A 1-Dimensional Physiological Signal Prediction Method Based on Composite Feature Preprocessing and Multi-Scale...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Accurate, real-time monitoring of physiological signals like intracranial pressure (ICP) and arterial blood pressure (BP) is critical for clinical care.
• Traditional invasive monitoring methods pose risks such as infection and procedural complexity, limiting continuous measurement.
• Learning-based prediction using observable signals offers a promising non-invasive alternative, but existing models face challenges in capturing multi-scale features and long-range dependencies efficiently.

Purpose of the Study:

• To develop an efficient and accurate method for non-invasive physiological signal prediction, addressing limitations of existing models.
• To improve the capture of both local waveform details and long-range temporal dependencies in physiological signals.
• To provide a computationally efficient solution for real-time clinical demands in physiological monitoring.

Main Methods:

• A composite feature preprocessing step constructs a seven-dimensional feature matrix to enhance discriminative power and mitigate phase mismatch.
• A novel CNN-LSTM-Attention (CBAnet) architecture integrates multiscale convolutions, Long Short-Term Memory (LSTM), and attention mechanisms.
• The model is evaluated on GBIT-ABP, CHARIS, and a self-built PPG-HAF dataset, comparing performance against BiLSTM, CNN-LSTM, Transformer, and Wave-U-Net.

Main Results:

• CBAnet demonstrates competitive performance across Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²).
• The proposed method effectively captures multi-scale local features and long-range temporal dependencies in physiological waveforms.
• Experimental results validate the model's superior accuracy and temporal consistency compared to baseline methods.

Conclusions:

• The developed CBAnet offers a promising and efficient approach for non-invasive, continuous physiological parameter prediction.
• This method addresses the limitations of existing models in handling complex physiological signal dynamics and computational demands.
• The findings support the potential of advanced machine learning techniques for improving real-time patient monitoring and clinical decision-making.