Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Learning-Based Hypoglycemia Classification Across Multiple Prediction Horizons

Beyza Cinar, Jennifer Daniel Onwuchekwa, Maria Maleshkova

TL;DR

This work targets hypoglycemia risk in type 1 diabetes by classifying time-to-onset across multiple horizons (0–24 h) within a single deep learning framework using CGM, insulin, and activity data from the OhioT1DM dataset. It compares LSTM, ResNet, and a hybrid model under population-based and subject-specific training, finding the LSTM consistently superior for nine-class tasks, especially for short horizons, while long-horizon predictions are challenging due to inter-subject variability and limited data. A six-class configuration improves macro-recall and early-event detection but still falls short of state-of-the-art performance, indicating a need for larger datasets or separate short- and long-horizon models. The study highlights practical challenges in deploying multi-horizon DL hypoglycemia classifiers and suggests directions like separate models for short- and long-term predictions and improved data imputation to enhance performance.

Abstract

Type 1 diabetes (T1D) management can be significantly enhanced through the use of predictive machine learning (ML) algorithms, which can mitigate the risk of adverse events like hypoglycemia. Hypoglycemia, characterized by blood glucose levels below 70 mg/dL, is a life-threatening condition typically caused by excessive insulin administration, missed meals, or physical activity. Its asymptomatic nature impedes timely intervention, making ML models crucial for early detection. This study integrates short- (up to 2h) and long-term (up to 24h) prediction horizons (PHs) within a single classification model to enhance decision support. The predicted times are 5-15 min, 15-30 min, 30 min-1h, 1-2h, 2-4h, 4-8h, 8-12h, and 12-24h before hypoglycemia. In addition, a simplified model classifying up to 4h before hypoglycemia is compared. We trained ResNet and LSTM models on glucose levels, insulin doses, and acceleration data. The results demonstrate the superiority of the LSTM models when classifying nine classes. In particular, subject-specific models yielded better performance but achieved high recall only for classes 0, 1, and 2 with 98%, 72%, and 50%, respectively. A population-based six-class model improved the results with at least 60% of events detected. In contrast, longer PHs remain challenging with the current approach and may be considered with different models.

Deep Learning-Based Hypoglycemia Classification Across Multiple Prediction Horizons

TL;DR

This work targets hypoglycemia risk in type 1 diabetes by classifying time-to-onset across multiple horizons (0–24 h) within a single deep learning framework using CGM, insulin, and activity data from the OhioT1DM dataset. It compares LSTM, ResNet, and a hybrid model under population-based and subject-specific training, finding the LSTM consistently superior for nine-class tasks, especially for short horizons, while long-horizon predictions are challenging due to inter-subject variability and limited data. A six-class configuration improves macro-recall and early-event detection but still falls short of state-of-the-art performance, indicating a need for larger datasets or separate short- and long-horizon models. The study highlights practical challenges in deploying multi-horizon DL hypoglycemia classifiers and suggests directions like separate models for short- and long-term predictions and improved data imputation to enhance performance.

Abstract

Type 1 diabetes (T1D) management can be significantly enhanced through the use of predictive machine learning (ML) algorithms, which can mitigate the risk of adverse events like hypoglycemia. Hypoglycemia, characterized by blood glucose levels below 70 mg/dL, is a life-threatening condition typically caused by excessive insulin administration, missed meals, or physical activity. Its asymptomatic nature impedes timely intervention, making ML models crucial for early detection. This study integrates short- (up to 2h) and long-term (up to 24h) prediction horizons (PHs) within a single classification model to enhance decision support. The predicted times are 5-15 min, 15-30 min, 30 min-1h, 1-2h, 2-4h, 4-8h, 8-12h, and 12-24h before hypoglycemia. In addition, a simplified model classifying up to 4h before hypoglycemia is compared. We trained ResNet and LSTM models on glucose levels, insulin doses, and acceleration data. The results demonstrate the superiority of the LSTM models when classifying nine classes. In particular, subject-specific models yielded better performance but achieved high recall only for classes 0, 1, and 2 with 98%, 72%, and 50%, respectively. A population-based six-class model improved the results with at least 60% of events detected. In contrast, longer PHs remain challenging with the current approach and may be considered with different models.

Paper Structure

This paper contains 15 sections, 4 equations, 6 figures, 15 tables.

Table of Contents

  1. Introduction
  2. State of the Art
  3. Methodology
  4. Dataset: OhioT1DM
  5. Preprocessing
  6. Time Series Generation
  7. Model Architectures
  8. Metrics Used for the Model Evaluation
  9. Results
  10. Population-Based Models Using 9 Classes
  11. Subject-Specific and Population-Based Models Using 9 Classes
  12. Population-Based Models Using 6 Classes
  13. Subject-Specific and Population-Based Models Using 6 Classes
  14. Discussion
  15. Conclusion and Future Work

Figures (6)

  • Figure 1: Architecture of the Applied RNN Models a) LSTM, b) BiLSTM
  • Figure 2: Architecture of the Applied CNN Models a) ResNet, b) 1DCNN
  • Figure 3: Architecture of the Applied Hybrid Model: ResNet + LSTM
  • Figure 4: Population-Based Confusion-Matrix Across 9 and 6 Classes of Subject 567 Trained with the LSTM Model
  • Figure 5: Population-Based and Subject-Specific Confusion-Matrices Across 9 Classes
  • ...and 1 more figures