Classification of Pulse Waveforms Using Edit Distance with Real Penalty

Dongyu Zhang, Wangmeng Zuo, David Zhang, Hongzhi Zhang, Naimin Li

Introduction

Traditional Chinese Pulse Diagnosis (TCPD) has served as a primary diagnostic tool for millennia, yet its clinical application remains highly subjective, dependent on the tactile sensitivity and experience of the practitioner. Modernizing this practice requires the objective classification of pressure waveforms into distinct patterns such as "taut," "smooth," or "hollow."

However, automating TCPD presents significant technical hurdles. Biological signals exhibit high intraclass variation—where pulses from the same category differ due to individual physiology—and local time shifting, where peaks and valleys do not align perfectly across different heartbeats. To address these challenges, researchers have developed novel classifiers that combine "elastic" time-series matching with advanced $k$-nearest neighbor (KNN) frameworks.

System Architecture

The proposed diagnostic system integrates hardware acquisition with a multi-stage software pipeline. The process involves both online real-time classification and offline database management for model training.

System architecture for pulse waveform acquisition, preprocessing, and classification.

As illustrated in the system flowchart, the pipeline begins with pulse-waveform acquisition, followed by a rigorous preprocessing stage. This stage includes denoising, baseline wander removal, and period segmentation to ensure the signal is clean and standardized before feature extraction and classification.

Methodology: Elastic Similarity Measures

Rather than relying on rigid Euclidean distance measures, which perform poorly on misaligned biological signals, this study utilizes Edit Distance with Real Penalty (ERP). ERP is a metric capable of aligning sequences of different lengths by penalizing gaps based on actual values, satisfying the triangle inequality—a property critical for stable kernel-based learning.

Core Technical Contributions

  • ERP-Induced Inner Product: The authors derived a method to transform ERP distance into an inner product, allowing elastic matching to be used within linear algebraic frameworks.
  • Gaussian ERP (GERP) Kernel: By embedding ERP distance into a Gaussian radial basis function, the researchers enabled non-linear classification in high-dimensional feature spaces.
  • Advanced KNN Frameworks: These measures were integrated into Difference-Weighted KNN (DFWKNN). Unlike standard majority-vote KNN, this framework assigns optimal weights to neighbors to minimize reconstruction error, resulting in two primary classifiers: EDKC (ERP-based) and GEKC (Gaussian ERP Kernel-based).

Data Acquisition and Preprocessing

The study utilized a substantial dataset of 2,470 pulse waveforms collected at Harbin Binghua Hospital using a custom-built motor-embedded pressure sensor. This represents one of the largest pulse datasets reported in the literature.

Raw pulse waveform showing baseline wander caused by physiological factors like respiration.

Preprocessing is vital because raw signals often contain artifacts. As shown above, baseline wander (often caused by breathing) must be corrected. Signals were denoised using Daubechies 4 wavelets and segmented into single periods of 150 points. Each sample was labeled based on the consensus of three TCM experts across five categories: Moderate, Smooth, Taut, Hollow, and Unsmooth.

Key Results and Insights

The proposed methods were benchmarked against standard 1NN-Euclidean and Dynamic Time Warping (DTW) approaches.

  • Top Performance: The GEKC method achieved the highest average accuracy of 91.74%.
  • Reliability: The EDKC followed closely at 90.36%, both significantly outperforming 1NN-Euclidean (87.36%) and 1NN-DTW (83.19%).
  • Class Specifics: The system was particularly effective at identifying "Taut" pulses (96.88% accuracy), which are clinically significant in TCM for diagnosing liver conditions.

The success of the Gaussian ERP kernel suggests that the relationships between different pulse patterns are highly non-linear. By allowing the algorithm to "slide" and "stretch" the time axis, the system effectively mimics the human expert's ability to focus on the overall shape and force of a pulse while ignoring minor temporal variations.

Conclusion and Future Work

While the ERP-based approach offers superior accuracy, it is computationally more intensive than simpler measures. Future research will focus on optimizing the search process using "lower bounds" for ERP distance to facilitate real-time clinical deployment. Additionally, the researchers plan to expand the system's scope to include a wider variety of pulse patterns beyond the initial five studied.

Summary updated

Abstract

Advances in sensor and signal processing techniques have provided effective tools for quantitative research in traditional Chinese pulse diagnosis (TCPD). Because of the inevitable intraclass variation of pulse patterns, the automatic classification of pulse waveforms has remained a difficult problem. In this paper, by referring to the edit distance with real penalty (ERP) and the recent progress in -nearest neighbors (KNN) classifiers, we propose two novel ERP-based KNN classifiers. Taking advantage of the metric property of ERP, we first develop an ERP-induced inner product and a Gaussian ERP kernel, then embed them into difference-weighted KNN classifiers, and finally develop two novel classifiers for pulse waveform classification. The experimental results show that the proposed classifiers are effective for accurate classification of pulse waveform.
BibTeX
@article{zhang2010classification,
  author = {Dongyu Zhang and Wangmeng Zuo and David Zhang and Hongzhi Zhang and Naimin Li},
  title = {{Classification of Pulse Waveforms Using Edit Distance with Real Penalty}},
  journal = {EURASIP J. Adv. Signal Process.},
  year = {2010},
  doi = {10.1155/2010/303140},
}
Figure