Using AI social determinants of health to identify risk
There is increased recognition for better methods to measure, predict adjust for social risk factors in healthcare population health. Current performance quality measures generally do not take SRFs into account in the bonus/penalty structure, nor are SRFs generally included in most risk adjustment formulas.
This can lead to unintended consequences, including the potential to perpetuate bias disparities in health outcomes. To mitigate these issues, RTI International is developing an “artificially intelligent” approach to risk adjustment for SRFs using random forests to understlife expectancy variances at the census tract level.
Where AI comes in
So how can AI help recognize how local area factors are independently associated with many health outcomes may be informative either in conjunction with individual-level data or on their own?
“While individual-level predictors are best for individual-level analyses, neighborhood-level conditions are often independently associated with potentially avoidable costs undesirable health outcomes at the population level,” said Lisa M. Lines, senior health services researcher at RTI International.
“In our approach to creating geospatial, small-area indices based on supervised machine learning to predict population outcomes, we have found that we can explain between 73 99% of variation in outcomes using publicly available data on social determinants of health.”
With a combination of statistical data science methods, researchers can drill down into different domains of SDOH start to understwhat is driving local variation in population outcomes such as life expectancy at birth or infant mortality, she added.
“One specific tool we have used is variable importance from random survival forests,” she explained. “However, VI may not provide accurate insights if VIs are split across several highly correlated predictors. VI is a tool for understanding error rates, but what is important to an algorithm’s accuracy predictive power may be less important than understanding what factors are most actionable and/or policy-relevant.
“This suggests two crucial ingredients in geospatial, SDOH-focused machine learning: a sound conceptual model a team of experts in multiple subjects – health geography, statistics, data science, health services research, health policy economics, more – involved from the beginning.”
AI improves incentives for providers
Lines says that artificial intelligence tools improve incentives for providers to treat more difficult patients. The question is how?
“The idea is grounded in principles of risk adjustment,” she explained. “In healthcare, risk adjustment has been used for decades in managed care to provide incentives to both plans providers to care for sicker patients. What has only recently begun to be accepted in health systems is the outsized role of social factors in health outcomes.
“In every setting disease area, social environmental factors like unemployment, lack of education, substandard housing, food insecurity lack of transportation can affect how difficult it is for people to access care follow a treatment plan, often, how sick people are.”
These factors are largely outside of providers’ control, despite efforts in recent years to connect people to social services through contacts with their health insurance plan or healthcare provider.
“One general principle of quality improvement is that individuals entities should not be held accountable for factors outside their control,” Lines said. “Thus, the field of risk adjustment is beginning to look at including social factors in payment formulas in some way.
“Similarly, the broader field of quality measurement is starting to think about stratifying and/or controlling for social risks when looking at quality-related outcome measures,” she continued. “The idea is to avoid punishing providers who see people with more complex or serious health-related needs – whether those needs are clinical, social, or both.”
A tension arises between the desire to better adjust for social factors the available data, she noted.
“Some people are, understandably, reluctant to provide their personal information to a faceless bureaucracy,” she said. “Instead, publicly available, anonymized data can be used to shed light on neighborhood factors without requiring the difficult, labor-intensive expensive task of collecting private information from millions of individual patients.”
Healthcare outcomes variations
On another front, commonly available area-level deprivation or vulnerability indices only partially explain the variation seen in healthcare outcomes.
“Of the publicly available indices we have studied, none were specifically designed or tested the way ours was – using more than 150 variables a random forest model to predict life expectancy,” Lines explained. “Instead, these older indices were developed using approaches like factor analysis qualitative component analysis.
“Studies have been published showing how these indices are associated with various outcomes, but none have shown the amount of variance explained in a core population health outcome like life expectancy at birth, to our knowledge.”
Now, with the power of modern computing, researchers can go beyond indices based on a handful of broad measures.
“We can get closer to understanding what is driving social health inequities by exploring the variation in life expectancy at the census tract level, which was only recently estimated for the first time,” she noted.
“With finer-grained data, we are able to show the relative contributions of food, housing healthcare-provider availability, for example, to the variation in life expectancy at the local level. Future efforts are needed to understthe value of AI in producing actionable data to policymakers others interested in improving health social equity.”
Lines will offer more detail during her HIMSS21 session, At-Risk Identification Using AI Social Determinants. It’s scheduled for August 11, from 11:30 a.m-12:30 p.m. in Venetian Marcello 4501.