Computable Social Factor Phenotyping Using EHR and HIE Data

Social factors are an important driver of morbidity, mortality, utilization, disparities, and healthcare costs. Measurement of these factors is a key component of defining patient risk and understanding population health. Improving their collection can lead to improvements in risk-predication models, identification of those who need social services, and allow better understanding of underlying disparities in population health. While recognizing their utility and value, many health systems struggle in their collection. Today, collecting social factors is generally done via questionnaires and is associated with challenges such as time constraints, workflow issues, and stigma.

This research will assess the validity of patient-level computable social factor phenotypes to predict a patient’s risk of increased healthcare costs and utilization. Computable phenotypes are created by analyzing collections of data elements and defining them as single data elements. For example, even when a patient is not coded as having diabetes, those with diabetes can be identified through a ‘phenotype’ that combines laboratory results and prescribed medications. This research will extend the concept of condition phenotypes to social factor phenotypes by determining the validity and usefulness of six social factor phenotypes computed from already-collected information within electronic health records (EHRs) and health information exchanges (HIEs).

The specific aims of the project are as follows:

• Assess the concurrent validity of patient-level computable social factor phenotypes.
• Assess the predictive validity of patient-level computable social factor phenotypes.
• Assess the reliability (bias) of patient-level computable social factor phenotypes across patient gender, race, ethnicity, and age.

The concurrent validity of three approaches to measuring social factors will be estimated and compared: 1) computed phenotypes of six social factors utilizing structured data obtained from EHRs and HIEs; 2) a 42-item self-administered, vendor-developed, EHR-embedded questionnaire; and 3) natural language processing of free-text clinical notes and documents from the EHRs. A longitudinal panel of adult primary care patients will be used to determine the predictive validity of the approaches in identifying subsequent utilization and cost. Two outcomes will be considered: the count of emergency department visits and the count of hospitalizations in the 6 months post-social factor measurement. Finally, the reliability or bias of patient-level computable social factor phenotypes across patient gender, race, ethnicity, and age will be assessed. By increasing the ease of collecting social factors, this research holds the potential to improve the prediction of a patient’s risk of increased healthcare costs and utilization.