Predicting overall survival in adults with cancer in the US using machine learning approaches integrating comprehensive social risk factors.

Abstract

1638 Background: Adults with cancer in the U.S. face an elevated mortality risk compared to the general population, with social risk factors playing a critical role – particularly among those with comorbidities. However, traditional mortality risk prediction models often focus on treatment exposures and basic demographic factors, overlooking social risk factors. We aim to develop a machine learning (ML) model that integrates comprehensive social risk factors with traditional predictors to predict overall survival for adults with cancer in the U.S. Methods: We analyzed data from 6,181 nationally representative adults diagnosed with cancer from the National Health Interview Survey (NHIS; 2013-2014). A total of 74 risk factors, including basic demographics (e.g., age at the survey, sex, marital status, body mass index [BMI]), personal and household socioeconomic status (SES; e.g., education, food insecurity), lifestyle, social support, and health status (e.g., chronic health conditions [CHCs], disability), were included in modeling. The primary endpoint was 5-year overall survival from the survey completion date, with secondary endpoints of 1- and 2-year survival. Death from any cause after the survey was defined as an event, and subjects were censored 5 years post-survey. The sample was randomly split into 70% training and 30% testing. A random survival forest (RSF) model predicted survival. The time-dependent area under the receiver operating characteristic (AUROC) curve and the Brier score (BS) assessed the model performance. Both AUROC and BS range from 0 to 1, with higher AUROC for higher accuracy (discrimination) and lower BS for better alignment between predicted and observed risk (calibration). The Shapley additive explanations (SHAP) values were used to interpret variable importance in the established RSF model. Results: The mean age of subjects during the survey was 65.6±13.8 years, and 40.2% were male. For the established RSF model, the AUROC (mean ± standard deviation) for predicting 1-, 2-, and 5-year survival was 0.795 ± 0.026, 0.810 ± 0.018, and 0.831 ± 0.011, respectively, reflecting high and improved predictive accuracy over time. The BS for 1-, 2-, and 5-year survival was 0.039 ± 0.004, 0.065 ± 0.005, and 0.119 ± 0.005, respectively, indicating excellent calibration. The top five variables ranked by SHAP values include age at the survey (0.048), use of special equipment due to health problems (0.029), employment status (0.020), number of CHCs (0.016), and BMI (0.015). Conclusions: By integrating social risk factors with traditional risk predictors, we developed an ML model that predicts overall survival with high accuracy and excellent calibration for adults with cancer in the U.S. Identifying key risk social factors enables targeted interventions, potentially improving health outcomes and management for the adult cancer population.