A Machine Learning Framework for Biometric Authentication using Electrocardiogram

Song-Kyoo Kim, Chan Yeob Yeun, Ernesto Damiani, Nai-Wei Lo

TL;DR

This paper introduces a comprehensive machine learning framework for ECG-based biometric authentication, addressing challenges in data quality and standardization. It proposes novel pre-processing techniques, a time-slicing method for dataset generation, and four new data quality metrics, along with a publicly available MATLAB toolbox to facilitate research and development in this area.

What Problem Does This Paper Address?

Biometric authentication using Electrocardiograms (ECGs) is a promising area due to its ability to use live body signals. However, existing research often lacks standardization in data acquisition, pre-processing, and evaluation, hindering the development of robust and reliable ECG-based authentication systems. Specifically, challenges include:

  • Data Quality: ECG signals can be noisy due to various factors like baseline drift, power line interference, and signal flipping caused by improper electrode placement.
  • Dataset Standardization: Defining appropriate datasets for training and testing machine learning models is crucial for achieving high accuracy.
  • Evaluation Metrics: Existing metrics may not fully capture the nuances of ECG data quality for authentication purposes.
  • Application Scenarios: Different use cases (e.g., hospital, security, personal devices) have varying requirements and constraints that need to be considered.

This paper aims to bridge these gaps by providing a structured framework and tools to improve the design and evaluation of ECG-based biometric authentication systems.

Methods

The paper proposes a framework that encompasses three main stages:

  1. Use Case Categorization: Three distinct use cases are identified to tailor authentication schemes to specific environments:
  2. Hospital (HOS): For patient identification within a healthcare setting.
  3. Security Check (SCK): For access control at building entrances.
  4. Wearable Devices (WD): For continuous authentication on personal devices like smartwatches.
  5. Pre-processing Techniques: To enhance ECG data quality, the authors introduce three novel pre-processing methods:
  6. Baseline Adjustment: Using polynomial curve fitting to remove low-frequency baseline drift.
  7. Power Line Interference (PLI) Noise Removal: Employing techniques to filter out 50/60 Hz interference common in medical equipment.
  8. Flipping Signal Correction: A method to detect and correct flipped ECG signals due to incorrect electrode placement.
  9. Core Process (Machine Learning):
    • Time Slicing: ECG data is segmented into fixed-length slices (e.g., 0.6 seconds) anchored by the R-peak. This creates numerous samples for machine learning training.
  10. Machine Learning Models: The framework supports various ML techniques for both regression (Decision Tree, Support Vector Machine) and classification (Artificial Neural Network, Convolutional Neural Network).
  11. Data Quality Metrics: Four new metrics are introduced to evaluate the quality of training and testing data:
  12. Mean Absolute Error Rate (MAER): Measures the error rate relative to a moving mean, designed to avoid division by zero.
  13. Upper/Lower Range Control Limits (UCL, LCL): Based on statistical process control, these define acceptable data ranges.
  14. Accuracy Percentage within Ranges (APR): The proportion of ECG data falling within the defined control limits.
  15. Accuracy per UCL (APU): A ratio of APR to UCL, indicating how much training has improved accuracy relative to the control limit.

A MATLAB toolbox named amgecg Toolbox is developed and made publicly available, containing functions for these pre-processing steps, time slicing, and data quality analysis.

Datasets

The paper mentions the use of ECG datasets from the PhysioBank database for evaluating regression models (specifically comparing DT and SVM). No new benchmark datasets are introduced, but the proposed framework and toolbox are designed to work with various ECG datasets.

Key Insights and Ideas

  • Use Case Driven Design: Recognizing that ECG authentication needs vary significantly across applications (hospital vs. wearable) is crucial for effective system design.
  • Data Quality is Paramount: The paper strongly emphasizes that robust pre-processing and rigorous data quality assessment are fundamental for successful ML-based ECG authentication, rather than solely relying on complex ML models.
  • Time Slicing for Data Augmentation: The R-peak anchored time-slicing technique is an effective way to generate a large number of training samples from limited ECG recordings, making ML training more feasible.
  • Novel Metrics for Nuanced Evaluation: The introduction of MAER, UCL/LCL, APR, and APU provides a more granular and context-aware way to assess data quality and model performance, going beyond traditional metrics like MSE or RMSE.
  • Open-Source Tooling: The release of the amgecg Toolbox is a significant contribution, lowering the barrier to entry for researchers and developers working on ECG biometrics.

Key Results

The paper focuses on presenting the framework and its components rather than reporting specific quantitative performance results of the ML models themselves. However, it does provide a comparison of Decision Tree (DT) and Support Vector Machine (SVM) for regression tasks using the PhysioBank dataset.

  • DT vs. SVM for Regression: The paper indicates that the Decision Tree (DT) technique delivered better performance compared to the Support Vector Machine (SVM) technique for building the regression model on the PhysioBank dataset. Specific metrics like RMSE, R-Squared, MSE, and MAE are presented in Table II, showing DT generally outperforming SVM in this specific comparison. For instance, DT achieved an RMSE of 33.761 compared to SVM's 36.168.
  • Prediction Speed: DT demonstrated significantly faster prediction speeds (~740,000 obs/sec) compared to SVM (~600 obs/sec).
  • Training Time: DT also had a much shorter training time (5.2194 sec) than SVM (2008.2 sec).

The authors highlight that the framework is designed to guide researchers in preparing ML setups and datasets, and the toolbox can be used to evaluate the applicability of the proposed framework.

Limitations and Future Work

  • Optimization of Time Slicing: The paper acknowledges that the optimal time-slicing window length is not determined and suggests that optimizing this parameter for biometrics purposes is an interesting area for future research.
  • Specific ML Model Performance: While the framework supports various ML models, the paper does not delve into optimizing or comparing the performance of specific Neural Network architectures (ANN, CNN) for classification. It states that designing proper NN models is another research topic.
  • Use Case Specific Metrics: The paper notes that while the authentication use cases are similar, the performance measure schemes shall be different, implying that further research might be needed to tailor the proposed metrics to each specific use case.
  • Generalizability of ML Model Choice: The authors mention that the optimal selection of a proper ML model for a regression approach might vary by dataset or computing system, suggesting that further empirical studies are needed.

Summary updated

BibTeX
@article{kim2019machine,
  author = {Song-Kyoo Kim and Chan Yeob Yeun and Ernesto Damiani and Nai-Wei Lo},
  title = {{A Machine Learning Framework for Biometric Authentication using Electrocardiogram}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2927079},
}
Figure