Management of diabetes is a daily challenge facing nearly 40 million Americans. These include engaging in food intake tracking, timing medications and physical activity. If you make that mistake, it can lead to serious health issues. Therefore, developing better predictive tools is an important part of effective diabetes care.
To support better diabetes management, researchers, funded by multiple National Science Foundation grants, are developing innovative tools that help patients predict blood glucose levels more accurately without compromising the privacy of their health data. This cutting-edge approach can change the way people with diabetes monitor and manage their condition in real time.
At the core of this technology is a method known as federation learning. This allows artificial intelligence models to train on many patient devices without sending personal data to a central server. This setup is ideal for healthcare where data privacy is paramount and patients often use smart devices that are limited to batteries and memory. However, early federal learning systems struggled to adapt to individual differences, such as how people eat, move, and respond to insulin.
To address this challenge, the researchers grouped patients based on carbohydrate (sugar and starch) intake levels. The idea is that people who eat in similar ways tend to exhibit similar glucose patterns. By training AI on these grouped behaviors, the model has become effective in making personalized blood glucose predictions.
To test their approach, the team evaluated two machine learning models using data generated from an FDA-approved Type 1 diabetes simulator. The model's accuracy was improved as simulated data was accumulated. In particular, the system can build personalized models even with limited inputs. This is an important benefit for newly diagnosed patients and for patients who have begun to manage their care using digital tools.
Federated learning offers a solution suitable for this field, as traditional AI systems typically require large amounts of data to be collected in central locations, which can pose privacy risks, especially when dealing with sensitive health information. Not only does it hold personal data on each individual device, such as mobile phones and wearable sensors, and shares only model learning, but there is no raw data. This will allow the system to improve over time and protect patient privacy.
Although early results are promising, researchers point out that the model still relies on detailed food intake data. Not all patients can easily provide. They plan to expand their system to test it in a larger group of patients, including other factors such as exercise and medication. In the long term, researchers hope to extend AI approaches that provide this personalized privacy for other chronic diseases, such as heart disease and asthma, which are equally important.
With diabetes costs more than $300 billion a year for the US economy, previous interventions and innovations that enable personalized care can reduce long-term costs and improve population health outcomes.
The project highlights how public investment in cutting-edge research drives innovation that benefits not only individual patients but also the entire US health system.
