Computational Intelligence in Biomedical and Health Care Informatics HCA 590 (Topics in Health Sciences) Rohit Kate Neural Networks: A Sample Medical Application 1 Reading • Chapter 9, Text 5: The Application of Neural Networks in the Classification of the Electrocardiogram 2 Role of Electrocardiogram (ECG) • Heart disease is the largest single cause of premature deaths • If detected, some causes of heart disease can be foreseen and prevented through lifestyle changes • Clinical techniques may evaluate the status of the heart • ECG is one of the most common such a clinical technique 3 ECG • Simple, inexpensive and non-invasive • Records the electrical activity of the heart • Correlates with the fundamental behavior of the heart • Wave shapes describe state of the working muscle masses • Rate of cardiac cycle provides rhythm statements 4 Diagnostic Utilities of ECG • Provides sufficient detail to diagnose a number of cardiac abnormalities including potentially fatal ones, for example, Myocardial Infarction, Left Ventricular Hypertrophy • Not all cardiac abnormalities can be identified by ECG • But in combination with other clinical techniques, for example, angiography, echocardiography, ECG can give a more complete picture of the heart • No standards are currently available for diagnostic classification using ECG 5 12-Lead ECG • Six limb leads measure the cardiac activity in the frontal plane • Six chest leads measure the cardiac activity in the horizontal plane • Together all 12 leads give a three-dimensional picture of the heart • There are 12 electrical signal waves 6 Computerized Classification of the 12-Lead ECG • Input: 12-lead ECG signals • Output: Assign patient to one of the possible diagnostic classes • Computerized classification of ECG is one of the earliest examples of use of computers in medicine • Meant to assist clinicians, not replace them 7 Steps for Computerized Classification • One cannot feed an entire waveform to a classification technique • This is an instance of time-series classification problem • Steps for computerized classification of ECG: – Beat detection – Feature extraction – Possible feature selection – Classification 8 Beat Detection • Automatically locate each cardiac cycle in each of the leads • Insert reference markers for the beginning and end of each inter-wave component • These are used in the feature extraction step 9 Feature Extraction • Generate feature from inter-wave measurements of: – Intervals – Durations – Amplitudes • No standard features have been agreed upon • Decided mainly based on expert medical opinions, medical criteria and some trial and error • Could be potentially hundreds of features from each lead 10 Features and Pathologies • Some ECG deviations from normal indicate certain pathologies, for example: – Q-wave location for specific types of Myocardial Infarction – Large QRS complexes indicate ventricular hypertrophy – R-R intervals tell about heart rate variability • Computers can be more accurate in detecting these and relating them to variuos pathologies 11 Feature Selection • Too many features can confuse a classification method • Too few features may miss some important information • Feature selection is sometimes done to select the most contributing features • Generally done by systematically trying combinations of features and measuring their impact of a validation part of the training set 12 Neural Networks for ECG Classification • Each feature is made an input node • Each of the diagnostic class is made an output node: 1 implies present 0 implies not present • Intermediate numbers may tell the degree of the diagnostic class present • Multiple diagnosis may be obtained • Number of hidden nodes, number of hidden layers, learning rate etc. are determined using the a validation set 13 Training Data • The training data consists of ECG features of several patients and their know diagnostic classes • Important to include wide variety of training examples to ensure generalization of the trained network 14 Results • Several studies have applied neural network techniques for ECG classification • Results vary based on the particular classification classes and features used, vary from 66% to 95% accuracy • Training multiple neural networks and combining their results usually improves the accuracy 15 Neural Networks for ECG Classification • Generally performed competitive with physicians • Neural networks could identify relationships between ECG features that may not have been identified by physicians • Being a statistical method, it is not able to explain final diagnosis – Some methods to extract rules from neural netowrks have been tried 16 Homework 3 Due by 2 pm, next class, Tuesday 10/8 Submit .txt, .doc, .pdf, ppt or a scanned image through D2L Answer each question in 2-3 sentences 1. Both Support Vector Machines and neural networks learn a separator for classification. How do they differ in choosing the separator? 2. How do they differ in their mechanism to go from a linear to a non-linear separator? 17