Benchmarking ResNet for Short-Term Hypoglycemia Classification with DiaData

Researchers Beyza Cinar and Maria Maleshkova have published a new benchmark for predicting dangerous hypoglycemic events in Type 1 Diabetes patients. Their work, presented at the BHI 2025 conference, centers on the DiaData repository—an integration of 15 separate datasets containing glucose values from 2510 subjects. A key contribution is a rigorous data cleaning pipeline that identifies outliers using an interquartile range (IQR) method and imputes missing values. For small data gaps (≤25 minutes), they use linear interpolation, while for larger gaps (30-120 minutes), they employ Stineman interpolation, which they found provides more realistic glucose estimates.

Using this refined dataset, the team established a correlation between glucose levels and heart rate 15 to 60 minutes before a hypoglycemic event. They then trained a state-of-the-art ResNet model, a deep learning architecture common in computer vision, to classify the onset of hypoglycemia (≤70 mg/dL) up to two hours in advance. The results are significant: training with more data improved model performance by 7%, while using their quality-refined data instead of raw data yielded an additional 2-3% gain in predictive accuracy. This benchmark provides a crucial foundation for developing reliable, individualized warning systems.