Disaggregating Latino nativity using machine learning on electronic health records: Insights for colorectal cancer screening disparities

Background: Advancements in colorectal cancer (CRC) prevention have not been equitable with studies showing reduced CRC screening rates and later-stage diagnoses among Latino patients. Latino subgroups vary in their cancer-related risk factors but also differ widely in their sociodemographic characteristics, migration histories, insurance coverage, and health care access patterns that may impact cancer prevention. However, large-scale datasets seldom contain granular data needed to study Latino subgroup-specific differences in cancer risk and outcomes. This study evaluated a machine learning approach designed to infer nativity and country of birth and advance cancer prevention research to better evaluate health equity among Latinos. Methods: We used comprehensive electronic health record data from 1,500,191 Latino patients receiving care at 1,876 community health centers across 28 states, along with geocoded census-tract-level neighborhood composition data, and surname-based data to develop machine learning models of Latino subgroups. Multiple supervised learning algorithms were trained and tested to predict nativity and country of birth. Model predictive performance was evaluated using area under the receiver operating curve (AUC). As a case example, we used model-predicted probabilities of Latino subgroups and nativity to evaluate CRC screening disparities by foreign-born status, both known and predicted. Results: Among 1,500,191 Latinos in the network, country of birth was self-reported by Latino patients for only 173,278 (11.6%), underscoring the challenges of using existing EHR data for studying Latino heterogeneity. Prediction models for nativity showed excellent discriminatory prediction performance across all groups (US-born vs. foreign-born: AUC=0.90; Mexican vs. non-Mexican: AUC=0.87; Guatemalan vs. non-Guatemalan: AUC=0.84; Cuban vs. non-Cuban: AUC=0.84). In our case example, using known foreign-born status of Latino patients, we observed that US-born Latinos had lower odds of CRC screening compared to foreign-born Latinos (OR=0.55, 95% CI=0.50-0.617). We observed high concordance between known and model-predicted estimates of CRC screening odds ratios. Conclusion: National calls for data disaggregation, including among Latinos, have numerous challenges. We developed and validated novel prediction models to infer Latino nativity and country of birth for use in population-based cancer disparities research. These methods present an opportunity to evaluate cancer disparities in data where Latino nativity is not collected.