BACKGROUND: Financial toxicity (FT), the economic stress from medical care, is common among people with cancer and is associated with worse health outcomes. While risk factors for FT are known, personalized FT risk prediction tools to help mitigate FT are lacking. METHODS: This study developed and evaluated machine learning models to predict FT risk using data from the Medical Expenditures Panel Survey-Experiences with Cancer Survivorship Supplement for patients undergoing or within one year of cancer treatment. FT was defined as the presence of > 1 of the following: bankruptcy, unpaid medical bills, payment concerns, or debt. Several models were trained using demographic, clinical, economic, and social variables. Fine-tuning was performed to enhance sensitivity in predicting FT. The Shapley additive explanations (SHAP) were used for interpretability of the model. RESULTS: Among 793 people with cancer, 283 (36%) experienced FT. A fine-tuned random forest algorithm achieved an AUROC of 0.84 (95% CI = 0.78 to 0.91) and an accuracy of 0.78 (95% CI = 0.71 to 0.85), with a sensitivity of 0.84 (95% CI = 0.72 to 0.92) and specificity of 0.75 (95% CI = 0.66 to 0.83), demonstrating balanced classification performance. SHAP values identified key predictors of FT risk, including younger age, lower income, higher medical expenditures, and poorer health status. To support clinical implementation, we developed a web-based FT risk calculator (https://hd-research.shinyapps.io/ftriskcalc/). CONCLUSION: The fine-tuned random forest algorithm resulted in promising results for predicting personalized FT risk. Integrated into a web-based calculator, the model has strong potential to help mitigate FT by identifying high-risk patients early in the cancer care continuum.
Damgacioglu et al. (Fri,) studied this question.