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

Area of Science:

• Oncology
• Health Informatics
• Machine Learning

Background:

• Tamoxifen citrate adherence is crucial for improving survival and minimizing recurrence in metastatic breast cancer patients.
• Real-world data and machine learning (ML) offer potential for identifying nonadherence patterns.

Purpose of the Study:

• To classify tamoxifen nonadherence using real-world data and ML methods.
• To identify predictive factors for tamoxifen nonadherence in metastatic breast cancer.

Main Methods:

• A cohort of 3,022 women with metastatic breast cancer (2012-2017) was analyzed using IBM MarketScan and Medicare claims.
• Machine learning models (logistic regression, random forest, neural networks) were trained on baseline clinical and healthcare encounter data.
• Nonadherence was defined as <80% proportion of days covered in the year post-treatment initiation.

Main Results:

• Forty percent of patients were classified as nonadherent.
• All models demonstrated moderate predictive accuracy; logistic regression achieved an AUC of 0.64.
• Key predictors included age ≥55 years and specific pretreatment procedures (lymphatic nuclear medicine, radiation oncology, arterial surgery).

Conclusions:

• Machine learning models utilizing baseline administrative data can predict tamoxifen nonadherence.
• Screening at treatment initiation could enable personalized care, improve outcomes, and reduce costs.
• Further validation with enriched longitudinal data is recommended to enhance model performance.