Deployment of a Spectrogram-Based Classifier for Physical Activity in Type 1 Diabetes on Embedded Systems

• Autores: Saavedra, Marcos David; Estrebou, César Armando; Inthamoussou, Fernando Ariel; Garelli, Fabricio
• Año de publicación: 2025
• Idioma: inglés
• Tipo de recurso: documento de conferencia
• Materia: Ciencias Informáticas; physical activity classification; embedded systems; spectrogram; convolutional neural network; type 1 diabetes
• Institución: Universidad Nacional de La Plata

Individuals with Type 1 Diabetes (T1D) face significant challenges in maintaining glycemic control, particularly during and after physical activity (PA), due to its variable impact on blood glucose levels. Accurately identifying the type of PA in real-time is essential for developing effective decision support systems (DSS) and automated insulin delivery (AID) systems that can dynamically adjust therapy. Recent work has demonstrated the effectiveness of Convolutional Neural Networks (CNNs) for classifying PA types (aerobic, interval, resistance) in T1D using heart rate and step count data. However, for practical application in wearable devices or medical equipment like insulin pumps, the computational demands of such models must be significantly reduced. This work presents the deployment of an efficient spectrogram-based CNN PA classifier into resource-constrained embedded systems, specifically microcontrollers (MCUs). The optimization of the spectrogram generation process and the CNN model to suit the limitations of MCU hardware using the EmbedIA framework for MCU code generation is detailed. The classifier was successfully deployed on four MCU platforms (ESP32, ESP32-C3, ESP8266, Raspberry Pi Pico) using floating-point (float), 32- bit fixed-point (fixed32) and 16-bit fixed-point (fixed16) arithmetic. The float and fixed32 implementations maintained the original model’s high accuracy (87.64%), while the fixed16 version showed a minor but acceptable drop to 85.16%. Resource analysis revealed significant reductions in memory usage and, notably, substantial decreases in inference time for fixed-point implementations on MCUs without dedicated floating-point units. These results suggest that the proposed implementation is suitable for low-power embedded systems, highlighting its potential to be integrated into wearable devices and AID/DSS systems for more responsive and personalized glucose management in T1D.
Red de Universidades con Carreras en Informática