Modeling the Impact of Social Determinants on Breast Cancer Screening: A Data-Driven Approach

Guofang Ma¹Miranda G Scully²Jiahui Luo³Jiazuo H Feng¹Christine M Gunn¹Roberta M DiFlorio Alexander¹Anna N.A Tosteson¹Sally A Kraft¹Wesley Marrero^3*

• ¹Dartmouth College Geisel School of Medicine, Hanover, United States
• ²Dartmouth College, Hanover, United States
• ³Dartmouth College Thayer School of Engineering, Hanover, United States

Background: This study addresses the critical science challenge of operationalizing social determinants of health (SDoH) in clinical practice. We develop and validate models demonstrating how SDoH predicts mammogram screening behavior within a rural population. Our work provides healthcare systems with an evidence-based framework for translating SDoH data into effective interventions. Methods: We model the relationship between SDoH and breast cancer screening adherence using data from over 63,000 patients with established primary care relationships within the Dartmouth Health System, an academic health system serving northern New England through seven hospitals and affiliated ambulatory clinics. Our analytical framework integrates multiple machine learning techniques including light gradient boosting machine, random forest, elastic-net logistic regression, Bayesian regression, and decision tree classifier with SDoH questionnaire responses, demographic information, geographic indicators, insurance status, and clinical measures to quantify and characterize the influence of SDoH on mammogram scheduling and attendance. Results: Our models achieve moderate discriminative performance in predicting screening behaviors, with an average Area Under the Receiver Operating Characteristic Curve (ROC AUC) of 71% for scheduling and 70% for attendance in validation datasets. Key social factors influencing screening behaviors include geographic accessibility measured by the Rural-Urban Commuting Area, neighborhood socioeconomic status captured by the Area Deprivation Index, and healthcare access factors related to clinical sites. Additional influential variables include months since the last mammogram, current age, and the Charlson Comorbidity Score, which intersect with social factors influencing healthcare utilization. By systematically modeling these SDoH and related factors, we identify opportunities for healthcare organizations to transform SDoH data into targeted, facility-level intervention strategies while adapting to payer incentives and addressing screening disparities. Conclusions: Our model provides healthcare systems with a data-driven approach to understanding and addressing how SDoH shape mammogram screening behaviors, particularly among rural populations. This framework offers valuable guidance for healthcare providers to better understand and improve patients’ screening behaviors through targeted, evidence-based interventions.