ECG Analysis for the Human Identification By Tsu-Wang Shen Department of Biomedical Engineering University of Wisconsin - Madison Problem Description By using the neural network technologies, my goal is tried to discover the essential features from the only “one-lead” resting ECG signals to identify human. Once the first goal is achieved, to minimize the number of features in order to apply in real world applications. Project Outline Goal: looking for if ECG analysis is a secure, fast, easily applied, and low-cost method to identify people Build an ECG database. Pre-process ECG and feature extraction Design a system to identify people by using only one-lead ECG. Use the database to train the ANN system. After the training is done, the system is tested for the correct classified rate. People have their own identical heart beat System Diagram Feature extraction ECG database Template match Candidates ECG signals from sensors Pre-process (LP/HP Filtering, and normal beat selection) Pre-screen Identification Decision based neural network (DBNN) Pre-process Remove the interference: (ECG signal frequency range: 0.01-250 Hz) Baseline wander filter Power line interference cancellation Highpass filter Detect Normal beats In this project, the beats is judged by physicians (MIT/BIH database). Template match results Candidates ECG feature Extraction The problem of feature extraction The feature extraction plays a key role of this project. Normal ECG vs. Abnormal ECG A person’s ECG signal may not have all the components, such as P wave and T wave. The selected features should be less correlation between each other. That makes the features have less redundant information. Heart beats change slightly all the time, so it is very hard to set observation points. Decision Based Neural Network Result of Recognition MAXNET 1 w11 x1 2 w21 wn1 w12 x2 xn x1 L w22 wn2 x2 xn w13 x1 w23 wn3 x2 xn x1, x2, … , and xn are features of ECG signals. DBNN Structure Train the system in advance. This is a supervised neural network. Reinforced learning is applied for the correct class neuron. Anti-reinforced learning is applied for the misclassified neurons. Pick the maximum value from all the class outputs as the final result. Conclusion It is possible to identify people by use only one-lead ECG. Pre-processing and pre-screening are important to limit the possible candidates. In this project, all ECG signals are in the ideal condition. (Normal ECG signals, Noise removed totally.) Need more ECG database in the future.