Technical Summary: A Systematic Review of ECG Arrhythmia Classification

Problem Statement

The classification of electrocardiogram (ECG) signals is vital for the early detection of arrhythmias and cardiac conditions. However, despite advancements in machine learning, a significant portion of the literature fails to adhere to key standardization protocols, leading to inconsistent performance evaluations and limited real-world applicability. Two primary issues undermine the reliability of current research:

1. Evaluation Bias: Many studies utilize "intra-patient" data partitioning, where heartbeats from the same patient appear in both training and testing sets. This practice inflates accuracy metrics by allowing models to memorize patient-specific features rather than learning generalizable arrhythmia patterns.
2. Embedded Feasibility Gap: Numerous studies prioritize classification accuracy without assessing the feasibility of deploying these models on resource-constrained embedded systems (e.g., pacemakers, Holter monitors, wearable patches). High computational complexity and memory footprints often render these models impractical for low-power, real-time cardiac monitoring devices.

Methodology

This study conducted a systematic review of ECG classification research published between 2017 and 2024. The methodology followed a rigorous multi-stage selection process:

• Data Source: An initial dataset of 1,427 articles was retrieved from the Web of Science database using the query TS=("ECG classification" OR "arrhythmia detection").
• Filtering Criteria:
  • Index: Restricted to the Science Citation Index Expanded (SCI-EXPANDED) to ensure high-quality, peer-reviewed research.
  • Timeframe: 2017–2024.
  • Content: Titles and abstracts were screened for relevance to arrhythmia or ECG classification.
  • Quality Control: Excluded non-English papers, retracted articles, duplicates, and survey/review papers.
• Selection Strategy: The final set of 122 articles was selected based on two criteria:
  1. Citation Impact: The 70 most cited articles, plus the top 10 most cited from 2022, 2023, and 2024 to capture recent high-impact work.
  2. Embedded Feasibility: An additional 22 studies were included specifically for their focus on lightweight architectures, quantization, pruning, FPGA implementations, or on-device learning.
• Evaluation Framework (E3C): The review introduced the E3C (Embedded, Clinical, and Comparative Criteria) framework to assess studies. To be considered fully compliant, a study must:
  • Adhere to AAMI (Association for the Advancement of Medical Instrumentation) recommendations, including the use of the MIT-BIH Arrhythmia Database and specific beat class definitions.
  • Implement the Inter-Patient Paradigm (proposed by De Chazal et al.), ensuring strict separation of patient data between training and testing sets.
  • Evaluate Embedded Feasibility, demonstrating the model's viability on low-power hardware (e.g., microcontrollers, FPGAs).

Key Contributions

1. Identification of Critical Gaps: The review reveals that only 4.1% (5 out of 122) of the analyzed studies simultaneously satisfy all E3C criteria. The majority of high-impact studies rely on intra-patient evaluation or fail to address hardware constraints.
2. Citation Network Analysis: The authors mapped the influence of studies, noting that works not adhering to the inter-patient paradigm (often using intra-patient splits) have historically dominated the field in terms of citations, potentially skewing the perception of model robustness.
3. State-of-the-Art (SOTA) Identification: The paper identifies and analyzes the five studies that meet all E3C criteria, providing a comparative analysis of their performance, energy consumption, and inference times.
4. Optimization Techniques Review: The paper synthesizes current methods for enabling embedded deployment, including:
   • Pruning: Removing redundant connections to reduce model size.
   • Quantization: Reducing bit-width (e.g., 32-bit to 8-bit) to lower memory and computational demands.
   • Lightweight Architectures: Utilizing models like MobileNet, SqueezeNet, and Spiking Neural Networks (SNNs).
   • FPGA Acceleration: Leveraging reconfigurable hardware for parallel processing.
   • On-Device Learning: Enabling real-time model adaptation to patient-specific variations without external retraining.

Results

• Adherence Statistics:
  • 55.7% of studies followed AAMI guidelines.
  • 30.3% adhered to the inter-patient paradigm.
  • 78.7% utilized the MIT-BIH database.
  • 31.1% considered embedded feasibility.
  • Only 4.1% met all E3C criteria.
• Performance of E3C-Compliant Studies: The five identified studies represent the current frontier of robust, deployable ECG classification:
  • Farag et al. [15]: Proposed a Matched Filter-based 1D CNN optimized for edge devices. It achieved the highest average accuracy (98.18%) and the fastest inference time (<1 ms) on a Raspberry Pi, with a model size of 15 KB. It demonstrated strong sensitivity for minority classes (SVEB and VEB).
  • Mao et al. [17]: Developed an ultra-low energy processor integrating Artificial Neural Networks (ANN) and Spiking Neural Networks (SNN). It featured on-chip learning, allowing patient-specific calibration that improved accuracy from 93.67% to 97.36%. It achieved the lowest energy consumption (0.3 µJ/beat) and the smallest model size (8 KB).
  • Yan et al. [16]: Utilized a two-stage CNN converted to SNN for inference, achieving high energy efficiency (77 mW) but moderate accuracy (90%).
  • Xing et al. [18]: Implemented an SNN with a Channel-Wise Attention Mechanism on an FPGA, achieving 92.07% accuracy under inter-patient evaluation with low latency (1.32 ms).
  • Raj and Ray [31]: Combined Discrete Orthogonal Stockwell Transform (DOST) with an optimized SVM, achieving 86.89% accuracy in inter-patient settings.
• Trade-offs: The analysis highlights a trade-off between accuracy and efficiency. While Farag et al. offer superior accuracy and speed, Mao et al. provide the most energy-efficient solution with the unique capability of continuous on-device adaptation.

Significance and Claims

The paper claims that its primary significance lies in shifting the focus of ECG classification research from purely accuracy-driven metrics to clinically viable and practically deployable systems. By highlighting the scarcity of studies that adhere to AAMI standards, the inter-patient paradigm, and embedded constraints, the authors argue that the field is currently overestimating model performance due to biased evaluation protocols.

The study asserts that future research must prioritize the E3C framework to ensure that machine learning models are not only accurate but also robust to patient variability and feasible for deployment in resource-constrained medical devices. The authors propose standardized reporting guidelines that include class-specific metrics (Sensitivity, Positive Predictive Value, F1-score) for all five AAMI beat classes, alongside hardware-specific metrics (energy, memory, inference time), to facilitate fair comparisons and accelerate the development of reliable, real-time cardiac monitoring solutions.