Vera F. Lehmann, Switzerland
University Hospital Bern Department of Diabetes, Endocrinology, Nutritional Medicine and MetabolismPresenter of 1 Presentation
HEADWIND: DESIGN AND EVALUATION OF A VEHICLE HYPOGLYCEMIA WARNING SYSTEM IN DIABETES – RESULTS FROM A DRIVING SIMULATOR STUDY
- Vera F. Lehmann, Switzerland
- Thomas Zueger, Switzerland
- Mathias Kraus, Switzerland
- Martin Maritsch, Switzerland
- Stefan Feuerriegel, Switzerland
- Felix Wortmann, Switzerland
- Tobias Kowatsch, Switzerland
- Caroline Albrecht, Switzerland
- Caterina Bérubé, Switzerland
- Naïma Styger, Switzerland
- Sophie Lagger, Switzerland
- Markus Laimer, Switzerland
- Elgar Fleisch, Switzerland
- Christoph Stettler, Switzerland
Abstract
Background and Aims
Hypoglycemia is one of the most relevant acute complications of diabetes mellitus and is associated with an increased risk of driving accidents. Today's cars gather a broad spectrum of real-time driving parameters. Based on changes in driving behaviour during hypoglycemia we aim at establishing algorithms capable of discriminating eu- and hypoglycemic driving patterns using artificial intelligence.
Methods
We included active drivers with type 1 diabetes mellitus (T1DM). Using an adapted hypoglycemic clamp protocol and a professional driving simulator, driving data was recorded in 2 glycemic states: euglycemia (deu, 5-8 mmol/l) and hypoglycemia (dhypo, 2.0-2.5 mmol/l). In each glycaemic state the participants drove for 15 min through a random sequence of 3 environments: highway, rural and town. Car-based sensor data were sliced into overlapping 45-second windows. Finally, predictive performance in hypoglycemia detection was evaluated using gradient boosted decision trees.
Results
The study encompassed 15 participants with T1DM (11 male, HbA1c 7.2±0.6 %). Mean blood glucose in deu and dhypo was 5.9±0.6mmol/l and 2.4±0.25 mmol/l (p<0.001), respectively. Car-based data provided 466,303 measurements in deu and 481,497 samples in dhypo. 1-fold cross-validation on subject level resulted in a ROC-AUC in hypoglycemia prediction of 0.85. ROC-AUC for the highway, rural and town environment was 0.80, 0.98, and 0.78, repectively.
Conclusions
Our study applying machine learning models on driving simulator-based data shows robust between-subject predictability of hypogylcemia. This confirms the effectiveness of artificial intelligence in hypoglycemia detection while driving and may represent a promising novel approach to increase traffic safety in people with diabetes.