A Two-Branch ResNet-BiLSTM Deep Learning Framework for Extracting Multimodal Features Applied to PPG-Based Cuffless...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Cardiovascular Health

Background:

• Cardiovascular disease poses a significant health risk, with blood pressure levels being a key indicator.
• Continuous blood pressure monitoring is crucial but hindered by the inconvenience of traditional cuff-based devices.
• Existing deep learning approaches for cuffless blood pressure estimation often lack interpretability, limiting accuracy.

Purpose of the Study:

• To develop a novel, accurate, and interpretable deep learning framework for cuffless blood pressure estimation using photoplethysmography (PPG).
• To combine the strengths of Residual Networks (ResNet) and Bidirectional Long Short-Term Memory (BiLSTM) for enhanced PPG signal analysis.
• To validate the proposed method against established clinical standards.

Main Methods:

• A two-branch deep learning architecture integrating ResNet and BiLSTM was designed.
• The ResNet branch processed 60 features selected via Support Vector Machine-Recursive Feature Elimination (SVM-RFE), including novel trend features.
• The BiLSTM branch analyzed complete PPG waveforms.
• The model was trained and tested on the MIMIC-IV dataset, comprising 220 waveform segments from 218 patients.

Main Results:

• The proposed ResNet-BiLSTM model achieved a mean absolute error of 3.47 mmHg for systolic blood pressure and 2.81 mmHg for diastolic blood pressure.
• Standard deviations were 5.06 mmHg (systolic) and 4.11 mmHg (diastolic).
• Performance met the Association for the Advancement of Medical Instrumentation (AAMI) standards and achieved an 'A' rating by British Hypertension Society (BHS) standards.

Conclusions:

• The novel two-branch deep learning framework offers a promising solution for accurate and convenient cuffless blood pressure estimation.
• The combination of ResNet and BiLSTM effectively leverages both extracted features and raw PPG waveforms.
• The achieved performance validates the clinical applicability and potential of this advanced PPG-based monitoring technique.