This PhD thesis presents some contributions to Statistical Signal Processing with applications in biomedical engineering. The thesis is separated into two parts. In the first part, the detection of protein interface hotspots ¿ the residues that play the most important role in protein interaction ¿ is considered in the Machine Learning framework. The Random Forests is used as the classifier. A new family of protein hotspot descriptors is also introduced. These descriptors are based exclusively on the primary one-dimensional amino acid sequence. No information on the three dimensional structure of the protein or the complex is required. These descriptors, capturing the protein frequency characteristics, make it possible to get an insight into how the protein primary sequence can determine its higher structure and its function. In the second part, the RDT (Random Distortion Testing) robust hypothesis testing is considered. Its application to signal detection is shown to be resilient to model mismatch. We propose an extension of RDT in the sequential decision framework, namely Sequential RDT. Three classical signal deviation/distortion detection problems are reformulated and cast into the RDT framework. Using RDT and Sequential RDT, we investigate the detection of AutoPEEP (auto-Positive End Expiratory Pressure), a common ventilatory abnormality during mechanical ventilation. This is the first work of that kind in the state-of-the-art. Extension to the detection of other types of asynchrony is also studied and discussed. These early detectors of AutoPEEP and asynchrony are key elements of an automatic and continuous patient-ventilator interface monitoring framework.
