High School Student Homestead High School Mequon, Wisconsin, United States
Introduction: Hypertension, often referred to as the "silent killer," is the primary cause of cardiovascular and cerebrovascular diseases. A recent report by the World Health Organization (WHO) indicates that the global population with hypertension has doubled in the past two decades (World Health Organization 2023). Nearly half of those affected remain unaware of their condition due to the absence of noticeable symptoms. Early detection and consistent monitoring of blood pressure are crucial for hypertension management. While cuff blood pressure monitors are commonly recommended for managing hypertension at home, they pose inconvenience due to the need for dedicated and consistent time commitment. Research suggests that pulse waves, measured through Photoplethysmogram (PPG) or Electrocardiogram (ECG), may exhibit correlations with blood pressure. The advent of portable and wearable PPG and ECG sensors, integrated into smartwatches, offers an opportunity to develop a method for estimating blood pressure from these signals (Sergio et al., 2023). This study aims to develop and train a deep learning (DL) model for predicting blood pressure from PPG and/or ECG signals. Successful implementation of such an approach could enable continuous blood pressure monitoring with minimal human intervention.
Materials and
Methods: Continuous waveforms of PPG, ECG, and arterial blood pressure (ABP) data were sourced from PhysioNet (https://physionet.org/). Specifically, the MIMIC III dataset was chosen to train the model. Signals were collected from approximately 10,000 de-identified patients admitted to the ICU between 2001 and 2012 at a hospital in Boston, MA.
All original signals were segmented into 20-second intervals. Systolic blood pressure (SBP) and diastolic blood pressure (DBP) were calculated based on the median values of the peaks and troughs in each arterial blood pressure (ABP) data segment. All data segments were divided into a training set (~37,000 segments) and a validation set (~9,300 segments).
A one-dimensional convolutional neural network (1D-CNN) was designed and trained to estimate blood pressure as a regression problem. The CNN model features an input layer (a 20-second PPG and/or ECG signal), an output layer (SBP and DBP values), and 16 inner layers. The CNN was trained with different types of input signals: ECG alone, PPG alone, or a combination of PPG and ECG as a two-channel input.
The trained CNN model was tested using a prototype PPG sensor. A Python program was developed to acquire the PPG signals from the sensor and calculate blood pressure using the trained CNN model. The prototype and model were tested on three human subjects (the student and family members). For each test, the subject's SBP and DBP values were also measured using a commercial-grade cuff blood pressure monitor for comparison.
Results, Conclusions, and Discussions: The validation test results indicate that SBP and DBP values predicted by the trained CNN model correlate well with the ground truth values (Figure 1). The model's performance was evaluated using three key metrics: mean error (ME), standard deviation of errors (STDEV), and mean absolute error (MAE). Table 1 presents these metrics for the CNN model when different input signals were used, alongside comparable metrics from other ML-based models in the literature, all trained on the MIMIC III dataset. The validation tests demonstrate that the model performs slightly better with ECG input than with PPG input. The most accurate results were obtained when both ECG and PPG signals were used as inputs to the CNN model. It achieved a Mean Absolute Error (MAE) of approximately 3.9 mmHg for systolic pressure and 2.0 mmHg for diastolic pressure, outperforming most published machine learning models that use the same dataset. However, in terms of STDEV, the Transformer-based model (Chu et al., 2023) exhibits the best performance.
Using PPG data collected by the prototype PPG sensor, the CNN model's predicted blood pressure values demonstrated a good correlation with those measured by a cuff-based blood pressure monitor (refer to Figure 2). Across 85 measurements from three subjects, the mean errors were 0.3 mmHg for SBP and -0.5 mmHg for DBP, while the mean absolute errors were 6.2 mmHg and 4.0 mmHg for SBP and DBP, respectively.
The proposed method can be integrated into smartwatches, enabling continuous and cuff-less blood pressure monitoring, potentially revolutionizing hypertension management in healthcare technology.
Acknowledgements (Optional): I would like to express my gratitude to my mentor, Professor Jun Zhang from the University of Wisconsin-Milwaukee, who have guided me throughout this project.
References (Optional): Chu, Y., Tang, K., Hsu, Y. C., Huang, T., Wang, D., Li, W., Savitz, S. I., Jiang, X. & Shams, S. (2023). Non-invasive arterial blood pressure measurement and SpO2 estimation using PPG signal: A deep learning framework. BMC Medical Informatics and Decision Making, 23(1), 131. Kachuee, M., Kiani, M. M., Mohammadzade, H., & Shabany, M. (2016). Cuffless blood pressure estimation algorithms for continuous health-care monitoring. IEEE Transactions on Biomedical Engineering, 64(4), 859-869. Slapničar, G., Mlakar, N., & Luštrek, M. (2019). Blood pressure estimation from photoplethysmogram using a spectro-temporal deep neural network. Sensors, 19(15), 3420. Schrumpf, F., Frenzel, P., Aust, C., Osterhoff, G., & Fuchs, M. (2021). Assessment of non-invasive blood pressure prediction from ppg and rppg signals using deep learning. Sensors, 21(18), 6022. Sergio, Gonzalez, Wan-Ting Hsieh, and Trista P. Chen. 2023. “A benchmark for machine-learning based non-invasive blood pressure estimation using photoplethysmogram.” Scientific Data 10 (1): 149. Huang, B., Chen, W., Lin, C. L., Juang, C. F., & Wang, J. (2022). MLP-BP: A novel framework for cuffless blood pressure measurement with PPG and ECG signals based on MLP-Mixer neural networks. Biomedical Signal Processing and Control, 73, 103404. World Health Organization (2023). “First WHO report details devastating impact of hypertension and ways to stop it.” Pan American Health Organization.
