Contributions to Biomedical Signal Analysis Using Nonlinear Dynamics and Machine Learning

Machine Learning and Signal Processing have myriad applications in healthcare from automating the administrative tasks to speeding up medical diagnosis and drug discovery. This domain brings computer and medical sciences in a single thread, which can pull together a massive amount of techniques to improve the efficiency and cost of healthcare in the sector. In this interdisciplinary field, Machine Learning solutions are needed to be designed explicitly for each task. Thus, the developed solutions in this area must be more prudent about compliance compared to other domains. It is crucial to know how the clinical data are collected or for what purposes the data are gathered to deliver reliable solutions. This study focuses on two primary objectives. The first objective is to propose a novel set of nonlinear descriptors for biosignals using nonlinear dynamics. The second objective is to develop feature extraction and classification schemes to help medical experts improve the diagnosis of prevalent medical conditions. The first contribution of the thesis is to introduce a novel set of nonlinear descriptors for capturing the dynamic signature of time series. The proposed features are inspired by the Poincaré section and nullclines concepts in nonlinear dynamics. The discriminative power of the proposed features is evaluated in the epileptic seizure detection task. For this purpose, a two-layer classification approach is developed to detect seizure events using multichannel electroencephalogram signals. The extensive comparative analysis shows the superiority of the proposed framework over other state-of-the-art methods, although further research is needed for reducing the high false alarm rate. The second contribution is to develop three novel feature extraction and classification schemes to tackle clinically relevant conditions. These medical applications encompass the identification of heart anomalies using heart sounds, automatic classification of atrial fibrillation using hand-held electrocardiogram devices, and sepsis prediction from 6 to 12 hours before clinical recognition. The experimental results show the feasibility of these assistive diagnostic tools. The developed solutions are assessed against a variety of solutions proposed by several teams from academia and industry and ranked among the top three methods. The proposed feature extraction and classification techniques are stand-alone contributions to the field. The proposed nonlinear dynamics techniques can be used in a variety of domains due to their generic nature. Moreover, the developed automatic classification and predictive models in this thesis showed significant improvement in the identification of challenging clinical events. According to the conducted evaluation analysis, the proposed methods exhibit relatively robust performance over the collected data across different institutions. We conclude that Machine Learning assistive diagnostic models designed with sufficient care can provide viable help for expert clinicians