A Low-Cost, Interpretable Machine-Learning System for Sleep Apnea Detection Using ECG and SpO₂ Signals

Sleep apnea affects close to one billion people worldwide, yet an estimated 80 to 90 percent of cases remain undiagnosed, largely because the diagnostic gold standard, inlaboratory polysomnography, is expensive and inconvenient. Existing at-home screeners rely on simple oxygen-desaturation thresholds and achieve limited accuracy, ignoring the nonlinear physiology of apnea and the cardiovascular changes visible in the electrocardiogram (ECG). This study presents a low-cost, interpretable machinelearning system that detects apnea from ECG and pulse oximetry (SpO₂). Using the expert-annotated HuGCDN2014-OXI database of 83 overnight recordings, I engineered six physiologically grounded features capturing heart-rate variability and oxygen desaturation across non-overlapping one-minute windows, and compared five machinelearning models against a threshold-based baseline that emulates current home kits. Models were trained and tested with participant-level splits to prevent subject leakage, and generalization was assessed on the independent PhysioNet Apnea-ECG dataset. XGBoost performed best, with a window-level AUC of 0.916 and a participant-level AUC of 0.958, far exceeding the baseline (0.599 and 0.694). Because the external dataset contains only ECG, external validation using the ECG features alone still reached an AUC of 0.86. A hardware prototype integrating both sensors was built for approximately twenty dollars. These results show that inexpensive, interpretable machine learning can substantially outperform threshold-based home screening and improve access to sleepapnea detection.