Prediction of Low-VALUE Cancer Care Among Older Men with Low-Risk Prostate Cancer: A Machine Learning Approach
Author(s)
Fiano R1, Merrick G2, Innes K3, LeMasters T4, Mattes M5, Shen C6, Sambamoorthi U6
1West Virginia University, Wheeling, WV, USA, 2Urologic Research Institute, Wheeling, WV, USA, 3West Virginia University, Morgantown, WV, USA, 4West Virginia University, School of Pharmacy, Morgantown, WV, USA, 5Rutgers Cancer Institute of New Jersey, New Brunswick, NJ, USA, 6Penn State College of Medicine, Hershey, PA, USA
Older men with incident prostate cancer are vulnerable to low-value prostate cancer treatment. Despite evidence-based support for conservative management (i.e., non-treatment), approximately 2 in 3 Medicare beneficiaries receive treatment for low-risk prostate cancer. Patients with multimorbidity who experience care fragmentation are vulnerable to departures from evidence-based medicine. A comprehensive analysis of clinical and non-clinical factors, such as life expectancy and care fragmentation, that may drive low-value prostate cancer treatment is lacking. OBJECTIVES : Use machine learning (ML) to identify leading predictors of cancer treatment within 12 months of diagnosis among older men with low-risk prostate cancer. Novel predictors included validated prostate-cancer specific life expectancy and care fragmentation. METHODS : In this retrospective cohort study we linked Surveillance, Epidemiology, and End Results cancer Registry (SEER), Medicare Claims, Census, and Area Health Resource files and included older men with incident low-risk prostate cancer from 2009 to 2014 (n=13,870). We used claims data to identify treatment (Yes/No) in the first 12 months after diagnosis. We used the XGboost algorithm and Shapley additive explanations (SHAP) to rank feature importance in treatment prediction. RESULTS : In our study cohort (n=13,870), 66.9% of older adults received cancer treatment. Age, multimorbidity, care fragmentation, social support, and social determinants were leading predictors of cancer treatment (Accuracy=0.70, Precision=0.71, Recall=.92, Precision-Recall Area Under the Curve = 0.78). Relationships of college education, income, and care fragmentation on low-value cancer treatment were nonlinear and complex. Life expectancy was a weak predictor of prostate cancer treatment. CONCLUSIONS : Our results suggest that non-clinical factors such as social determinants, care fragmentation, and social support are the most important predictors of treatment among older men diagnosed with low-risk prostate cancer. Despite a critical role in evidence-based treatment recommendations, life expectancy had limited impact on treatment selection.
Conference/Value in Health Info
2021-05, ISPOR 2021, Montreal, Canada
Value in Health, Volume 24, Issue 5, S1 (May 2021)
Code
PCN159
Topic
Health Service Delivery & Process of Care, Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics, Disease Management, Treatment Patterns and Guidelines
Disease
Oncology