Deep-learning algorithms using smart bed pressure signals accurately estimated heart rate (mean bias 0.05 bpm) and respiration rate (mean bias -0.16 breaths/min) compared to PSG and ECG.
Observational
Can deep-learning algorithms accurately estimate heart rate, respiration rate, and heart rate variability from smart bed pressure signals compared to standard PSG and ECG?
Deep learning algorithms applied to smart bed pressure signals can accurately and non-invasively estimate heart rate, respiration rate, and heart rate variability during sleep.
Effect estimate: mean bias 0.05 bpm (95% CI 0.01 to 0.08)
Abstract Introduction Deep-learning algorithms estimated heart rate (HR), respiration rate (RR), and HRV SDNN at 1 Hz from smart beds using ballistocardiography (BCG) signals captured by a pressure sensor embedded in the bed’s air chamber for non-invasive biometric monitoring. Methods BCG signals were sampled at 1kHz. The signals were processed with an 8 Hz low-pass filter and subsequently high-pass filtered at 0.1 Hz for RR, and at 0.5 Hz for HR and HRV. After resampling to 20 Hz, 60-second segments were normalized using z-score transformation. Supervised models were built using a convolutional neural network (CNN) backbone with three stacked blocks of either Attention Free Transformer (AFT), Transformer encoder, or LSTM applied for RR, HR and SDNN estimation respectively. The training set consisted of 127 sleep sessions, using polysomnography (PSG)-derived abdominal belts as the reference standard for RR and electrocardiography (ECG) for HR and HRV. Validation was performed on an independent dataset of 81 unseen sessions for biometrics and 45 sessions for SDNN. Model performance was evaluated using Bland–Altman analysis and coefficient of determination (R2) at 30-second epoch and session levels. Results HR estimation demonstrated close agreement with reference values, yielding epoch-level limits-of-agreement (LoA) from −9.42 to 9.51 bpm and a mean bias of 0.05 bpm 95% CI: 0.01 to 0.08. At the session level, LoA ranged from −2.96 to 3.47 bpm, with a bias of 0.26 bpm 95% CI: −0.11 to 0.62. R2 values were 0.78 and 0.97 respectively. RR estimation showed epoch-level LoA from −2.70 to 2.38 breaths/min and a bias of −0.16 breaths/min 95% CI: −0.17 to −0.15, while session-level LoA ranged from −0.77 to 0.47 breaths/min with a bias of −0.15 breaths/min 95% CI: −0.22 to −0.08. Corresponding R2 were 0.65 and 0.96. For SDNN, epoch-level LoA was −22.52 to 26.28 ms with a bias of 1.88 ms 95% CI 1.75 to 2.01, and session-level LoA was −9.89 to 13.68 ms with a bias of 1.89 ms 95% CI: 0.09 to 3.70. Corresponding R2 were 0.66 and 0.82. Conclusion This approach enables unobtrusive, continuous at-home monitoring of sleep biometrics, supporting scalable, personalized sleep assessment without the need for invasive instrumentation. Support (if any) n /a
Rao et al. (Fri,) conducted a observational in Sleep biometrics monitoring. Deep-learning algorithms using smart bed pressure signals (BCG) vs. Polysomnography (PSG) and electrocardiography (ECG) was evaluated on Heart rate (HR) estimation accuracy (epoch-level mean bias) (mean bias 0.05 bpm, 95% CI 0.01 to 0.08). Deep-learning algorithms using smart bed pressure signals accurately estimated heart rate (mean bias 0.05 bpm) and respiration rate (mean bias -0.16 breaths/min) compared to PSG and ECG.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: