/NPV), sensitivity, and specificity.

RESULTS

A total of 10,160 beneficiaries met inclusion and exclusion criteria. Subsequently, 64,197 dispense transactions were observed, with 8,547 (13.3%) considered nonadherent. Prior nonadherence was strongly associated with current dispense nonadherence (odds ratio, 2.65; 95% CI, 2.53-2.78). The random forest dispense-level model had fair predictive performance (AUC 0.739, accuracy 85.4%, sensitivity 17.7%, specificity 95.8%, PPV 39%, NPV 88.3%, F1 score 0.243). The bagging dispense-level model performed comparably with higher sensitivity (AUC 0.714, accuracy 74.9%, sensitivity 51%, specificity 78.6%, PPV 26.8%, NPV 91.3%, F1 score 0.351).

CONCLUSIONS

Machine learning models trained on Medicare fee-for-service claims data had fair predictive performance identifying nonadherent dispenses. The bagging model, which minimizes false negatives, may be most appropriate for future clinical decision support tools.