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

Peiquan Chen, Jie Li, Bo Peng, Zhaohui Liu, Liang YanSensors 2025 · DOI

TL;DR

This paper introduces CBAnet, a novel deep learning framework for predicting physiological signals like intracranial pressure (ICP) and blood pressure (BP) from observable signals like pulse waves. By combining a unique 7-dimensional feature preprocessing technique with a multi-scale CNN-LSTM-Attention architecture, CBAnet achieves superior accuracy and temporal consistency compared to existing methods, paving the way for more reliable non-invasive patient monitoring.

What Problem Does This Paper Address?

Precise, real-time monitoring of vital physiological signals like intracranial pressure (ICP) and arterial blood pressure (BP) is crucial for patient care, especially in critical situations. However, traditional invasive methods carry significant risks, including infection and procedural complexity, and are not suitable for continuous monitoring. Non-invasive methods, particularly those leveraging observable signals like photoplethysmography (PPG) or blood pressure (BP) waveforms, offer a promising alternative.

The core challenge lies in the inherent complexity of these physiological signals. They exhibit strong non-linearity, cross-periodic dependencies, and individual variability, making accurate prediction difficult. Existing deep learning models often struggle to simultaneously capture both the fine-grained local details (like waveform peaks and troughs) and the long-range temporal dependencies that characterize these signals. Furthermore, many current approaches are computationally intensive, hindering their deployment in real-time, edge-based monitoring scenarios. This paper aims to bridge this gap by developing a method that is both accurate and efficient for non-invasive physiological signal prediction.

Methods

The proposed solution, named CBAnet, is built upon two key innovations:

  1. Composite Feature Preprocessing: Instead of using raw one-dimensional signals, the authors construct a seven-dimensional feature matrix. This matrix is derived from physiological signal priors and includes:
  2. The raw waveform

    • Amplitude
    • Pulse width
    • Rise time
    • Fall time
    • First derivative (rate of change)
    • Second derivative (curvature)

    This multi-dimensional representation aims to capture richer morphological and trend information, reduce learning complexity, and mitigate phase mismatch issues.

  3. Multi-Scale Modeling Architecture (CBAnet): This network integrates three main components:

  4. Local Time-Domain Feature Extraction (LFE): Utilizes lightweight, two-layer convolutional networks to extract stable, low-level features like waveform geometric structures without sacrificing temporal resolution. This part focuses on capturing details like rising edges and reentrant notches.
  5. Long-Term Dependency Representation (TRN): Employs bidirectional Long Short-Term Memory (LSTM) networks to model long-range temporal dependencies and sequential biases. This captures patterns across multiple cardiac cycles.
  6. Global Reweighted Alignment (GRA): Incorporates a multi-head self-attention mechanism. This allows the model to directly "see" and align features from any two temporal steps within a sequence, effectively reweighting and aligning data within time windows to focus on critical segments and mitigate misalignment.

    The combination of these modules allows CBAnet to progressively model the signal from local details to global trends and then align them for accurate prediction.

Datasets

The authors evaluated their model on three datasets:

  • PPG-HAF (Self-Built): A newly established dataset created by the authors, featuring synchronized PPG signals acquired simultaneously from the head and fingertip. This dataset was developed to address the lack of publicly available data with concurrent head and fingertip PPG signals.
  • GBIT-ABP: A public dataset containing multi-channel bioimpedance signals, along with corresponding BP and PPG signals. The authors used the simultaneously recorded fingertip PPG and BP signals.
  • CHARIS: A public dataset comprising multi-channel physiological signals from patients diagnosed with traumatic brain injury (TBI), including ECG, BP, and ICP. The authors utilized the simultaneously recorded BP and ICP waveform data.

Key Insights and Ideas

  • Feature Engineering Matters: Expanding a simple 1D signal into a 7D feature space based on physiological priors significantly boosts prediction accuracy. This highlights the value of incorporating domain knowledge into data preprocessing for deep learning models.
  • Multi-Scale is Key: The success of CBAnet lies in its ability to simultaneously capture both the rapid, localized changes in waveforms (via CNNs) and the slower, overarching trends (via LSTMs and attention). This multi-scale approach is crucial for understanding complex physiological signals.
  • Attention for Alignment: The attention mechanism plays a vital role in aligning temporal segments, helping to overcome phase mismatches and improve the fidelity of waveform reconstruction, especially in the presence of noise or individual variations.
  • Bridging the Gap to Real-Time: The moderate parameter count and efficient inference of CBAnet make it a viable candidate for real-time, bedside monitoring, addressing a critical need in clinical settings.

Key Results

CBAnet demonstrated superior performance across all three datasets and compared to several baseline models (BiLSTM, CNN-LSTM, Transformer, Wave-U-Net).

  • Accuracy Improvements: On the CHARIS dataset, CBAnet achieved a significant reduction in RMSE (by 45.9%) and MAE (by 50.3%), with an increase in R² (by 39%) compared to the next-best model.
  • Overall Performance: CBAnet consistently achieved the best RMSE, MAE, and R² metrics across the GBIT-ABP, CHARIS, and PPG-HAF datasets. For example, on the CHARIS dataset, it achieved an RMSE of 0.4903 and an R² of 0.8451.
  • Generalization: The model showed consistent improvements across different datasets, indicating good cross-dataset generalization capabilities.
  • Bland-Altman Analysis: CBAnet consistently exhibited the smallest bias, smallest standard deviation (SD), and narrowest consistency limits across all datasets, indicating its predictions closely align with actual signals and are robust to variations.

Limitations and Future Work

  • Limited Data Cohort: The current model has been validated using data from a single cohort. Future work should involve validation on more diverse populations and longer-term monitoring scenarios (beyond short-term).
  • Single Signal Prediction: The study primarily focused on predicting a single physiological signal. The authors suggest future research could explore multi-parameter cascade prediction frameworks (e.g., predicting ICP from PPG, then BP from ICP).
  • Clinical Applicability: While the model shows near-real-time processing potential, further validation is required for clinical application, especially concerning extended recording durations, diverse cohorts, and consistency analysis.
  • Dataset Diversity: As high-quality, diverse datasets become more available, model architectures can be further optimized to reduce dependence on specific data distributions, enhancing both accuracy and interpretability.

Summary updated

Abstract

The real-time, precise monitoring of physiological signals such as intracranial pressure (ICP) and arterial blood pressure (BP) holds significant clinical importance. However, traditional methods like invasive ICP monitoring and invasive arterial blood pressure measurement present challenges including complex procedures, high infection risks, and difficulties in continuous measurement. Consequently, learning-based prediction utilizing observable signals (e.g., BP/pulse waves) has emerged as a crucial alternative approach. Existing models struggle to simultaneously capture multi-scale local features and long-range temporal dependencies, while their computational complexity remains prohibitively high for meeting real-time clinical demands. To address this, this paper proposes a physiological signal prediction method combining composite feature preprocessing with multiscale modeling. First, a seven-dimensional feature matrix is constructed based on physiological prior knowledge to enhance feature discriminative power and mitigate phase mismatch issues. Second, a network architecture CNN-LSTM-Attention (CBAnet), integrating multiscale convolutions, long short-term memory (LSTM), and attention mechanisms is designed to effectively capture both local waveform details and long-range temporal dependencies, thereby improving waveform prediction accuracy and temporal consistency. Experiments on GBIT-ABP, CHARIS, and our self-built PPG-HAF dataset show that CBAnet achieves competitive performance relative to bidirectional long short-term Memory (BiLSTM), convolutional neural network-long short-term memory network (CNN-LSTM), Transformer, and Wave-U-Net baselines across Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R<sup>2</sup>). This study provides a promising, efficient approach for non-invasive, continuous physiological parameter prediction.
BibTeX
@inproceedings{chen2025dimensional,
  author = {Peiquan Chen and Jie Li and Bo Peng and Zhaohui Liu and Liang Yan},
  title = {{A 1-Dimensional Physiological Signal Prediction Method Based on Composite Feature Preprocessing and Multi-Scale Modeling}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25216726},
}
Figure