Evaluating generalizability of population-based and age-segmented models for hypoglycemia classification
Publication date
2026-06-01
Document type
Conference paper
Author
Organisational unit
Scopus ID
Conference
Conference on Artificial Intelligence 2026 (IEEE CAI 2026) ; Granada, Spain ; May 8–10, 2026
Publisher
IEEE
Book title
2026 IEEE Conference on Artificial Intelligence (CAI)
Part of the university bibliography
✅
Language
English
Keyword
Age
Diabetes
DiaData
Generalization
Hypoglycemia Classification
Individualization
Abstract
Artificial intelligence (AI) augments medical care by providing personalized predictions not restricted to fixed rules and parameters. This data-driven improvement is gaining importance for chronic diseases like type 1 diabetes (T1D), where disease progression varies with demographic and genetic factors. Notably, patients with T1D depend on insulin therapy, which can lead to hypoglycemia, a dangerous condition of low blood glucose levels (≤70 mg/dL) associated with increased mortality risk. Hypoglycemia risk can be reduced through AI-based early alarms leveraging data from continuous glucose monitoring (CGM) devices. However, while glucose variability and hypoglycemia occurrence differ across demographics, hypoglycemia classification most often relies on population-based models restricted to specific age ranges. Thus, we classify hypoglycemia 0, 5-15, 20-45, and 50-120 minutes before onset using DiaData, a large CGM dataset of patients with T1D, including children, teenagers, adults, and seniors. In particular, we investigate: 1) the performance of a Fully Convolutional Network (FCN) trained on the whole population and tested separately on subjects across four age groups, 2) the impact of age-segmented models, where the same model architecture is trained separately for the age groups, and 3) the effect of individualization through transfer learning, where the model is fine-tuned with the test subjects' data. The results show that the FCN model trained on the whole population captures temporal patterns leading to hypoglycemia with similar or superior performance compared to models trained on age-specific subsets, indicating strong cross-demographic generalization. Likewise, age-segmented models generalize well across age groups. However, models specialized on children's data achieve the highest performance on test sets of children but the lowest on other age groups. Thus, a use case restricted to children should train only on children's data.
Version
Published version
Access right on openHSU
Metadata only access