Poster #35: Tugce Kinik PhD (c), MS

Explainable Machine Learning for Predicting Functional Limitations Among U.S. Adult Cancer Survivors Using National Health Interview Survey Data

Kinik T, PhD Candidate, Department of Epidemiology and Biostatistics, College of Public Health, Temple University, Philadelphia, PA, USA

Functional limitations are common concerns among cancer survivors, yet their prediction may involve complex and nonlinear relationships that conventional regression approaches may not fully capture. We evaluated machine-learning approaches for predicting functional limitations among U.S. adult cancer survivors and used explainable artificial intelligence to identify influential predictive features. Methods: We analyzed 11,980 adult cancer survivors from the 2019–2023 National Health Interview Survey. Functional limitation was modeled as a binary outcome using demographic, behavioral, health, and cancer-related characteristics, including age, sex, race/ethnicity, education, marital status, smoking, body mass index, comorbidity burden, general health, age at cancer diagnosis, and years since diagnosis. Data were stratified into 80% training and 20% held-out testing samples. Logistic regression, Elastic Net, Random Forest, and XGBoost were evaluated using 10-fold cross-validation. Held-out performance was assessed using AUROC, precision-recall AUC, classification metrics, and Brier scores. AUROCs were compared using paired DeLong tests. SHapley Additive exPlanations (SHAP) characterized global feature importance and predictive relationships. Results: XGBoost demonstrated the highest test-set discrimination (AUROC=0.760, 95% CI: 0.738–0.782; PR-AUC=0.572), compared with Random Forest (0.755), logistic regression (0.750), and Elastic Net (0.750). XGBoost significantly exceeded logistic regression (p=0.035) and Elastic Net (p=0.032), but not Random Forest (p=0.145). General health was the dominant SHAP feature, followed by comorbidity, age, sex, BMI, and age at cancer diagnosis. SHAP analyses identified nonlinear predictive patterns for comorbidity burden and age-related characteristics. Excluding age at diagnosis did not alter discrimination (AUROC=0.760; p=0.818), whereas excluding general health reduced AUROC from 0.760 to 0.687 (ΔAUROC=0.073, 95% CI: 0.054–0.093; p<0.001). Conclusions: XGBoost modestly improved discrimination over conventional regression while identifying nonlinear predictive patterns. General health contributed substantially to prediction, demonstrating the potential utility of explainable machine learning for characterizing predictors of functional limitations among cancer survivors.