Abstract
Background and Aims
BGL predictive algorithms can improve T1DM treatment preventing glucose excursions. An autonomous system based on neural networks was developed to personalize blood glucose predictions based on blood glucose, insulin infusion, nutrient intakes and heart rates in a real world scenario.
Methods
20 T1DM patients were monitored with flash glucose monitor, activity tracker and mobile app (GlucoTrends) to collect meal and insulin data. Personalized prediction models were trained independently, without requiring specific settings for each individual. Patient's characteristics: 9 female, age: 32.4(SD:10.5), BMI: 26.0(3.8), BGL: 159.1(34.0), under the following therapies: 45% under fixed doses, 40% carbohydrate-count and 15% insulin pumps. Patients were monitored during on average 29.3 days (SD: 7.9), and the last 20% measurements were reserved for evaluation. Prediction accuracy for 1-hour prediction horizon was evaluated by Clarke Error Grid (CEG) compareding to BGL measured with flash monitor.
Results
The percentage of predictions within AB zones of CEG for all-day and only for night-time periods are 92.9% and 94.0%, respectively (Figure 1). Patients using Insulin Pumps, Analog insulin and Human insulin had 97.5%(0.7), 93.6%(4.5) and 90.0%(8.2) predictions within AB zone on average, respectively.
Conclusions
Our proposed algorithm has demonstrated to be capable to personalize glucose prediction in a real world scenario and poses as a potential solution for an autonomous support system. In our study different insulin regimen and molecules confirmed expected results. Higher than desired predictions in zone D may be caused by sensor's error biases within minimum glucose levels.
