Our generalizable framework included the following steps: data pre-processing and variable construction, model selection and implementation, performance evaluation, and model explainability analysis. We detailed these steps and presented their execution for predicting the probability of mammogram screening behaviors, including both scheduling and attendance.

The comprehensive methodological detailed for each framework component, including specific algorithms, parameter settings, and validation procedures, were provided in Supplementary Methods M1–M4. While this detailed framework was designed for broader application across healthcare systems, we demonstrated its implementation through a specific case study within the Dartmouth Health System.

While the framework was designed to be generalizable across different healthcare systems, we applied it specifically to the Dartmouth Health System to demonstrate its practical utility and effectiveness in a real-world setting. Dartmouth Health was an academic health system serving patients across northern New England and nearby communities through seven community hospitals, affiliated ambulatory clinics, and the academic facility Dartmouth Hitchcock Medical Center (19). The system encompassed facilities across Vermont and New Hampshire and utilized an integrated Epic electronic health record (EHR) system that enabled standardized data collection across most clinical sites (19). With over 16,000 employees including 2,300 providers, the system delivered approximately 3 million outpatient visits annually and was recognized as a nationwide leader in rural health (19). This application allowed us to assess the framework’s ability to generate actionable insights within a defined healthcare context before broader implementation in diverse healthcare environments.

To assess the robustness of our findings to different analytical assumptions, we conducted sensitivity analyses focusing on missing data handling approaches. Specifically, we performed complete case analyses using only patients with complete SDoH questionnaire data as sensitivity checks for our primary imputation-based approach. These analyses used identical modeling frameworks and performance evaluation metrics as described in Section 2.1 to ensure comparability with our primary results.

Additionally, we conducted comprehensive secondary analyses to provide deeper insights into factors influencing mammogram screening behaviors, including age-stratified evaluations, SDoH-only models, patient-level models, and clinic-level analyses. Detailed methodologies and results for all secondary analyses are presented in Supplementary materials S1.1–S1.5 (Scheduling analyses) and Supplementary materials S2.1–S2.5 (Attendance analyses).

Our analysis of SDoH questionnaire data revealed substantial variation in response completeness across the 37 administered questions. Missingness rates ranged from 10.2 to 92.4%, with a median missingness of 73.8% across all questions (Supplementary Table S1). This evaluation identified 11 questions that exceeded our pre-established 80% missingness threshold, which were subsequently excluded from model development to ensure implementation reliability. The excluded questions primarily addressed sensitive domains such as mental health status, substance use behaviors, and detailed information regarding past scheduling experiences.

Examination of the dataset revealed distinct patterns in both scheduling and attendance behaviors. Scheduling rates, calculated as the proportion of all eligible women aged 50–75 who had a mammogram scheduled, showed substantial variation across clinical sites (4.4–21.3%), insurance types (Medicare: 16%; Commercial: 12.9%), age groups, and neighborhood deprivation levels. For attendance, missed appointment rates varied by clinical site (1.9–9.1%), insurance status (Medicaid Managed: 13%; Blue Cross: 4%), and racial demographics (Asian: 0.8%; Hispanic: 8.1%) (Supplementary Table S4). We found a linear relationship between neighborhood deprivation and missed appointments (ADI 1: 1.6%; ADI 10: 11.4%) and higher attendance in urban areas compared to rural settings. SDoH questionnaire responses indicated that housing instability (multi-residence: 12.4% vs. single-residence: 4.8% missed appointments), transportation barriers (unable to work due to transportation: 18.8% vs. no barriers: 4.9%), food insecurity (often: 14.3% vs. never: 4.9%), and health literacy challenges were associated with lower scheduling rates and higher missed appointment rates (Supplementary Tables S2, S3, S5).

Our analysis revealed that the relationship between social determinants and screening behavior involved multiple interacting factors. The light gradient boosting model’s ability to capture these relationships (AUC = 0.709), together with similar performance from the random forest approach (AUC = 0.702), suggested that accounting for non-linear interactions between social drivers might help healthcare systems better understand scheduling behavior patterns.

Consistent with established literature, our models confirmed that geographic accessibility (RUCA) and socioeconomic factors (ADI state rank) influence screening behaviors (5, 8). However, our variable importance analysis revealed a hierarchy of influence that differs from traditional approaches focused primarily on individual-level social barriers. Temporal factors (months since last mammogram) and organizational factors (clinical site) emerged as the strongest predictors, suggesting that healthcare systems might achieve more immediate impact through system-level interventions rather than attempting to address individual patients’ social circumstances. This finding highlighted a different priority than much of the existing mammography literature, which emphasized individual-level barriers such as transportation, cultural beliefs, and health literacy (6, 9, 10). While these individual SDoH factors remained important in our descriptive analyses, our machine learning approach revealed that facility-level variations and care patterns were more predictive of scheduling and attendance behavior. This pattern was particularly true for traditional SDoH questionnaire response, where individual measures such as homelessness, food insecurity, and financial hardship each contributed less than 0.01% to variable importance. This minimal predictive power might have reflected the substantial missingness in SDoH data, potential underreporting of sensitive information in clinical settings, or that geographic and organizational indicators served as more reliable proxies for underlying social determinants.

For Dartmouth Health specifically, this suggested that standardizing practices across clinical sites might yield greater improvements in screening rates than traditional patient-education or transportation-assistance programs. For attendance behavior, the evidence suggested a dual approach combining system-level standardization with targeted interventions. This recommendation was supported by our clinic-level analysis for scheduling (Supplementary material S1.4), which confirmed substantial performance variation across sites. However, the clinic-level analysis for attendance (Supplementary material S2.4) showed inconsistent results, limiting conclusions about organizational effects.

Our findings also illuminated the complex interplay between organizational and social factors that traditional regression approaches often missed (11). The elastic-net logistic regression model’s performance in capturing attendance patterns (AUC = 0.698) demonstrated that patient health complexity (General Adult Risk Score) and digital engagement (portal activity) were critical factors that complement traditional socioeconomic predictors. This insight provided healthcare systems with a more nuanced understanding of how to target interventions across different patient populations.

Beyond these specific findings for breast cancer screening, our systematic framework laid the groundwork for analyzing SDoH’s influence on other preventive health behaviors, demonstrating the potential for broader applications in improving routine preventive care utilization. The methodological approach of combining individual-level social determinants with organizational and temporal factors could be adapted to examine colorectal cancer screening, cervical cancer screening, and other preventive services where similar complex interactions between social drivers and healthcare delivery factors likely influence patient engagement.

Our analytical approach offered several methodological contributions to healthcare delivery and the implementation science. First, we demonstrated a novel approach to operationalizing SDoH in breast cancer scheduling and attendance practices, providing healthcare systems with a framework to translate social determinant screening tools into actionable screening strategies. Unlike previous work analyzing nationwide census tract-level scheduling rates and focused primarily on geographic accessibility and demographics, our study examined individual-level data integrating clinical, behavioral, and social determinants within a healthcare system context (28). This focus allowed us to identify specific patient-level factors that directly influence scheduling decisions, rather than ecological correlations at the population level.

Second, our modeling framework effectively functioned as a poly-social risk score system, aggregating multiple social determinants to quantify their combined influence on screening adherence. This approach moved beyond examining isolated social determinants to consider how they collectively impact health behaviors. Third, the age-stratified analysis (Supplementary material S1.2) revealed important variations in these poly-social risk profiles across demographic groups, suggesting the need for age-specific implementation strategies that account for different SDoH impacts across the lifespan. Finally, our application of light gradient boosting models and elastic-net logistic regression models and partial dependence plots revealed important non-linear patterns in the relationship between months since the last mammogram and non-scheduling probability—a critical insight that traditional regression approaches would likely miss. For the attendance model, our elastic-net logistic regression approach similarly captured complex relationships between organizational factors, patient characteristics, and social determinants, though with different key predictors than the scheduling model. These findings demonstrated the value of machine learning approaches in capturing complex relationships between certain social determinants and screening behavior.

The development of a unified modeling framework that incorporated both individual-level social drivers and system-level factors provided healthcare organizations with a template for analyzing their own screening programs. This approach could be particularly valuable as healthcare systems work to improve cancer screening rates for their medically-homed populations while effectively integrating SDoH data into their quality improvement initiatives.

The substantial missingness in our SDoH questionnaire data, with a median of 73.8% across questions, reflected common implementation challenges in clinical settings. Our analysis excluded 11 questions that exceeded the 80% missingness threshold, which primarily addressed sensitive domains such as mental health status, substance use behaviors, and detailed information regarding past scheduling experiences. Health literacy, a factor that prior research has demonstrated to influence mammography screening adherence (6), represented another domain affected by substantial missingness, restricting our ability to comprehensively assess its influence on patient screening decisions. These exclusions represented a methodological consideration because these sensitive domains might be critical drivers of patient decision-making regarding mammogram scheduling and attendance. More complete data on these domains could potentially alter our understanding of the factors driving screening behavior within healthcare systems. These data collection challenges highlighted the need for alternative approaches to capture important behavioral determinants. However, our framework was designed to be adaptable and can incorporate these variables when improved collection methods make such data available in future implementations.

These implementation challenges underscored the importance of systematic approaches to translating our findings into practice. The Consolidated Framework for Implementation Research (CFIR) offered a valuable lens for future efforts to translate our findings into practice. Though our current work focused on quantitative modeling rather than a full CFIR implementation, our findings provided a foundation for subsequent mixed-methods approaches that could more fully leverage implementation science frameworks.

For example, the facility-level variations identified in our model aligned with CFIR’s ‘inner setting’ domain, suggesting that organizational culture and readiness for implementation played important roles in both scheduling and attendance behaviors. Future work could build on our quantitative findings by using qualitative methods to explore how these organizational factors influenced practices and how interventions might be tailored to different clinical settings. Similarly, our findings related to geographic and socioeconomic factors corresponded to CFIR’s ‘outer setting’ domain, highlighting the importance of understanding patient needs and resources within their communities. Further investigations could provide deeper insights into how these community factors shaped decisions and how healthcare systems might better address them.

As healthcare systems consider implementing SDoH-informed interventions, CFIR and other implementation science frameworks could provide valuable guidance for assessing feasibility, sustainability, and potential barriers. Our work represented an important first step in this direction by providing quantitative evidence of key relationships that future implementation efforts should consider.

Our model was developed and validated within the Dartmouth Health system, which served a population with limited racial, ethnic, and linguistic diversity. This demographic homogeneity might have limited our ability to capture important language-related barriers to screening access and communication, and might have restricted the generalizability of our findings to more diverse healthcare settings and populations. Although our findings indicated small racial differences, future work should validate these approaches in healthcare systems serving more diverse communities to ensure broader applicability. It was important to note, however, that we had designed our work as a generalizable framework that could perform well in other situations and could incorporate more diverse racial groups and other demographic aspects if the necessary data were available.

Moreover, our modeling approach assumed that the ratio between screening and non-screening populations would remain stable over time. This assumption might not have held in different implementation contexts or as scheduling programs evolve. Healthcare systems implementing similar approaches should carefully consider their local population characteristics and mammogram scheduling patterns. Additionally, while our model demonstrated modest predictive performance within our system, its generalizability to other healthcare settings might be limited by differences in organizational structure, population characteristics, screening protocols, and the substantial missingness in our SDoH data, which might have limited our ability to fully capture social determinant influences. Future research should explore how these models could be adapted and calibrated for different healthcare contexts.

Beyond expanding the population and removing assumptions, we identified several priority areas for future research. First, the development of dynamic modeling approaches that could adapt to changing population characteristics and scheduling patterns would enhance the robustness of our framework. Additionally, integrating SDoH-informed scheduling models with other preventive care programs could create more comprehensive implementation strategies. Investigation of facility-level variations in scheduling patterns would have further identified best practices for implementation. Finally, extending our analytical framework to other scheduling programs, such as colorectal and cervical cancer scheduling, would increase the broader applicability of SDoH-informed modeling approaches and strengthen the overall impact of our generalizable framework across diverse healthcare settings.

Our study provided healthcare systems with a data-driven approach to understanding and addressing how social determinants shape breast cancer scheduling practices. Our findings suggested that machine learning approaches could help healthcare systems develop more effective, targeted implementation strategies. As healthcare systems worked to meet cancer screening targets for their medically-homed populations, approaches that systematically analyzed and addressed social determinants of health could have become increasingly valuable for improving adherence and reducing disparities.

For the scientific community, these findings offered two primary contributions. First, our results demonstrated the relative influence of different predictors on screening behaviors, highlighting that healthcare systems might achieve greater impact by focusing on factors within their direct control rather than attempting to address individual patients’ social circumstances alone. Second, our framework enabled identification of patients at highest risk of not scheduling or attending appointments, providing a practical tool for targeted intervention strategies.

Looking ahead, our quantitative findings provided a foundation for future implementation science approaches that could more fully leverage frameworks like CFIR to translate these insights into practice. By combining machine learning approaches with implementation science, healthcare systems could develop more comprehensive strategies for addressing the complex interplay between social determinants and screening behaviors, ultimately improving health outcomes for diverse patient populations. Healthcare systems and researchers could adapt this approach using their own data to develop targeted interventions and improve mammography adherence within their specific patient populations and organizational contexts. Through such systematic approaches to understanding and addressing screening behaviors, healthcare system could potentially work toward more effective, evidence-based strategies for reducing disparities and improving preventive care delivery.