Development of a long-term AI-enabled monitoring device using reliable non-contact sensors for sleep apnea

In this prospective observational study, we developed and validated a contactless diagnostic system combining 60-GHz frequency-modulated continuous-wave (FMCW) radar with artificial intelligence (AI) for obstructive sleep apnea (OSA) detection and evaluated it against polysomnography (PSG). Eighty-two adults (aged 18–65 years) underwent overnight PSG for suspected OSA with simultaneous 60-GHz radar monitoring of thoracoabdominal movement and autonomic surrogates. Radar signals were processed into respiratory and heart rate variability features and analyzed using convolutional neural networks and a random forest. The resulting probabilities were combined through an unweighted ensemble; before analysis, a consensus quality-control step excluded segments that all three streams failed to classify confidently. We assessed agreement with PSG using the intraclass correlation coefficient (ICC) and Bland–Altman analysis. We evaluated patient-level performance at apnea–hypopnea index (AHI) thresholds of 5, 15, and 30 events/h and across four severity tiers. Radar-derived AHI correlated with PSG-derived AHI (r = 0.80; p < 0.001; ICC = 0.80), with a mean bias of − 2.7 events/h and 95 % limits of agreement from − 38.1 to + 32.7 events/h. For severe OSA (AHI ≥ 30), sensitivity was 82.9 % and specificity 97.6 %; severity cross-classification achieved 64.6 % exact-tier and 93.9 % within-one-tier agreement (linear weighted κ = 0.60). The ensemble reached a segment-level accuracy of 82.7 %, a macro-F1 of 0.796, and an AUROC of 0.902 across a continuous 0.5–1.5 m operating range. Therefore, an AI-augmented, contactless FMCW radar can provide reliable segment-level detection and useful patient-level severity estimates, and they support its further development as a home-based tool for OSA screening and longitudinal follow-up in populations that current testing underserves.