TL;DR
This paper introduces a novel human identification system using electrocardiogram (ECG) signals. By employing Ensemble Empirical Mode Decomposition (EEMD) for robust feature extraction and K-nearest neighbors (K-NN) for classification, the system achieves high accuracy (95%) and demonstrates robustness against noise and heart rate variability, making ECG a viable biometric.
What Problem Does This Paper Address?
Traditional biometric systems like fingerprints or facial recognition can be vulnerable to spoofing or environmental factors. ECG, as a physiological signal generated internally, offers unique advantages in terms of universality, permanence, and difficulty to mimic. However, ECG signals are susceptible to noise and variations in heart rate (HRV), which can hinder accurate identification. This paper aims to develop an ECG-based identification system that effectively handles these challenges to provide a reliable biometric solution.
Methods
The proposed system involves several key stages:
- Preprocessing: A robust preprocessing pipeline is implemented to address noise and HRV. This includes:
- Noise Elimination: A combination of detrending and wavelet minimax thresholding is used to remove various noise sources like baseline drift, power line interference, and electrode motion artifacts.
- Heartbeat Normalization: A method to linearly normalize heartbeats to a standard rate (75 bpm) is applied to mitigate the effects of HRV. This involves QRS complex detection, segmentation of heartbeats, and resampling.
- Quality Measure: A periodicity transform-based method is used to assess the quality of the normalized heartbeat signal, ensuring that only high-quality segments are used for further processing. A threshold of 0.8 is set for signal acceptance.
- Feature Extraction:
- Ensemble Empirical Mode Decomposition (EEMD): This technique is used to decompose the preprocessed ECG signal into intrinsic mode functions (IMFs). EEMD is chosen over standard EMD to overcome the "mode mixing" problem, leading to more meaningful and distinct components.
- Welch Spectral Analysis: This method is applied to estimate the power spectrum of each IMF component.
- MIMFs Selection: The most important IMFs (MIMFs), identified as those containing the most significant signal energy (typically IMFs 1-4), are selected.
- Feature Fusion: The temporal waveforms and power spectra of the selected MIMFs are fused to create a feature set.
- Principal Component Analysis (PCA): PCA is employed as a dimensionality reduction technique to reduce the size of the feature space, improving classification speed and efficiency.
- Classification:
- K-Nearest Neighbors (K-NN): The K-NN algorithm is used as the classifier. The similarity between features is measured using Euclidean distance.
Datasets
The system was evaluated using data from three public ECG databases:
- MIT-BIH ST change database: Contains 28 ECG recordings, some from exercise stress tests.
- Long-term ST database: Includes 86 lengthy recordings from 80 subjects, focusing on ST segment changes.
- PTB database: Compiles ECG signals from 294 subjects with various clinical conditions, including healthy individuals.
The experiments involved selecting ECG segments from these databases, with a total of 90 subjects used for the final evaluation.
Key Insights and Ideas
- EEMD's Advantage: EEMD proves superior to standard EMD in ECG processing by effectively mitigating mode mixing, leading to more interpretable and discriminative IMFs.
- Heart Rate Independence: The preprocessing steps, particularly heartbeat normalization, are crucial for making the identification system independent of heart rate variability, a significant challenge in ECG biometrics.
- Feature Fusion Power: Combining both temporal waveforms and spectral information from the EEMD components (MIMFs) yields a richer feature set, leading to higher identification accuracy than using either alone.
- PCA for Efficiency: PCA significantly reduces the dimensionality of the fused features, making the system more computationally efficient without a substantial loss in accuracy.
Key Results
- The proposed ECG identification system achieved an overall identification accuracy of 95% for 90 subjects.
- The system demonstrated robustness across different ECG databases, with accuracies of 98.00% (MIT-BIH ST change), 95.75% (Long-Term ST), and 96.00% (PTB).
- The fused feature coefficients (MIMFs and their power spectra) provided the best performance (98.1%) compared to using MIMFs alone (95.3%) or power spectra alone (80.9%).
- PCA reduced the feature space from an 8x200 matrix to an 8x20 matrix, improving efficiency.
- The K-NN classifier with k=1 (nearest neighbor) achieved the highest accuracy of 99.78% in optimal parameter selection experiments.
- A 10-second ECG segment was found to be an optimal time interval for identification.
Limitations and Future Work
- ECG Signal Collection Factors: The paper acknowledges that factors like collection position, emotional state, and cardiac irregularities can influence ECG signals and require further investigation.
- Database Expansion: The authors suggest the need for larger ECG databases encompassing a wider range of individuals with varying demographics and conditions.
- Feature Extraction and Classifier Improvement: Further research is needed on more advanced feature extraction techniques and classifier designs to handle more noisy and complex ECG signals.
- Handling Cardiac Irregularities: The current system is validated on typical chest ECGs. Future work should focus on robustly handling ECGs with various cardiac irregularities.
Summary updated