Metabolic Syndrome Detection Based on Classification of Electrocardiography Signals

Metabolic syndrome (MS) components, mainly correlated with insulin resistance and diabetes, constitute physiological disturbances that are objectively detectable based on physiological and anatomical measurements. In particular, the scientific literature indicates clear associations between features extracted from electrocardiograph (ECG) signals and MS. However, there exist few scientific studies related to MS detection by means of ECG signals, specially in automatic computer aided systems. This paper aims at developing and evaluating automatic tools for possible MS detection based on ECG signals. To evaluate how accurately and precisely the developed classifier systems detect MS from ECG signals, we use the following procedures. Initially, we use algorithms that automatically extract Q, R, and S peaks from ECG waveforms. Subsequently, we extract temporal features mainly associated with averages and variances of intervals and ratios between successive Q, R, and S peaks. We also use features describing the cardiac axis. The features are then used for training and testing classifier systems, including Support Vector Machines (SVMs) and RobustBoost classifiers. We also test the use of classifiers operating on raw ECG signals, without preliminary explicit feature extraction. The tested models constitute different configurations of Convolutional Neural Networks (CNNs). Our results indicate that it is possible to classify ECG signals in two different classes, separating people with MS from a control group, with statistically significant results. SVM, RobustBoost, and CNN models obtained average accuracy values equal to 94%, 89%, and 98%, respectively. These results indicate that automatic computer-aided diagnositcs of MS can be added to standard ECG clinical exams.