TL;DR
This paper introduces a novel deep neural network (DNN) model that estimates blood pressure (BP) using only photoplethysmogram (PPG) signals, eliminating the need for more complex and costly ECG sensors. The model achieves high accuracy, meeting stringent BHS and AAMI standards, and demonstrates superior performance compared to existing PPG-only methods on a large dataset.
What Problem Does This Paper Address?
Cardiovascular diseases (CVDs) are a leading cause of death globally, and hypertension is a major risk factor. Continuous, real-time blood pressure (BP) monitoring is crucial for early detection and management. Traditional cuff-based BP measurement devices are time-consuming and can be inconvenient. While some non-invasive methods exist, they often require multiple sensors, such as electrocardiogram (ECG) and photoplethysmogram (PPG) signals, increasing cost and complexity. This paper addresses the need for a more accessible, cost-effective, and user-friendly BP estimation method by developing a generalized model that relies solely on PPG signals.
Methods
The proposed approach involves a three-stage pipeline:
- Data Preprocessing: Raw PPG signals undergo noise removal (using FFT and IFFT to filter frequencies outside 0-8 Hz), min-max normalization to scale values between 0 and 1, and calculation of the first and second derivatives (dPPG and sdPPG). Crucially, feature points like systolic peaks, onset/offset valleys, and dicrotic notches are detected on the PPG, dPPG, and sdPPG waveforms. Abnormal heart cycles are then removed.
- Feature Extraction and Selection: A comprehensive set of 65 features, previously identified as relevant to BP estimation, are extracted. A novel selection index, denoted by $\gamma$, is introduced to evaluate the correlation between each feature and BP, and to select the 32 most relevant features. This index compares the distribution of standardized features to a standard normal distribution.
- Deep Neural Network Predictor: A fully connected deep neural network (DNN) is employed for BP prediction. The network consists of an input layer receiving the 32 selected features, four hidden layers with an increasing number of neurons (2048, 4096, 8192, 2048), and an output layer predicting Systolic Blood Pressure (SBP) and Diastolic Blood Pressure (DBP). The ReLU activation function is used, and the model is trained using the Nadam optimizer.
Datasets
The study utilizes the MIMIC II online waveform database, which contains data from approximately 12,000 subjects and over 4 million heart cycles. For training, validation, and testing, the researchers preselected 9000 data instances from this dataset. The final evaluation uses approximately 2.17 million PPG, dPPG, and sdPPG data fragments of heart cycles.
Key Insights and Ideas
- PPG-Only Sufficiency: The research strongly suggests that PPG signals alone, when processed with advanced feature extraction and deep learning, are sufficient for accurate blood pressure estimation, obviating the need for ECG.
- Feature Engineering Matters: The paper highlights the importance of carefully selected physiological features derived from PPG and its derivatives, rather than just using raw waveform segments. The introduction of the $\gamma$ index for feature selection is a novel contribution.
- Large-Scale Data is Key: The use of a massive dataset (over 2 million heart cycles) is critical for training a generalized model that performs well across a diverse population.
- DNN Architecture Choice: The selection of a fully connected DNN over recurrent architectures like LSTM is justified by its simpler implementation for wearable devices and avoidance of time delays in critical situations.
Key Results
- High Accuracy: The model achieves a Root Mean Square Error (RMSE) of 4.64 mmHg for SBP and 3.31 mmHg for DBP. The Mean Absolute Error (MAE) is 3.21 mmHg for SBP and 2.23 mmHg for DBP.
- BHS and AAMI Compliance: The model meets the stringent standards set by the British Hypertension Society (BHS) (Grade A for both SBP and DBP) and the Association for the Advancement of Medical Instrumentation (AAMI).
- Error Distribution: 80.63% of estimated SBP records have absolute errors less than 5 mmHg, and 90.19% of estimated DBP records have absolute errors less than 5 mmHg.
- Comparison: The proposed model demonstrates superior performance compared to other PPG-only BP estimation methods on the MIMIC II dataset, achieving better MAE and BHS grades.
Limitations and Future Work
- Dataset Specificity: While the MIMIC II dataset is large, the authors acknowledge that the diversity of data affects model generalization. Future work could explore even more diverse datasets.
- Model Generalization: The paper notes that comparing with other works is challenging due to differences in datasets, input signals (ECG+PPG vs. PPG only), and subject numbers.
- Future Directions:
- Investigate other Recurrent Neural Network (RNN) related models (e.g., sequence-to-sequence) to potentially further reduce MAE and RMSE.
- Explore the correlation between BP estimation and natural time analysis to account for phase changes in blood pressure.
- Implement the model in wearable devices for continuous monitoring in real-world, non-hospital settings.
Summary updated