Primary Submission Category: Health care/services
Explainable Machine Learning Insights into Multi-Level Social Determinants of Treatment Refusal Among HPV-Associated Cancer Patients
Authors: Ryan Suk, Maryam Kheirandish,
Presenting Author: Ryan Suk*
Purpose: This study employs machine learning and SHAP (SHapley Additive exPlanations) values to examine the predictive impact of multi-level social determinants of health (SDoH) and clinical factors on the refusal of recommended treatment among HPV-associated cancer patients.
Method: A retrospective analysis was conducted using the SEER-Medicare data (2004–2017) for cervical, vaginal, vulvar, penile, anal, and oropharyngeal cancers. We incorporated individual-level socio-demographic (age, sex, race and ethnicity, marital status, insurance) and clinical factors (cancer site and stage, comorbidity) as well as zip code-level factors (education, income), census-tract-level poverty, and county-level rural-urban status. The XGBoost (eXtreme Gradient Boosting) classifier predicted treatment refusal, addressing class imbalance with adaptive synthetic (ADASYN) sampling approach. Model performance was evaluated using SHAP values.
Results: The cohort included 67,421 patients. The model achieved 98% accuracy, with F1 scores of 0.67 (radiotherapy refusal) and 0.72 (surgery refusal). Key predictors of radiotherapy refusal included cancer stage, cancer site, area-level income, comorbidity, race, and state of residence. Surgery refusal was influenced by age, cancer stage, marital status, insurance, cancer site, and area-level factors such as poverty, education, and income. Patients with localized malignancies, cervical or oropharyngeal cancer, aged 65–75, and living in areas with <20% poverty were more likely to refuse radiotherapy. Surgery refusals were more common among younger patients, those with in-situ or localized malignancies, and those in areas with low-income or low-education levels.
Conclusion: These findings highlight the role of multi-level SDoH in treatment decisions and can guide tailored interventions and policies to improve equitable cancer care. Personalizing approaches based on key predictors may help reduce disparities and improve HPV-associated cancer outcomes.