کنفرانس مهندسی برق ایران

صفحه اصلی / سی و سومین کنفرانس بین المللی مهندسی برق

Implementation of an Optimized Deep Learning Model to Assess Pediatric Sleep Apnea Severity Using SpO2 Signals on Resource-Limited Microcontrollers

نویسندگان :

Erfan Mortazavi¹ Hanieh Mohammadi² Bahram Tarvirdizadeh³ Khalil Alipour⁴ Mohammad Ghamari⁵

1- دانشگاه تهران 2- دانشگاه تهران 3- دانشگاه تهران 4- دانشگاه تهران 5- California Polytechnic State University

کلمات کلیدی :

Sleep apnea-hypopnea (SAH)،Blood oxygen saturation (SpO2)،Convolutional neural network (CNN)،Long short-term memory (LSTM)،Attention،Apnea–hypopnea index (AHI)،Microcontroller

چکیده :

Accurately diagnosing pediatric sleep apnea-hypopnea (SAH) is a complex task in pediatric healthcare. Traditional methods like polysomnography (PSG) test, though effective, can be uncomfortable and impractical for children. This study explores a less invasive approach using deep learning to analyze blood oxygen saturation (SpO2) signals. A total of 844 SpO2 signals from the CHAT dataset were utilized, with the data split into 60% for training, 30% for testing, and 10% for validation to train a convolutional neural network (CNN)-long short-term memory (LSTM)-Attention model. The model achieved a kappa score of 0.62 and a four-class accuracy of 74.91% in estimating the apnea-hypopnea index (AHI) and classifying sleep apnea severity. The primary challenge was deploying the model on the resource-constrained STM32H743IIT6 microcontroller. TensorFlow toolkit optimization techniques were implemented to minimize model size and resource usage while preserving satisfactory accuracy, despite the microcontroller's sufficient 1MB data RAM. Each technique was evaluated in memory-constrained environments, leading to post-quantization deployment. The best optimized model maintained an accuracy of 73.63% and a kappa score of 0.60, demonstrating the feasibility of portable, non-invasive diagnostic tools.