Similarity Based Approaches for the Analysis and Prediction of Physiologic Times Series

Abstract

Cardiovascular disease (CVD), a general name for a wide diversity of diseases, disorders and conditions that affect the heart and often the blood vessels, is the largest cause of death in the European Union. Since it is well known that heart health is linked to behaviour and lifestyle, the focus should be on prevention. In the context of preventive medicine, telemonitoring solutions are making a huge impact by enabling remote patient monitoring for the healthy and for those requiring management of chronic diseases. One of the projects that address CVDs management by means of telemonitoring is HeartCycle, a European Integrated Project (FP7-216695) that aims at researching, developing and clinically validating innovative solutions for this purpose. Particularly, the goal of HeartCycle is to improve the quality of care for coronary artery disease (CAD) and heart failure (HF) patients. Integrated in the third workpackage (WP3-Multi-parametric Analysis and Decision Support), the present thesis is centred on the development of specific clinical applications, which target cardiovascular conditions identified as relevant for the CAD/HF management, such as ischemia, arrhythmias and hypertension, based on the analysis and processing of the electrocardiogram (ECG) and blood pressure (BP) signals daily collected by home telemonitoring. Namely, investigation is made on techniques for the diagnosis of the referred conditions, and for the analysis of future trends of these signals enabling the early detection of critical events. Specifically, this thesis presents methodologies for similarity detection and prediction in biosignal time series, which are mainly founded on the representation of signals as linear combinations of a set of orthogonal basis and on the time-frequency analysis of those signals. Particularly, it proposes a new strategy for diagnosing ischemia comprising a measure for evaluating the ST deviation based on the time-frequency analysis of the ECG through the Wigner-Ville transform, and the use of Hermite basis functions to capture the most relevant morphologic characteristics of the QRS complex. This methodology was tested using the European Society of Cardiology ST-T public database, and the relevant results achieved, namely a sensitivity of 96.7% and a positive predictivity of 96.2%, confirmed its potential. Additionally, a new similarity measure based on a combination of the wavelet transform with the Karhunen-Loève transform for temporal patterns detection in biosignal time series, mainly to support prediction methodologies, was developed. The respective validation was performed by quantitatively comparing the proposed measure with other three common measures through the use of data from a public dataset of Physionet (MIMIC-II) and from a private telemonitoring platform (TEN-HMS). The obtained results confirm that the proposed similarity is particularly appropriate to deal with noise, trends and signals that are not perfectly aligned in time. Moreover, an iterative implementation allows for its efficient computational implementation. In terms of predictive strategies two approaches are explored. The first, based on generalized regression neural networks integrated into a multi-model structure is designed for the accurate prediction of time series future values. It was applied in the prediction of acute hypotensive episodes (AHE) and validated in the context of the 2009 Physionet/Computers in Cardiology Challenge using data from MIMIC-II dataset. A correct prediction of 10 out of 10 AHE for test set A and of 37 out of 40 AHE for test set B was achieved, corresponding to the best results of all entries in the two events of the challenge. The main advantage of the second approach is that it does not require the development of a model. It exploits the multi-resolution analysis provided by the wavelet transform to estimate future evolution trend of biosignals, based on the trend evolution of similar historic signals. Its validity was demonstrated by the comparison with other common predictive methodologies. It was employed in the evaluation of the hypertension risk using data from TEN-HMS and MyHeart studies. The obtained results, in terms of Sensitivity-Specificity, were of 84.2%-75.5% and of 85.7%-91.8%, respectively, for the TEN-HMS and the MyHeart datasets, confirming the capability of the approach in this type of application.