This article first appeared January 16, 2018 on HealthLeaders Media.
This technology can accurately predict 30-day readmissions and clarify the AI process for clinicians.
“A lot of the resistance from clinicians … is because people don’t quite understand how [AI analytics] work,” says Kamal Jethwani, MD, MPH, senior director of Connected Health Innovation at Partners HealthCare.
That’s why Partners Connected Health developed technology with Hitachi to create what it calls “explainable AI,” which clarifies the AI process for users in an effort to make it easier to understand, and ultimately, trustworthy, all toward a goal of adoption.
How it works
The technology assigns scores that can accurately predict the risk of 30-day readmission for patients with heart failure. Their research has shown this technology has the potential to save $7,000 per patient per year. What makes this system different than other risk prediction systems is the transparency behind the score, says Jethwani.
“Very rarely are [clinicians] able to figure out why that score is what it is,” he says. As a result, clinician decision-makers often don’t know whether to trust the technology.
With explainable AI, on the other hand, clinicians not only see the score, but also the top contributing factors behind that score.
“It’s telling us why a person is likely to come back to the hospital,” Jethwani says.
Moreover, he says that the “why” behind the readmission prediction score is actionable for the clinician and the care team. Because of this, the algorithm can contribute to better care, not just better predictions, Jethwani says.
For example, the new technology uses an algorithm that considers roughly 3,000 variables that might contribute to readmission, drawing from both structured and unstructured data, from sources ranging from medication adherence, to clinic notes, to imaging reports.
“Adding this level of detail into the algorithm is improving its accuracy,” Jethwani says.
Each patient has their own unique set of variables that determine their readmission score.
While there are variables that clinicians have no control over, such as age or gender, others can be improved with intervention, allowing clinicians to see which actions can help which patient.
The tool can also help clinicians choose which patients are the best candidates for participating in readmission prevention programs following hospital discharge.
“We want this to be used as an important data point in their decision-making,” Jethwani says, while also “respecting their clinical judgement.”
To test the technology, the Partners Connected Health Innovation team simulated the readmission prediction program among heart failure patients participating in the Partners Connected Cardiac Care Program, a remote monitoring and education program designed to improve the management of heart failure patients at risk for hospitalization.
Results were compared to data from approximately 12,000 heart failure patients hospitalized and discharged from the Partners HealthCare hospital network in 2014 and 2015.
The analysis showed the prediction algorithm achieved a high accuracy of approximately AUC 0.71, and can significantly reduce the number of patient readmissions. (AUC, area under the curve, is a measure of prediction model performance with an ideal value range from 0 to 1.)
Jethwani says the next step is to present the technology to clinicians to figure out how it can best be used and to determine where in the medical record and workflow clinicians prefer to see the data.
“As people are designing analytics and smart systems [it’s important], to really think about end user engagement first,” Jethwani says.