Prof develops AI with goal to better predict diabetes
You’ve just been diagnosed with prediabetes: Wouldn’t you want to know if you were in danger of actually getting diabetes?
Wouldn’t you want to know if the recommended intervention would actually benefit you, especially if the interventions are medications that cost money and could have side effects?
A University of Texas McCombs School of Business study used artificial intelligence to better predict whether someone with a prediabetes diagnosis is at risk for developing Type II diabetes.
Professor Maytal Saar-Tsechansky, who teaches information, risk and operations management at the university, developed the algorithm to predict which patients are most likely to benefit from prevention measures such as medication or one-to-one coaching, and which ones do not need such interventions because they are unlikely to develop diabetes.
Saar-Tsechansky used the electronic health records of almost 90,000 people in Israel who had a prediabetes diagnosis from 2003 to 2012. The health records had data on body measures such as height, weight, lab tests, disease diagnoses, prescriptions and demographics.
Then, Saar-Tsechansky used a computer to sort out who was given preventative medicine, such as metformin, and whether they developed diabetes or not.
Putting the algorithm to the test, Saar-Tsechansky estimates the computer learning would have prevented 25% more diabetes diagnoses than using the current Framingham diabetes risk score developed by the U.S. National Institutes of Health – a score many doctors use.
“These personalized predictions really leverage intricate details,” she said. “It allows us to make more accurate predictions.”
Saar-Tsechansky used Israeli data because of the information in the public health records. Countries that have universal health care systems like Israel are more likely to have a person’s whole health history in their health records, she said, because there aren’t frequent changes in health systems or changes in insurance carriers that people in the U.S. regularly experience.
Of universal health systems, she said, “that’s a treasure. At a population level, you can find patterns.”
Using machine-learning decision making could lead to cost savings and better health resource allocations because interventions wouldn’t be wasted on people who were unlikely to become diabetic, Saar-Tsechansky said, and the right interventions could be targeted to the type of people who would respond well to those interventions.
“We need to be more proactive,” Saar-Tsechansky said. When it comes to using medication to prevent diabetes, “we need to know who should we give it to and who shouldn’t we.”
There were some basic differences between the people whom the AI model noted should have interventions and the clinical score indicated for interventions. In general, the machine identified people with a higher body mass index and people who were older.
What Saar-Tsechansky developed could be used for any disease using any set of health records, she said.