Personalized food consumption detection with deep learning and Inertial Measurement Unit sensor

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Diabetes Technology

Background:

• Accurate carbohydrate counting is essential for diabetes mellitus management, particularly for individuals using artificial pancreas systems.
• Manual carbohydrate tracking is often overlooked, impacting glycemic control and insulin dosing.
• Artificial pancreas systems rely on precise carbohydrate intake data for effective insulin pump activation.

Purpose of the Study:

• To develop and validate a personalized deep learning model for accurate carbohydrate intake detection.
• To improve automated insulin delivery in artificial pancreas systems through precise meal monitoring.
• To address the challenge of inconsistent manual carbohydrate tracking in diabetes management.

Main Methods:

• Utilized a publicly available dataset from an Inertial Measurement Unit (IMU) with accelerometer and gyroscope data sampled at 15 Hz.
• Preprocessed the sensor data and employed a recurrent neural network architecture with Long short-term memory (LSTM) layers for patient-tailored modeling.
• Evaluated model performance using metrics such as F1 score and confusion matrix analysis.

Main Results:

• The deep learning model achieved a median F1 score of 0.99, demonstrating high accuracy in carbohydrate intake detection.
• The model exhibited performance consistently above 90%, with most results ranging from 98% to 99%.
• Confusion matrix analysis showed minimal discrepancies (6 seconds), and average prediction latency was 5.5 seconds.

Conclusions:

• The developed personalized deep learning model accurately detects carbohydrate intake, offering a significant advancement for diabetes management.
• Recurrent neural networks, particularly LSTMs, substantially enhanced problem-solving capabilities for this task.
• Future work may involve transformer networks and shorter time windows to further improve responsiveness and accuracy, with multi-day data collection recommended.