Advancing Cuffless Arterial Blood Pressure Waveform Estimation: Time-Series Deep Neural Network Approach

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Current deep learning models for arterial blood pressure (ABP) estimation often use sequence-to-sequence (seq2seq) tasks.
• These models may not fully capture temporal dynamics in physiological signals for accurate ABP waveform prediction.

Purpose of the Study:

• To introduce and evaluate a novel time-series training strategy for ABP waveform prediction.
• To compare the performance of seq2seq versus time-series training on deep learning models of varying complexity (gMLP and UtransBPNet).

Main Methods:

• Implemented two deep learning architectures, gMLP and UtransBPNet.
• Trained and evaluated these models using both traditional seq2seq and the proposed time-series training strategies.
• Assessed performance based on mean absolute error (MAE) and out-of-distribution data prediction.

Main Results:

• Time-series training significantly improved ABP estimation performance for both models compared to seq2seq.
• MAE reductions of 2.0 mmHg (gMLP) and 0.9 mmHg (UtransBPNet) were observed.
• The performance gains were more substantial for the simpler gMLP model, and time-series training enhanced out-of-distribution predictive capabilities.

Conclusions:

• The proposed time-series training strategy offers a more effective approach for cuffless arterial blood pressure estimation.
• This method enhances model efficiency and predictive accuracy, making it suitable for edge wearable devices.
• The findings suggest a promising direction for developing next-generation non-invasive BP monitoring solutions.