HealthPrism^TM
Health analytics for the public sector

HealthPrism uses semi-supervised machine learning (ML) methods to train explainable predictive models. It generates modeling datasets that represent the United States population with respect to legally protected groups, including age, race, and gender. HealthPrism also extends bias checks to include representation across income groups and the Rural-Urban Continuum Codes. Additionally, each variable input into a model is examined for its distribution—with respect to age, race, gender, income, and the rural-urban continuum—to help ensure that it does not act as a proxy for any of those variables. This bias awareness and examination can also influence the strategy for model selection in which only explainable and interpretable models are used to train and predict risk.

HealthPrism is committed to maintaining the privacy of personal data, protecting user data and maintaining transparency with strict controls around data usage:

• Secure and controlled data: HealthPrism is controlled, encrypted, and protected through a Confidential Information Management Plan (CIMP) and is FedRAMP ready.
• Nationwide opt-in/opt-out access: HealthPrism complies with the strictest data privacy act in the US, the California Consumer Privacy Protection (CCPA) Act. HealthPrism only includes data that individuals have elected to share and provides the option for individuals to opt out of our dataset.
• No data from scraping or monitoring: We use consensual data and never process data that has been sourced through internet data scraping or monitoring.

HealthPrism data reflects predictive estimates based on multiple demographic and consumer marketing data sources. Underlying data and risk estimates are updated on a regular basis (typically every one to two months, depending on the source). In other words, the data in this dashboard is updated on a consistently recurring basis as opposed to a real-time basis.

Populations that suffer disproportionately high rates of death and disease—such as racially and ethnically diverse populations, people with disabilities, and rural populations—are underrepresented in some health research, and their status as part of those groups is often not recorded in critical health care records. This lack of data is a serious barrier to advancing equity.

Race and ethnicity data are particularly essential to understanding and improving health equity. However, this data on much of the US population is missing from many health-related datasets. Race and ethnicity are often poorly documented in commercial health records like insurance claims. Additionally, gaps in data and outdated information in public data sources like the US Census can hinder meaningful insights into race and ethnicity. These gaps can prevent recognition very real health disparities and delay action to improve equity.

To enable meaningful action on health equity, HealthPrism uses predictive models based on the Bayesian Improved Surname and Geocoding Method (BISG) to fill the gaps in missing race and ethnicity data (this method is used by government agencies like the Virginia Department of Health). In addition, HealthPrism uses other data science methods to improve race and ethnicity predictive model accuracy. HealthPrism includes self-reported race in which such data is available and displays predicted race where no self-reported data is available.