New research highlights machine learning as a potential early risk detection strategy
Chicago, June 20, 2025 /prnewswire/ – Development from two studies highlighting the possibilities of machine learning to leverage artificial intelligence (AI) technology Improved identification of early stages of type 1 diabetes was presented as a slow broken poster session at 85th American Diabetes Association Science Session® (ada) in Chicago.
Approximately 64,000 Americans are diagnosed with type 1 diabetes each year. Nearly 40% are unaware of the illness until they experience a life-threatening event that requires hospitalization. This is because the disease can progress quietly until symptoms are present, such as excessive thirst, frequent urination, and diabetic ketosidosis. By this point, serious and often irreversible damage has already occurred in the insulin-producing cells, highlighting the need for early detection and intervention strategies.
AI models reduce false positives and increase accuracy in type 1 diabetes risk assessments up until one year before diagnosis
The results of a new study presented as part of a slow, broken symposium show the possibility that AI could more accurately identify individuals at risk for type 1 diabetes by a year before diagnosis.
The retrospective cohort study developed two age-specific machine learning models. One is for individuals aged 0-24, and the other is for individuals aged 25 and over. The researchers applied certain criteria to identify confirmed cases of type 3 diabetes, including patients who recorded at least two medical claims for type 1 diabetes, a high frequency of type 1 versus type 2 diabetes claims, the use of insulin or continuous glucose monitors, and serial medical and pharmacy use activities over two years of diagnosis or treatment.
This model effectively identified the risk of type 1 diabetes up to 12 months prior to traditional screening methods. This model showed high sensitivity in correctly identifying people with type 1 diabetes. This is 80% in the younger group and 92% in adults. It also maintains improved accuracy compared to traditional screening methods, generally resulting in a positive percentage of just 0.3% in the general population.
“We are energized by the results of this study and what it means for risk detection for early type 1 diabetes, allowing for more efficient and targeted screening of diseases that are often undetected until serious events encourage medical evaluations.” Laura WilsonDirector Health Economics Results Research, Sanofi's Digital Health. “We believe that applying AI-driven predictive models to real-world data will provide opportunities to identify high-risk individuals much faster and plan and prepare for the future.”
Researchers will begin multiphase studies in close collaboration with key hospital sites and experts to validate and refine new clinical decision support tools for type 1 diabetes. This study aims to integrate advanced AI models with hospital electronic health records, enabling previous data-driven interventions for at-risk patients.
AI uses US open claims data to detect more than 18 times the risk of type 1 diabetes
Using the Symphony Health database, a large healthcare claims database covering 75 million patients, the researchers trained machine learning models to identify people at risk for type 1 diabetes before they show symptoms. Over 2.5 million people were compared to approximately 90,000 individuals with type 1 diabetes to define each group using specific inclusion and exclusion criteria. Record patterns were analyzed to determine who could develop type 1 diabetes. This model was tested in large real-world populations and evaluated using a variety of performance measures to determine how accurately it could predict risk.
The results showed that machine learning models could successfully identify people at risk for type 1 diabetes before symptoms appeared, increasing detection efficiency by more than 18 times. Of patients with type 1 diabetes, 29% were previously misclassified as having type 2 diabetes or other forms of type 2 diabetes, highlighting a significant gap in diagnostic accuracy that can delay appropriate treatment and increase the risk of complications.
Researchers have discovered that the best-performing AI models are bi-directional encoder representations from Transformers (BERT), a sophisticated tool originally designed to understand language. Burt correctly identifies 80% of true type 1 diabetes cases, and is more accurate than the other models, with a stronger odds ratio (97.27 vs. 38.01), meaning that its prediction is much more likely to be accurate.
“We have the opportunity to shift the entire timeline of care by identifying individuals with type 1 diabetes,” he said. Jared JocelynSenior Vice President and Global Edge Head, Sanofi. “These findings demonstrate how AI can help improve detection rates with the goal of uncovering hidden patterns in daily healthcare data and promoting more aggressive and scalable care before disease progression.”
Researchers should note that follow-up studies are needed to validate approaches using additional healthcare datasets in the US and internationally and to validate predictions in the clinical setting. Future work will explore improved model performance by supporting previous data-driven intervention strategies by incorporating more longitudinal, genome, and real data into a broader clinical workflow through multimodal AI technology.
Research presentation details:
Dr. Wilson presents the findings as a slow poster session.
- Identification of early stage autoimmune type 1 diabetes using machine learning algorithms
- was presented in Sunday, June 22nd in 12:30pm
Research presentation details:
Jared presents the findings as a slow poster session.
- Predictive modeling of syndrome type 1 diabetes detection using open claims data
- was presented in Sunday, June 22nd in 12:30pm
About the ADA Science Session
ADA's 85th The Science Session, the world's largest scientific conference focused on diabetes research, prevention and care, is Chicago, IllinoisJune 20th-23rd. Thousands of leading physicians, scientists and healthcare professionals around the world are expected to be directly and effectively convened to advance towards cutting-edge research, treatment recommendations and treatment of diabetes. Participants receive exclusive access to thousands of original research presentations and participate in provocative and engaging interactions with leading diabetes experts. Join science session conversations on social media using #Adascisessions.
About the American Diabetes Association
The American Diabetes Association (ADA) is one of the nation's leading voluntary health agencies fighting to end diabetes and help people thrive. This year, the ADA celebrates the '85 driving discovery and research, and will ultimately prevent, manage, treat and treat treatment. There are 136 million Americans living with diabetes or prediabetes. Through advocacy, program development and education, we are fighting for all of them. To learn more or to get involved, visit diabetes.org or call 1-800-Diabete at 800-342-2383. Join Facebook (American Diabetes Association), Spanise Facebook (Asociación Americana de La Diabetes), LinkedIn (American Diabetes Association), and Instagram (@Amdiabetesassn). For more information about how we advocate for everyone affected by diabetes, see us on X (@Amdiabetesassn).
Media Contact: Mimi Carmody, [email protected]
Source American Diabetes Association
