EvoApneaFormer: an IoT and prognostic evolutionary deep learning-based framework for real-time multi-event sleep apnea disorder detection and remote monitoring

Introduction Sleep apnea is a prevalent, yet underdiagnosed, disorder associated with cardiovascular disease, hypertension, and cognitive decline. Although polysomnography (PSG) is the diagnostic gold standard, its high cost > $6,000 per unit, invasiveness, and limited availability hinder large-scale screening. Methods To overcome these barriers, we propose ApneaSense, a non-intrusive IoT-enabled diagnostic framework powered by EvoApneaFormer, an evolutionary transformer-based deep learning model for temporal bio-signal fusion and adaptive learning. EvoApneaFormer combines dynamic self-attention with neuroevolutionary optimization, enabling superior convergence, noise resilience, and generalization across heterogeneous signals. ApneaSense processes synchronized ECG, SpO2, respiratory effort, and motion data on low-power Raspberry Pi 4 hardware <2W using TensorFlow Lite quantization, with a companion Flutter app for real-time visualization, 72-h offline operation, and secure HL7/FHIR telehealth synchronization. Clinical validation on a hybrid dataset, the UCD Sleep Apnea Database (n = 25) and the ApneaSense Clinical Dataset (n = 61), with stratified patient-level splitting. Results On the held-out test cohort (n = 17), EvoApneaFormer achieved 99.98% accuracy, 99.91% precision, 99.95% recall, a 99.93% F1-score, and a 0.999 macro-AUC. Six apnea classes, namely, normal, obstructive, central, mixed, hypopnea, and respiratory effort-related arousal (RERA), were each identified with ≥99.8% accuracy. Real-world trials confirmed robustness in both clinical 99.8% and home 96.9% settings, even under motion artifacts and signal dropouts. SHAP-based interpretability and uncertainty quantification were deemed actionable by 82% of clinicians. Discussion Cost-effectiveness analysis indicated break-even in 6.2 months at $5,200/QALY. Compared to wearable oximeters, ApneaSense suggested a 9.98-point accuracy gain at one-third the cost per test, representing a paradigm shift in edge-based respiratory diagnostics for scalable and personalized sleep apnea monitoring in underserved communities.