Researchers in Italy have demonstrated a no-cost, AI-based technique for detecting the presence of previously undiagnosed abnormalities in blood sugar stability. Such abnormalities, known collectively as dysglycemia, can precede the onset of type 2 diabetes.
Enrico Buccheri, MD, of the University of Catania and colleagues describe their work in a study posted Feb. 27 in Diabetes Research and Clinical Practice, which is published by Elsevier for the International Diabetes Federation.
The team used data on 47 variables of health status captured over 10 years to develop their method, which applies Darwinian evolutionary theory to health data.
Commenting on this unlikely aspect, the authors write:
Our novel approach implements a well-known scheme based on the theories formulated by Charles Darwin about 200 years ago. Through the mechanism theorized by Darwin, a heterogeneous population evolves towards those individuals whose characteristics better adapt to the environment. Similarly, in the evolutionary computing, the Darwinian scheme is simulated in order to maximize a generic function called fitness. The latter numerically describes the capabilities of an individual to fit with the environment.
In testing and validation, the team found its system capable of detecting dysglycemia using only two of the 47 variables—age and waist circumference—with good sensitivity and specificity.
Comparing their model’s performance with that of the current gold standard for dysglycemia screening, Buccheri and colleagues found little difference between the two.
“Despite its outstanding simplicity, the accuracy of our model turns out to be equivalent to that of more complex tools previously published in the literature and widely used to perform cross-sectional studies,” the authors comment.
They suggest their AI-based innovation could help screen for dysglycemia at the population level while incurring little to no costs.
The advance, they write, “also indicates that the use of approaches based on novel artificial analysis techniques might lead to a larger discovery power than traditional health data analysis methods widely adopted in the literature.”