Probabilistic Models for Automated ECG Interval Analysis

This thesis proposes a new approach for the automated analysis of electrocardiogram (ECG) signals, based on the framework of probabilistic modelling. The approach makes use of a number of techniques from machine learning, speech recognition and time-frequency analysis. The problem is to estimate from an ECG signal a number of timing intervals which occur in every heartbeat. The accurate measurement and assessment of these intervals, and in particular the QT interval, is currently the gold standard for evaluating the cardiac safety of new drugs in clinical trials. The approach adopted in this thesis is to train a hidden Markov model (HMM) using a data set of ECG waveforms and the corresponding expert interval measurements. The trained model then serves two complementary purposes. Firstly, it enables the segmentation of test ECG signals through the use of the Viterbi algorithm. Secondly, it serves as a statistical description of ECG waveform normality. This allows the derivation of a confidence measure in the automated ECG interval values inferred by the model. This thesis contributes to the literature on automated ECG interval analysis in three different ways. Firstly, it shows how the undecimated wavelet transform can be used in conjunction with derivative features, to provide a sample-wise encoding of the ECG waveform which is more suited to analysis by a hidden Markov model. Secondly, the problem of double-beat segmentations by an HMM is investigated in detail, and an effective solution proposed based on the use of minimum state duration constraints. Finally, this thesis shows how the probabilistic generative nature of the HMM can be leveraged to provide a confidence measure in the model segmentations derived from the joint log likelihood. Evaluation of the approach on two challenging ECG data sets from clinical drug studies demonstrates the ability of the algorithm to overcome many of the problems associated with existing automated systems, and hence provide a more accurate and robust analysis of ECG signals.