Electrocardiogram and hybrid support vector algorithms for detection of hypoglycaemia in patients with type 1 diabetes
- Publication Type:
- Issue Date:
Hypoglycaemia is the most acute and common complication of type 1 diabetes. Physiological changes occur when blood glucose concentration falls to a certain level. A number of studies have demonstrated that hypoglycaemia causes electrocardiographic (ECG) alteration. The serious harmful effects of hypoglycaemia on the body motivate research groups to find an optimal strategy to detect it. Detection of hypoglycaemia can be performed by puncturing the skin to measure the blood glucose level. However, this method is unsuitable as frequent puncturing may produce anxiety in patients and periodic puncturing is difficult to conduct, not to mention inconvenient, while the patient is sleeping. Therefore, a continuous and non-invasive technique can be considered for hypoglycaemia detection. Several techniques have been reported, such as reverse iontophoresis and absorption spectroscopy. Another approach to hypoglycaemia detection is based on the physiological effects of hypoglycaemia on the various parts of the body such as the brain, heart and skin. Physiological effects of hypoglycemia to the brain are studied by investigating electroencephalography (EEG) features. Hypoglycemic effects to the heart include alteration of electrocardiographic (ECG) parameters such as heart rate, QT intervals and T-wave amplitude alteration. Several algorithms were developed to process ECG parameters for hypoglycemia detection. The algorithms include neural network and fuzzy system based intelligent algorithms. Furthermore, hybrid systems were also developed, such as fuzzy neural network and genetic-algorithm-based multiple regression with fuzzy inference systems. So far, hypoglycaemia detection systems which are based on the physiological effects still require extensive validation before they can be adopted for worldwide clinical practices. The research in this thesis introduces several ECG parameters especially which relate to the repolarization phase and could contribute to hypoglycaemia detection. Furthermore, this research aims to introduce novel computational intelligent techniques for hypoglycaemia detection. The detection is based on electrocardiographic (ECG) parameters. A support vector machine (SVM) is the first algorithm introduced for hypoglycaemia detection in this research. The second algorithm is a hybrid of SVM with particle swarm optimization (PSO), which is called an SSVM algorithm. This algorithm is intended to improve the performance of the first algorithm. PSO is an evolutionary technique based on the movement of swarms. It is employed to optimize SVM parameters in order that the SVM perform well for hypoglycaemia detection. The third algorithm is for the improvement of the second algorithm where a fuzzy inference system (FIS) is included. This algorithm involves SVM, FIS and a PSO, which is called SFSVM. The FIS is used to process some ECG parameters to find a better performance of hypoglycaemia detection. FIS is an effective intelligent system which employs fuzzy logic and fuzzy set theory. Its frameworks are based on the concepts of fuzzy set theory, fuzzy if-then rules, and fuzzy reasoning. In addition, the proposed algorithms are compared with the other algorithms. All the algorithms are investigated with clinical electrocardiographic data. The data is collected from a hypoglycaemia study of type 1 diabetic patients. This study shows that the selected ECG parameters in hypoglycaemia differ significantly from those in nonhypoglycaemia (p<0.01). This difference might consider that the ECG parameters are part of repolarization, in which repolarization prolongs hypoglycaemia. It implies that the ECG parameters are important parameters which possibly contribute to hypoglycaemia detection. Therefore, the ECG parameters are used for inputs of hypoglycaemia detection in this study. The result also shows that the hypoglycaemia detection strategy which uses SSVM performs better than that which uses SVM (80.04% vs. 73.63%, in terms of geometric mean). Furthermore, the SFSVM performs better than the SSVM (87.22% vs. 80.45% in terms of sensitivity and 79.41% vs. 79.64% in terms of specificity). In summary, SFSVM performs better than the other two algorithms (SVM and SSVM), with acceptable sensitivity, specificity and geometric mean of 87.22%, 79.41% and 83.22%, respectively.
Please use this identifier to cite or link to this item: