Nocturnal Blood Pressure Estimation from Sleep Plethysmography Using Machine Learning

Gizem Yilmaz, Xingyu Lyu, Ju Lynn Ong, Lieng Hsi Ling, Thomas Penzel, B. T. Thomas Yeo, Michael W. L. Chee

TL;DR

This paper presents a machine learning approach to estimate nocturnal blood pressure (BP) using only fingertip photoplethysmography (PPG) signals during sleep. The developed random forest model achieved promising accuracy, suggesting a potential for non-invasive, continuous BP monitoring during sleep, which is crucial for cardiovascular health.

What Problem Does This Paper Address?

Hypertension, particularly elevated nocturnal blood pressure, is a significant risk factor for cardiovascular disease and mortality. Traditional cuff-based blood pressure monitoring during sleep is often uncomfortable, can disrupt sleep, and is not conducive to continuous monitoring. While cuffless BP estimation methods using PPG and machine learning have emerged, most focus on daytime measurements and often require calibration. There's a need for accurate, non-invasive, and calibration-free methods to monitor BP specifically during sleep, a critical period for cardiovascular risk assessment.

Methods

The researchers developed a machine learning model to predict nocturnal systolic blood pressure (SBP) and diastolic blood pressure (DBP) using features extracted from single-channel fingertip PPG signals.

  • Data Acquisition: Overnight polysomnography (PSG) data, including fingertip PPG and ambulatory blood pressure monitoring (ABPM), were collected from 68 healthy adults.
  • Feature Extraction: PPG waveforms were pre-processed (filtered, motion artifact reduction) and then divided into 30-second windows with 20-second overlap. Fiducial points (onset, peak, notch, etc.) of the PPG pulse were identified using a combination of derivative analysis and Gaussian fitting. From these points, 75 PPG waveform features were extracted, focusing on amplitude and duration characteristics.
  • Machine Learning Model: A random forest regression model was employed. This model is known for its ability to capture non-linear relationships between features and targets.
  • Training and Validation: A 5-fold cross-validation strategy was used, with data split into training, validation, and testing sets. Hyperparameters for the random forest model were tuned using random search on the validation set. To ensure stability, this process was repeated 50 times with different random splits.
  • Feature Importance: Shapley Additive Explanations (SHAP) values were used to identify the most influential PPG waveform features for BP prediction.

Datasets

The study utilized data from 68 healthy adult participants who underwent overnight polysomnography (PSG) and ambulatory blood pressure monitoring (ABPM) in a sleep laboratory. This dataset was collected specifically for this study and is not a publicly available benchmark.

Key Insights and Ideas

  • Nocturnal BP is Crucial: The paper emphasizes that nocturnal BP is a stronger predictor of cardiovascular events than daytime BP, highlighting the importance of monitoring this specific period.
  • PPG as a Rich Source: The study demonstrates that even a single-channel PPG signal, when analyzed for its intricate waveform morphology, contains sufficient information to estimate BP during sleep.
  • Stiffness Index is Key: The "stiffness index" (SI), a feature reflecting arterial stiffness, emerged as the most important predictor of BP from the PPG waveform. This underscores the link between vascular health and BP estimation.
  • Sleep as an Ideal Monitoring Window: Sleep offers a unique opportunity for BP monitoring due to reduced movement artifacts, making PPG signals more reliable compared to daytime recordings.
  • Calibration-Free Potential: The use of machine learning on PPG waveform features shows promise for a calibration-free approach, a significant advantage over existing methods.

Key Results

  • Accuracy: The random forest model achieved a mean absolute error (MAE) of 5.72 ± 4.51 mmHg for SBP and 4.52 ± 3.60 mmHg for DBP using a 7-second analysis window. The root mean square error (RMSE) was 6.47 ± 1.88 mmHg for SBP and 4.62 ± 1.17 mmHg for DBP.
  • Correlation: The mean correlation coefficient (R) between measured and predicted BP was 0.87 for SBP and 0.86 for DBP.
  • Optimal Window Length: A 7-second window length for PPG analysis yielded the best out-of-sample performance across all tested window lengths (60s, 30s, 15s, 7s, 3s).
  • Feature Importance: The stiffness index (SI) was identified as the most important PPG waveform feature. When demographic variables (age, sex, BMI) were included, BMI became the most important feature.

Limitations and Future Work

  • Sample Homogeneity: The study's sample consisted of young, healthy adults. The generalization of the model to older individuals, those with hypertension, or those with peripheral vascular disease needs to be investigated.
  • BP Range: The BP values assessed in the sample were relatively restricted, which might affect the model's performance on individuals with a wider range of BP.
  • Proportional Bias: A slight proportional bias was observed, where the model tended to overestimate BP at lower true values and underestimate at higher true values. This suggests the method might be more suited for studying the nocturnal physiology of normotensive individuals.

Future work should focus on validating the model in diverse populations, including those with hypertension, and exploring its utility for early identification of individuals at risk for cardiovascular disease. The potential for large-scale deployment using wearable devices with PPG sensors is also a promising avenue.

Summary updated

Abstract

Background: Elevated nocturnal blood pressure (BP) is a risk factor for cardiovascular disease (CVD) and mortality. Cuffless BP assessment aided by machine learning could be a desirable alternative to traditional cuff-based methods for monitoring BP during sleep. We describe a machine-learning-based algorithm for predicting nocturnal BP using single-channel fingertip plethysmography (PPG) in healthy adults. Methods: Sixty-eight healthy adults with no apparent sleep or CVD (53% male), with a median (IQR) age of 29 (23–46 years), underwent overnight polysomnography (PSG) with fingertip PPG and ambulatory blood pressure monitoring (ABPM). Features based on pulse morphology were extracted from the PPG waveforms. Random forest models were used to predict night-time systolic blood pressure (SBP) and diastolic blood pressure (DBP). Results: Our model achieved the highest out-of-sample performance with a window length of 7 s across window lengths explored (60 s, 30 s, 15 s, 7 s, and 3 s). The mean absolute error (MAE ± STD) was 5.72 ± 4.51 mmHg for SBP and 4.52 ± 3.60 mmHg for DBP. Similarly, the root mean square error (RMSE ± STD) was 6.47 ± 1.88 mmHg for SBP and 4.62 ± 1.17 mmHg for DBP. The mean correlation coefficient between measured and predicted values was 0.87 for SBP and 0.86 for DBP. Based on Shapley additive explanation (SHAP) values, the most important PPG waveform feature was the stiffness index, a marker that reflects the change in arterial stiffness. Conclusion: Our results highlight the potential of machine learning-based nocturnal BP prediction using single-channel fingertip PPG in healthy adults. The accuracy of the predictions demonstrated that our cuffless method was able to capture the dynamic and complex relationship between PPG waveform characteristics and BP during sleep, which may provide a scalable, convenient, economical, and non-invasive means to continuously monitor blood pressure.
BibTeX
@inproceedings{yilmaz2023nocturnal,
  author = {Gizem Yilmaz and Xingyu Lyu and Ju Lynn Ong and Lieng Hsi Ling and Thomas Penzel and B. T. Thomas Yeo and Michael W. L. Chee},
  title = {{Nocturnal Blood Pressure Estimation from Sleep Plethysmography Using Machine Learning}},
  booktitle = {Sensors},
  year = {2023},
  url = {https://openreview.net/forum?id=uXYDnMJESK},
}
Figure