The PSG-2-BPNet deep-learning model predicted nocturnal mean arterial blood pressure from polysomnography signals with a mean absolute error of 9.41 mmHg (95% CI 9.15-9.67).
Can a deep-learning model (PSG-2-BPNet) accurately infer nocturnal mean arterial blood pressure from polysomnography signals in adults with OSA?
A novel deep-learning model can estimate nocturnal mean arterial blood pressure from standard polysomnography signals with an MAE of 9.41 mmHg, offering a potential cuff-free BP monitoring approach during sleep studies.
Estimación del efecto: MAE 9.41 mmHg (95% CI 9.15-9.67)
Abstract Rationale Obstructive sleep apnea (OSA) is highly prevalent and causes night-time spikes and loss of normal “dipping” in blood pressure (BP), key patterns linked to hypertension and cardiovascular risk. We develop a deep-learning model, PSG-2-BPNet, that utilizes clinical polysomnography (PSG) signals to predict nocturnal mean arterial blood pressure (MAP) profile in adults with OSA. Methods We studied 56 adults with untreated moderate-severe OSA undergoing PSG with concurrent beat-to-beat BP monitoring. Following artifact removal and signal preprocessing, the PSG data (inputs to PSG-2-BPNet) were structured into a supervised rolling-window dataset, with a 30-second window (matching the conventional PSG epoch) and a 1-second stride. The proposed PSG-2-BPNet model predicted the average MAP over the subsequent 10 seconds, enabling continuous, real-time nocturnal MAP estimation directly from PSG signals. A 51/5 sample split was used for training and testing. The MAP input values were primarily concentrated within the 80-120 mmHg range (68.47%), with fewer samples in the 60-80 mmHg range (8.00%) and a smaller proportion in the 120-200 mmHg range (23.53%), highlighting the limited representation of values in the very high or low physiological range. The PSG-2-BPNet architecture consisted of 20 per-signal sequence autoencoders that extract signal-specific latent features, which are then fused through a regression head for continuous MAP prediction. Results The sample consisted of middle-aged individuals (mean ± standard deviation, 45.1 ± 9.6 years), with 67% being men (38/57), and had a body mass index (BMI) of 37.1 ± 7.1 kg/m². On the test samples (MAP 60-180 mmHg), the model achieved a mean absolute error (MAE) of 9.41 mmHg (bootstrap mean 9.41; 95% CI 9.15-9.67), a bias of 0.07 ± 12.63 mmHg, and coverage of 36.4% within ±5 mmHg and 62.6% within ±10 mmHg. Performance was strongest in the 80-120 mmHg range (8.3 mmHg MAE), followed by the 120-180 mmHg range (10 mmHg MAE). Accuracy declined for lower pressures (80 mmHg), which represented 8% of the training data, reflecting a limited amount of training data in that range. Conclusion PSG-2-BPNet provides a cuff-free approach to reconstructing the nocturnal MAP profile from standard PSG signals, showing acceptable performance across individuals with OSA. PSG-2-BPNet should be further trained across a larger sample with a broader range of MAP values to enhance its clinical reliability and for external validation. This abstract is funded by: NHLBI 1R56HL157182
Prasad et al. (Fri,) conducted a other in Obstructive sleep apnea (OSA) (n=56). PSG-2-BPNet (deep-learning model) vs. Concurrent beat-to-beat BP monitoring was evaluated on Mean absolute error (MAE) of predicted nocturnal mean arterial blood pressure (MAP) (MAE 9.41 mmHg, 95% CI 9.15-9.67). The PSG-2-BPNet deep-learning model predicted nocturnal mean arterial blood pressure from polysomnography signals with a mean absolute error of 9.41 mmHg (95% CI 9.15-9.67).