Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Optimized Hybrid Feature Engineering for Resource-Efficient Arrhythmia Detection in ECG Signals: An Optimization Framework

Moirangthem Tiken Singh, Manibhushan Yaikhom

TL;DR

This work argues that domain-grounded, hybrid feature engineering can render complex ECG arrhythmia data linearly separable, enabling ultra-lightweight, interpretable linear classifiers that approach deep-learning performance. By fusing wavelet-based time-frequency analysis with graph-theoretic descriptors and HRV metrics, the authors construct a compact feature space that supports efficient edge AI in IoMT contexts. Across MIT-BIH and INCART benchmarks, Linear SVC and Logistic Regression achieve high diagnostic accuracy with model footprints below 9 KB and per-beat latency in the microsecond range, significantly reducing energy and memory requirements compared to DL baselines. The study validates a data-centric paradigm for real-time, battery-less cardiac monitoring, while identifying limitations and outlining hardware acceleration as a path to broader deployment.

Abstract

Cardiovascular diseases, particularly arrhythmias, remain a leading global cause of mortality, necessitating continuous monitoring via the Internet of Medical Things (IoMT). However, state-of-the-art deep learning approaches often impose prohibitive computational overheads, rendering them unsuitable for resource-constrained edge devices. This study proposes a resource-efficient, data-centric framework that prioritizes feature engineering over complexity. Our optimized pipeline makes the complex, high-dimensional arrhythmia data linearly separable. This is achieved by integrating time-frequency wavelet decompositions with graph-theoretic structural descriptors, such as PageRank centrality. This hybrid feature space, combining wavelet decompositions and graph-theoretic descriptors, is then refined using mutual information and recursive elimination, enabling interpretable, ultra-lightweight linear classifiers. Validation on the MIT-BIH and INCART datasets yields 98.44% diagnostic accuracy with an 8.54 KB model footprint. The system achieves 0.46 $μ$s classification inference latency within a 52 ms per-beat pipeline, ensuring real-time operation. These outcomes provide an order-of-magnitude efficiency gain over compressed models, such as KD-Light (25 KB, 96.32% accuracy), advancing battery-less cardiac sensors.

Optimized Hybrid Feature Engineering for Resource-Efficient Arrhythmia Detection in ECG Signals: An Optimization Framework

TL;DR

This work argues that domain-grounded, hybrid feature engineering can render complex ECG arrhythmia data linearly separable, enabling ultra-lightweight, interpretable linear classifiers that approach deep-learning performance. By fusing wavelet-based time-frequency analysis with graph-theoretic descriptors and HRV metrics, the authors construct a compact feature space that supports efficient edge AI in IoMT contexts. Across MIT-BIH and INCART benchmarks, Linear SVC and Logistic Regression achieve high diagnostic accuracy with model footprints below 9 KB and per-beat latency in the microsecond range, significantly reducing energy and memory requirements compared to DL baselines. The study validates a data-centric paradigm for real-time, battery-less cardiac monitoring, while identifying limitations and outlining hardware acceleration as a path to broader deployment.

Abstract

Cardiovascular diseases, particularly arrhythmias, remain a leading global cause of mortality, necessitating continuous monitoring via the Internet of Medical Things (IoMT). However, state-of-the-art deep learning approaches often impose prohibitive computational overheads, rendering them unsuitable for resource-constrained edge devices. This study proposes a resource-efficient, data-centric framework that prioritizes feature engineering over complexity. Our optimized pipeline makes the complex, high-dimensional arrhythmia data linearly separable. This is achieved by integrating time-frequency wavelet decompositions with graph-theoretic structural descriptors, such as PageRank centrality. This hybrid feature space, combining wavelet decompositions and graph-theoretic descriptors, is then refined using mutual information and recursive elimination, enabling interpretable, ultra-lightweight linear classifiers. Validation on the MIT-BIH and INCART datasets yields 98.44% diagnostic accuracy with an 8.54 KB model footprint. The system achieves 0.46 s classification inference latency within a 52 ms per-beat pipeline, ensuring real-time operation. These outcomes provide an order-of-magnitude efficiency gain over compressed models, such as KD-Light (25 KB, 96.32% accuracy), advancing battery-less cardiac sensors.
Paper Structure (15 sections, 12 equations, 10 figures, 5 tables)

This paper contains 15 sections, 12 equations, 10 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Literature Survey
  3. Methodology
  4. Experimental Setup
  5. Data Preparation and Processing Analysis
  6. Hyperparameter Optimization Analysis
  7. Feature Space Analysis and Pipeline Efficacy
  8. Classification Performance and Resource Efficiency
  9. Diagnostic Accuracy
  10. Computational Resource Profiling and Clinical Feasibility
  11. Statistical Significance and Reliability
  12. Model Interpretability and Feature Relevance
  13. Comparative Analysis with State-of-the-Art
  14. Discussion
  15. Conclusion

Figures (10)

  • Figure 1: ECG Classification Pipeline for Arrhythmia Detection
  • Figure 2: Comparison of raw and preprocessed ECG signals, illustrating the effect of filtering and noise removal
  • Figure 3: Ensemble Approach for R-Peak Detection with SQI Weighting
  • Figure 4: Directed Graph for graph augmentation.
  • Figure 5: Hyperparameter Optimization Landscape for Adaptive Segmentation. (Upper Left) Mean loss heatmap for pre-RR ($\alpha$) and post-RR ($\beta$) fractions, showing the sensitivity of segmentation performance to $\beta$. The Exact Global Minimum (red star) and Robust Global Optimum (blue diamond) are marked. (Upper Right) Loss profiles vs. $\beta$ for fixed $\alpha$ slices, confirming a consistent minimum near $\beta \approx 0.37$. (Lower Left) Correlation between $\beta$ and segment length (ms); lower loss (darker points) clusters in the 900–1200 ms range. (Lower Right) Beat-specific optima, illustrating the stability of $\beta$ (orange) versus the variance of $\alpha$ (blue) across cardiac cycles.
  • ...and 5 more figures