An Adaptive Fuzzy ECG Classifier Presented by: Lei W ai Kei Dept. - PowerPoint PPT Presentation

An Adaptive Fuzzy ECG Classifier Presented by: Lei W ai Kei Dept. Electronic & Electrical Eng, Faculty of Science & Technology, University of Macau

Outlines � Introduction � ECG Signal and Patterns � A Fuzzy ECG Classifier � An Adaptive Fuzzy ECG Classifier � Demonstrations � Conclusions

Introduction � ECG signals reflect the health status of heart by completely recording the subtle cardiovascular circulation . � Automatic ECG analysis is essential to implement ECG monitoring in Home Healthcare . � Here an Adaptive Fuzzy ECG Classifier ( AFC - ECG ) is proposed to integrate the knowledge from medical expertise and by statistical analysis as well.

ECG Cardiology Left Atrium � P: Contraction of Atrium � QRS: Contraction of Ventricle � T: Relaxation of Heart Right Atrium R T P Right Ventricle Q Left Ventricle S ECG signal Heart

Characteristic ECG Features � Characteristic Features: ECG Features ECG Features R P Peak Value T R Peak Value P T Peak Value Prior Heart Rate PR Segment ST Segment Q Post Heart Rate S PR Segment QRS Interval

ECG Classification � ECG Classifier Structure Medical Knowledge Inference Result ECG ECG Intelligent Classification of Signal Parameterization Inference ECG Beats Machine Extracting ECG Features

Methods for ECG Classifiers � Fuzzy Inference Networks (FIN) Membership Function Performs imprecise and interpretable inference. Input Result But without self-adaptation mechanism. Rule Set � Neural Networks (NN) Capable of self-adaptation inference but blind to medical experts. Weights Input Result � Adaptive FIN Not only performs imprecise , interpretable inference like FIN , but also improves the classifier performance by statistics-based self-adaptation .

FIN ECG Classifiers � Fuzzification ECG Features Membership Grading � Classification Knowledge Rule-based fuzzy Inference Fuzzification Base � Results Over 10 kinds of heart beats are defined, 4 of them are chosen for Classification testing: (1) Normal Beat ( N ) (2) Premature Atrial Contraction ( PAC ) (3) Left Bundle Block Beat ( LBBB ) Result (4) Right Bundle Block Beat ( RBBB )

Fuzzification � Calculating the Membership Grades of the input ECG Features. � Obtaining the Linguistic Variables ECG Features Fuzzification Linguistic Variable P Peak Value Disappear, Early… R Peak Value Upward, Downward… T Peak Value Upward, Downward… RR0 Short, Normal… RR1 Short, Normal… PR Segment Reference Line

Applied Membership Functions for Fuzzification � S -Function = ≤ ⎧ RI 0 x a ⎪ ( ) ( ) ( ) = ⋅ − − < ≤ + 2 2 ⎪ RI 2 x a b a a x a b / 2 ⎨ ( ) ( ) ( ) = − ⋅ − − + < ≤ 2 2 RI 1 2 x a b a a b / 2 x b ⎪ ⎪ = ≤ ⎩ RI 1 b x � Z -Function Similar to S-Function Type Type a b � G aussian Function S, Z Lower Upper ( ) = − − Boundary Boundary 2 2 RI exp[ x a 2 b ] G Mean Standard ( RI: means Relationship Index ) Value Deviation

Illustration of Fuzzification � Membership Functions of “Normal Beat” Linguistic ECG Function Parameter a Parameter b Variable Feature Type P upward P Peak S 0.10mV 0.15mV Value QRS upward R Peak S 0.70mV 0.80mV Value T upward T Peak Value S 0.10mV 0.15mV RR0 normal Prior-HR Gaussian 80bpm 20bpm RR1 normal Post-HR Gaussian 80bpm 20bpm

Rule Template of Fuzzy Classification IF “Feature 1 Feature 1” ” is “Linguistic Variable 1” & “ “Feature 2 Feature 2” ” is “Linguistic Variable 2” & “ … “Feature N Feature N” ” is “Linguistic Variable N” “ THEN “ “Beat Type Beat Type” ” is “Class Name” Beat Type No. of Rules No. of Hypotheses Normal 1 5 RBBB 2 8 LBBB 2 8 PAC 2 8

Illustration of Fuzzy Classification � Rule for Normal Beat ( N ) IF “P is Upward ” & “QRS is Upward ” & “T is Downward ” & “RR0 is Normal ” & “RR1 is Normal ” THEN Type is “Normal”

Problems in FIN ECG Classifiers � Variations in ECG Signals An illustration of variation occurs in “Normal Beat”

Problems in FIN ECG Classifiers � Hard to define membership functions Linguistic ECG Function Parameter a Parameter b Variable Feature Type P upward P Peak S 0.10mV 0.15mV Value QRS upward R Peak S 0.70mV 0.80mV Value T upward T Peak Value S 0.10mV 0.15mV RR0 normal Prior-HR Gaussian 80bpm 20bpm RR1 normal Post-HR Gaussian 80bpm 20bpm

Adaptive Fuzzy ECG Classifier Features: Features: � Population Estimation of incoming signals: Obtaining the statistical parameters of incoming signals, e.g., mean values and standard deviations � Modifying membership functions : Base on above statistical parameters , adapting membership functions to current ECG Signals . current ECG Signals Rationales: Rationales: � Physiological: Incoming ECG signal rarely numerical match with the medical knowledge. The capacity of adaptation is necessary to deal with the variance which occurs in input. variance � Mathematical: Statistics is pertaining to the analysis , interpretation and presentation of data. It provides a way to draw inferences draw inferences about the population of incoming signal.

Statistical analysis for Population Estimation Two parameters are used to estimate the population: � Mean Value Mean Value � ⎛ ⎞ N ∑ μ = ⎜ ⎟ / x N i ⎝ ⎠ = i 1 � Standard Deviation (SD) Standard Deviation (SD) � SD Samples σ 68 . 27 % ⎛ ⎞ N ( ) ∑ σ σ = − μ ⎜ ⎟ 2 2 95 . 45 % / x N i ⎝ ⎠ σ 3 99 . 73 % = i 1 σ 4 99 . 99 %

Self-adaptation of Membership Function According to statistic analysis of ECG signals, about 95% of the values are within 2 standard deviation . � S , Z -function = μ − σ μ + σ 2 a 2 = μ + σ b 2 � G aussian μ − σ = μ 2 a = σ b 2

AFC-ECG Structure � Enhance the conventional FIN ECG classifiers by statistical learning methods . Adaptation Yes ECG Knowledge Fuzzification Adjust? Results Features Base No Classification

Adaptive Process The first 10 minutes of ECG signals are taken advantage for self-adaptation : � Population calculation by Statistical Analysis Modifying the parameters of membership � function μ , σ Modification by ( ) Learning set Normal Adaptation LBBB RBBB Yes PAC ECG Knowledge Fuzzification Adjust? Features Base

Classification Process � The remaining 20 minutes of ECG signals are used for further classification. Beat Type Normal LBBB Test Beats RBBB PAC ECG Knowledge Fuzzification Adjust? Result Features Base No Classification

Experiment Results Record Type Annotation Before Learning After Learning Normal 106 1507 1208 80.2% 1258 83.5% PAC 0 7 - 1 - RBBB 0 30 - 27 - LBBB 0 0 - 1 - Uncertain 0 262 - 220 - Normal 118 0 1 - 0 - PAC 96 49 51.0% 72 75.0% RBBB 2166 1602 74.0% 1831 84.5% LBBB 0 0 - 0 - Uncertain 0 610 - 359 - 207 Normal 0 19 - 2 - PAC 107 127 81.3% 105 98.1% RBBB 86 71 82.6% 80 93.0% LBBB 1457 1357 93.1% 1389 95.3% Uncertain 0 76 - 74 -

Experiment Analysis & Discussion � Accuracy Accuracy: � � The average accuracy of FIN ECG classifier : 77.0% � The average accuracy of AFC-ECG : 88.2% � Worst case is in the Record “118” because there are no enough learning samples. Even after self-adaptation, its accuracy is still low to 75.0% .

Significances � The proposed adaptive fuzzy ECG classifier integrates the advantages of Fuzzy Logics in human knowledge expression and statistical learning method for self-adaptation . � Advantages: � Eliminating the effects of physiological signal variation; � Refining the parameters of fuzzy rule by self-adaptation; � Disadvantages: � Dependent on initial fuzzy rules; � Invalid to unknown ECG signals. � Future work � Increasing the types of classifiable ECG beats . � Extending to Fuzzy Neural Networks