Machine Learning to Predict Tamoxifen Nonadherence Among US Commercially Insured Patients With Metastatic Breast Cancer.

Abstract

PURPOSE\nAdherence to tamoxifen citrate among women diagnosed with metastatic breast cancer can improve survival and minimize recurrence. This study aimed to use real-world data and machine learning (ML) methods to classify tamoxifen nonadherence.\n\n\nMETHODS\nA cohort of women diagnosed with metastatic breast cancer from 2012 to 2017 were identified from IBM MarketScan Commercial Claims and Encounters and Medicare claims databases. Patients with < 80% proportion of days coverage in the year following treatment initiation were classified as nonadherent. Training and internal validation cohorts were randomly generated (4:1 ratio). Clinical procedures, comorbidity, treatment, and health care encounter features in the year before tamoxifen initiation were used to train logistic regression, boosted logistic regression, random forest, and feedforward neural network models and were internally validated on the basis of area under receiver operating characteristic curve. The most predictive ML approach was evaluated to assess feature importance.\n\n\nRESULTS\nA total of 3,022 patients were included with 40% classified as nonadherent. All models had moderate predictive accuracy. Logistic regression (area under receiver operating characteristic 0.64) was interpreted with 94% sensitivity (95% CI, 89 to 92) and 0.31 specificity (95% CI, 29 to 33). The model accurately classified adherence (negative predictive value 89%) but was nondiscriminate for nonadherence (positive predictive value 48%). Variable importance identified top predictive factors, including age ≥ 55 years and pretreatment procedures (lymphatic nuclear medicine, radiation oncology, and arterial surgery).\n\n\nCONCLUSION\nML using baseline administrative data predicts tamoxifen nonadherence. Screening at treatment initiation may support personalized care, improve health outcomes, and minimize cost. Baseline claims may not be sufficient to discriminate adherence. Further validation with enriched longitudinal data may improve model performance.