Abstract Background Obstructive sleep apnea (OSA) is common and underdiagnosed. While polysomnography (PSG) is the gold standard, scalable, non-contact solutions are needed for screening and longitudinal follow-up. We validated a smart mattress system equipped with biometric sensors across diverse phenotypes (men, women, and pediatric patients). Aims To assess the diagnostic performance of a smart, non-contact mattress for apnea/hypopnea detection compared with gold-standard PSG, and to describe the signal processing and model selection approach used. Methods Prospective study conducted at a tertiary Sleep Unit, with a growing repository of recordings from more than 300 patients. Mattress-derived ballistocardiography (BCG) and embedded respiratory activity (ERA) signals were time-aligned with PSG and segmented into 30-second windows labeled as apnea/non-apnea. Statistical, temporal, and morphological features were extracted to create a supervised dataset. Several algorithmic configurations were tested across stratified train/validation/test splits. Performance was quantified using AUROC, AUPRC, and F1-score, with leave-one-patient-out (LOPO) cross-validation employed to avoid information leakage. Results The best-performing classifier achieved AUROC = 0.858 and AUPRC = 0.694. At the selected threshold, the model yielded F1 = 0.519, precision = 0.716, and recall = 0.407 (TP = 1,721; FP = 682; FN = 2,508; TN = 11,125). This reflects a conservative detection profile: high precision (71.6% of flagged events are true) with lower recall. Clinically, this supports confident detection of true apnea events suitable for screening and longitudinal monitoring. Continued signal integration and model refinement are expected to improve sensitivity while maintaining precision (Figure 1). Conclusions A smart mattress system can detect clinically relevant respiratory events with high precision compared with PSG across heterogeneous patient profiles. Dataset expansion and model retraining will further enhance alignment with gold-standard scoring, supporting scalable OSA screening and home-based follow-up. This abstract is funded by: This project received funding from the Economic Forward Agency of La Rioja (ADER) and the European Regional Procurement Fund (FEDER). It also benefitted from the collaboration of Welltech.
Lázaro et al. (Fri,) studied this question.