An Accurate, Low-Cost Machine Learning System for Sleep Apnea Detection

Overview

Sleep apnea, a disorder in which breathing repeatedly stops and starts during sleep due to airway obstruction or disrupted brain signaling, affects over one billion people worldwide.However, over 80-90% of these cases remain undiagnosed due to the high cost and inaccessibility of polysomnography, which requires 25+ sensors and costs over $3,000. This project developed and evaluated a low-cost machine learning system to detect sleep apnea events using only ECG and SpO₂ signals, targeting a final hardware cost of ~$20. Using the HuGCDN2014-OXI dataset of 83 participants, ECG and SpO₂ signals were segmented into one-minute windows and processed to extract six factors detailing cardiovascular variation and oxygen desaturation patterns. Five machine learning models - XGBoost, SVM-RBF, Hidden Markov Model, Gaussian Process Classifier, and Naive Bayes Network - were implemented and compared against a traditional SpO₂ threshold baseline using stratified 80/20 train-test splitting. XGBoost achieved the highest participant-level AUC of 0.958, outperforming all models and the baseline. To assess generalizability, the trained model was applied without retraining to the independent PhysioNet Apnea-ECG dataset, achieving 0.86 AUC using ECG features alone, confirming cross-dataset validity. All machine learning models significantly outperformed the threshold baseline, demonstrating that multi-feature classification captures apnea patterns that threshold methods miss. A low-cost hardware prototype integrating ECG and SpO₂ sensors was assembled using an Arduino microcontroller to demonstrate real-world deployment feasibility. These results suggest machine learning is a viable segue to affordable, accurate at-home sleep apnea screening, expanding the opportunity for diagnosis to the millions currently undetected.