During the last decade an unprecedented development of healthcare technologies has taken place. Advances in computerization, networking, imaging, robotics, micro/nano technologies, and bioinformatics, have contributed to significantly increasing the amount and diversity of information available to physicians for medical diagnosis, prognosis, treatment and monitoring. The increase in the amount and diversity of clinical data requires the continuous development of new techniques and procedures capable of integrating and analysing the complexity of the whole information generated as well as providing support in their interpretation.
In this context, this thesis has been focused on the analysis and processing of biomedical signals and their use in automatic classification problems. That is, the design and implementation of automatic biomedical signal pre- processing algorithms, the assessment of compatibility between biomedical signals, the design of novel feature extraction methods for biomedical signals, and the design of classification models for specific clinical problems. In most of the cases presented, these methods will be applied in the framework of clinical decision systems; computational systems for the decision support in the diagnosis, prognosis and treatment of patients.
One of the main contributions of this thesis consists of the evaluation of the compatibility between magnetic resonance spectra (MRS) obtained by two coexisting magnetic resonance scanner technologies (1.5T and 3T) in the context of brain tumour automatic classification. The results obtained in this work suggest that existing classifiers based on 1.5T datasets may be applicable to 3T 1H SV-MRS dataset. The compatibility between the now coexisting clinical MR scanners technologies of 1.5T and 3T plays a decisive role in the development of new classifiers for tumour diagnosis support and for the use of the existing ones based on 1.5T. Furthermore, this will make it possible to take more benefit from the effort made throughout years in the compilation of large MRS databases.
Another contribution of this thesis is the design and implementation of a new feature extraction method, RSFDA, for the classification of brain tumour MRS spectra. RSFDA is based on a functional description or MRS data by a non-uniformly distributed B-spline basis. The evaluation of RSFDA using simulated datasets and two large real muticenter datasets, shows that RSFDA had exhibited robustness against the most common artefacts in MRS. The application of RSFDA constitutes an interesting alternative to other FE methods, especially when using datasets with high baseline deformations. Furthermore, this thesis provides a new method for coincidence classification in positron emission tomography (PET) scanners. The presented method uses an artificial neural network (ANN) trained with simulated datasets, for the identification of true coincidences. This algorithm improved the performance of traditional coincidence sorting algorithm based on a time coincidence window and the application of geometrical conditions. Comparing both methods at the same efficiency values, the ANN based method always presents a sorted output with a smaller random fraction. Moreover the comparison of reconstructed images using the different methods for coincidence classification shows an improvement of all figures of merit defined in the study.
Finally, this thesis presents a multicenter study for the assessment of different feature extraction methods, classification algorithms, and aligning methods for arrhythmia automatic detection and classification by electrocardiography (ECG) signals. In this case it has been achieved to compile a relatively large dataset consisting on annotated 12-lead ECG records. This study has been focused on the classification of ventricular arrhythmia versus supra ventricular and control beats. The results show an improvement of the classification performance when using the algorithm for the beat alignment. Moreover, it is remarkable that the simplest classifiers based on k-nearest neighbor algorithm obtain the best performances.
The results of this PhD Thesis apply directly to the results of the following projects: eTumour (VI Framework Program, LSHC-CT-2004- 503094, 2004-2009), HealthAgents (VI Framework Program, IST- 2004- 27214, 2006-2009), DSSPROSTATA (PROFIT FIT- 340001-2007-14, 2007- 2009), DSS PROSTATA2 (AVANZA, TSI-020302-2009-65, 2009-2010), and Sistema Experto en Electrocardiograf’a (AVANZA I+D, TSI-020302-2010- 136, 2010-2011). Based on the scientific contributions studied in this thesis, different practical solutions have been developed including the design and implementation of: (1) semi-automatic pre-processing pipeline for 1H-MRS signals in CURIAM-BT (a clinical decision support system for brain tumours diagnosis by MRS); (2) ArrythLab, a software for automatic pre-processing, quality control, segmentation and beat labeling in arrhythmia studies by ECG; and (3) MRSILab, an open source MATLAB toolbox for automatic pre- processing, analysis and visualization of magnetic resonance spectroscopy imaging (MRSI) data.