AI Models Detect Type 1 Diabetes Risk Early

AI Identifies Early Type 1 Diabetes Risk

Developments from two studies highlighting the potential for machine learning using AI to improve early identification of type 1 diabetes were presented at the 85th Scientific Sessions of the American Diabetes Association (ADA) in Chicago.

Around 64,000 Americans are diagnosed with type 1 diabetes each year, and as many as 40% are unaware they have it until a life-threatening event occurs. The disease can progress silently until symptoms like excessive thirst, frequent urination, or diabetic ketoacidosis appear. By then, significant damage to insulin-producing cells may have already occurred, emphasizing the need for earlier detection.

AI Model for Risk Assessment

A new study demonstrated AI's potential to identify individuals at risk for type 1 diabetes up to a year before diagnosis, with greater accuracy and fewer false positives than standard methods. The retrospective study developed two age-specific machine learning models—one for those aged 0–24 and another for those 25 and older—using medical claims and lab test data from NorstellaLinQ.
Researchers used criteria to identify stage 3 type 1 diabetes cases, including at least two medical claims for type 1 diabetes, more type 1 versus type 2 diabetes claims, documented insulin or continuous glucose monitor use, and continuous medical and pharmacy claims activity in the two years before diagnosis or treatment.

The models effectively identified type 1 diabetes risk up to 12 months earlier than traditional screening. The models showed high sensitivity in correctly identifying people with type 1 diabetes—approximately 80% in the younger group and 92% in adults. They also maintained improved precision compared to conventional screening methods.

Researchers plan to launch a study to validate and refine a clinical decision support tool for type 1 diabetes, integrating AI models with hospital electronic health records to enable earlier interventions for at-risk patients.

AI Using Claims Data

Researchers used the Symphony Health Database, covering 75 million patients, to train a machine learning model to identify people at risk for type 1 diabetes before symptoms appear. Records from nearly 90,000 individuals with type 1 diabetes and over 2.5 million people without it were compared. Patterns were analyzed to determine who was likely to develop type 1 diabetes.

The model was tested on a real-world population to determine its prediction accuracy. Results showed machine learning models could identify people at risk for type 1 diabetes before symptoms appeared, increasing detection efficiency more than 18-fold. Among those with type 1 diabetes, 29% had been misclassified as having type 2 diabetes or other forms. The AI model that performed best, Bidirectional Encoder Representations from Transformers (BERT), correctly identified 80% of true type 1 diabetes cases and was more accurate than other models.

Researchers note follow-up studies are needed to validate the approach using additional healthcare datasets and in clinical settings. Future work will also explore enhancing model performance through multimodal AI techniques and by incorporating more longitudinal, genomic, and real-world data into clinical workflows.