Analysis and interpretation of electrocardiogram signals for the detection of hypoglycaemia.

Diabetes is a complication of metabolism where the glucose control system of the human body is impaired and cannot preserve the blood glucose levels in the normal range. This research investigated the relationship between abnormally low glucose levels (hypoglycaemia) and cardiac function in human subjects with Type 1 diabetes. The aim of the research was to detect the onset of spontaneous nocturnal hypoglycaemia indirectly through analysis of the subject's Electrocardiogram (ECG). The research hypothesis follows from previous studies, that suggested changes in ECG morphology, in particular prolongation of the QT interval and flattening of the T-wave, during hypoglycaemia.The research methodology involved ECG feature extraction and classification of extracted features into euglycaemic (normal glucose levels) and hypoglycaemic categories. A number of time-domain ECG features were evaluated and a few ECG annotation algorithms were investigated for detection of onsets, peaks and offsets of the ECG components. Autoregressive (AR) modelling was also employed as a means of describing and characterising post-QRS ECG segments. ECG segment classification was carried out using Multi-layer Perceptron (MLP) neural networks. Statistical classifiers were also employed namely, Linear Discriminant Analysis (LDA) and the k-Nearest Neighbour (kNN).This research proposed a new methodology for detection of spontaneous nocturnal hypoglycaemia by combining time-domain characterisation and classification of the post-QRS ECG segment. Two novel ECG features were introduced to characterise T-wave morphology. MLPs achieved better classification of ECG feature vectors compared to LDA. Also ECG representation by AR coefficients was marginally superior to individual ECG features, according to classification performance by LDA. Finally a Knowledge-Based System (KBS) was designed for ECG monitoring during the night. It was developed and tested onoffline data in a m anner that simulated an online monitoring scenario. The system was able to detect ECG abnormalities related to spontaneous nocturnal hypoglycaemia and to raise an alarm if necessary. In its optimal configuration, the system correctly monitored 30 out of the 32 recorded nights (originating from 19 patients) while there were 2 false alarms. This performance corresponds to accuracy, sensitivity and specificity of 93.75%, 100% and 91.30% respectively.The main contribution to knowledge from this research was successful detection of the onset of spontaneous nocturnal hypoglycaemia indirectly, using solely ECG information. This result supports the hypothesis stating that spontaneous hypoglycaemia affects the cardiac function and is manifested on the ECG. A. detailed analysis of the ECG signal for the detection of hypoglycaemia was carried out in the thesis. ECG features were extracted and assessed as predictors of the clinical condition. A number of approaches for ECG representation and classification (MLP, kNN, LDA) were examined and compared. Moreover, a KBS capable of achieving satisfactory monitoring performance on offline data from diabetic patients was designed. It was found that ECG changes in response to hypoglycaemia were short-time transients and incorporation of temporal information in the classification system caused significant improvement in performance. Successful continuation of this work may lead to a hypoglycaemia-detection system for the bedside.