Several AI algorithms are now available to improve patient care in Type 1 and Type 2 diabetes. Some of these digital tools may even personalize treatment using multimodal data.
By John Halamka, M.D., Diercks President, Mayo Clinic Platform and Paul Cerrato, MA, senior research analyst and communications specialist, Mayo Clinic Platform and a Professor at Northeastern University.
A recent systematic review of research studies concluded: “The emergence of artificial intelligence (AI) and wearable technology has revolutionized healthcare, offering innovative solutions for diabetes management.” That bold statement may seem overly optimistic but it’s based on an in-depth analysis of 60 studies that demonstrate the right AI algorithms, combined with wearables like smartphones, continuous glucose monitors, and fitness trackers, are capable of giving clinicians and patients personalized advice, suggestions on how to adjust their diet and activity level, and help to individualize insulin dosages.
For example, investigators used the camera on the Apple Watch, in conjunction with a deep neural network (DNN), to enable early detection of diabetes in over 53,000 individuals. The smartphone camera uses a technology called photoplethysmography to detect blood flow in the tissues below a person’s skin (changes in peripheral blood flow have been successfully used for years to serve as biomarkers like oxygen saturation and heart rate variability). Avram et al combined this tool with a DNN in a clinic cohort to generate a risk score, which was significantly correlated with patients’ hemoglobin A1c levels. In the final analysis: “These findings demonstrate that smartphone-based photoplethysmography provides a readily attainable, non-invasive digital biomarker of prevalent diabetes.”
Similarly, Fraser et al have studied the value of reinforcement learning to help improve insulin dosing; they also point out that fitness trackers can be used to integrate sleep and exercise data into diabetes management. Keep in mind, however, that several of the studies summarized by Frazer’s group have shortcomings, including racial and ethnic bias, and the so-called black box dilemma, i.e., there is too little explanation of how AI algorithms function in a way that can be readily understood by the average clinician.
Another problem with many of the AI tools used to manage diabetes is that they rely too heavily on HbA1a. Carletti et al highlighted this dilemma during a prospective study of over 1,000 US patients. They analyzed multimodal data from 347 individuals, including those who had normal glucose levels, were prediabetic, or had type 2 diabetes. The patients had wearable sensors, mobile apps, data from continuous glucose monitors, dietary logs, activity metrics, and data from their gut microbiome. Among patients with the same HbA1c levels, the investigators discovered “hidden glycemia heterogenicity,” in other words, their multimodal risk scores were more closely aligned with their risk of diabetes than the HbA1c levels. Their findings suggest that this so-called gold standard for diabetes management may have its shortcomings, especially if hour-by-hour plasm glucose levels fluctuate significantly. Their work also highlights the need to take into account numerous parameters that are usually ignored when clinicians focus almost solely on glucose levels to monitor patients’ course.
New AI-enhanced biomarkers are also being explored to improve the early detection of Type 1 diabetes (T1D). Currently, the best we can do to predict the onset of insulin-dependent T1D is to monitor autoantibodies to pancreatic islet cells and calculate a person’s genetic risk score. Both biomarkers fall short. Autoantibodies only develop after a patient has already reached stage one of T1D, and 80% of patients who develop the disease have no family history or genetic risk. With these weaknesses in mind, Australian researchers have developed a biomarker that relies on microRNA. Their report concluded: “miRNA-based DRS [dynamic risk score] … offered good predictive power (area under the curve = 0.84) for T1D stratification in a separate multicontext validation dataset (n = 662), and accurately predicted future exogenous insulin requirement at one hour of islet transplantation.”
AI continues to offer hope to patients with Type 1 and Type 2 diabetes, and the latest research is likely to bring an even brighter future.
Share this:
