OBJECTIVES : Ovarian cancer (OC) is difficult to detect and largely goes undiagnosed until late stages. Incidence of OC increases with age, and nearly half of women are diagnosed at age 65 or greater. This study aims to compare several models for predicting OC. The goal is to identify healthcare utilization patterns that are associated with an increased OC risk that may allow for intervention and reduce time to diagnosis.

METHODS : Using a de-identified claims database from 01/2017-05/2020, Medicare Advantage enrollees with OC were identified: 2+ claims with ICD-10 C56.x between 30 and 365 days apart. Patients were required to have continuous enrollment (CE) for ≥30 months prior to first OC claim and be between 66 to 85 years old at time of diagnosis. A control cohort was selected from women with equivalent age and CE criteria. We compared five algorithms in predicting likelihood of OC: LASSO logistic regression, decision tree, random forest, XGBoost, and CatBoost.

RESULTS : A total of 876 women met study criteria for the OC cohort and 1,667 met study criteria for the control cohort. Random forest (sensitivity 0.96, specificity 0.88, precision 0.75) and XGBoost (sensitivity 0.87, specificity 0.92, precision 0.85) were the best performing models. Length of continuous enrollment, age, and pharmacy utilization in the two years prior had the highest feature importance for both models. Ambulatory utilization and total payer costs were also identified as important features in random forest and XGBoost respectively.

CONCLUSIONS : Random forest and XGBoost performed well at predicting OC in a higher-risk Medicare Advantage population and may allow development of targeted interventions using claims data for early OC screening in women with higher healthcare utilization. Future iterations will explore implications of model sensitivity and specificity; minimum enrollment criteria will be assessed, and analysis will be extended to investigate model performance in a younger commercial population.