| Summary: | The healthcare system, which is ubiquitously recognized as one of the most influential system in society, is facing new challenges since the start of the decade.The myriad of physiological data generated by individuals, namely in the healthcare system, is generating a burden on physicians, losing effectiveness on the collection of patient data. Information systems and, in particular, novel deep learning (DL) algorithms have been prompting a way to take this problem. This thesis has the aim to have an impact in biosignal research and industry by presenting DL solutions that could empower this field. For this purpose an extensive study of how to incorporate and implement Convolutional Neural Networks (CNN), Recursive Neural Networks (RNN) and Fully Connected Networks in biosignal studies is discussed. Different architecture configurations were explored for signal processing and decision making and were implemented in three different scenarios: (1) Biosignal learning and synthesis; (2) Electrocardiogram (ECG) biometric systems, and; (3) Electrocardiogram (ECG) anomaly detection systems. In (1) a RNN-based architecture was able to replicate autonomously three types of biosignals with a high degree of confidence. As for (2) three CNN-based architectures, and a RNN-based architecture (same used in (1)) were used for both biometric identification, reaching values above 90% for electrode-base datasets (Fantasia, ECG-ID and MIT-BIH) and 75% for off-person dataset (CYBHi), and biometric authentication, achieving Equal Error Rates (EER) of near 0% for Fantasia and MIT-BIH and bellow 4% for CYBHi. As for (3) the abstraction of healthy clean the ECG signal and detection of its deviation was made and tested in two different scenarios: presence of noise using autoencoder and fully-connected network (reaching 99% accuracy for binary classification and 71% for multi-class), and; arrhythmia events by including a RNN to the previous architecture (57% accuracy and 61% sensitivity). In sum, these systems are shown to be capable of producing novel results. The incorporation of several AI systems into one could provide to be the next generation of preventive medicine, as the machines have access to different physiological and anatomical states, it could produce more informed solutions for the issues that one may face in the future increasing the performance of autonomous preventing systems that could be used in every-day life in remote places where the access to medicine is limited. These systems will also help the study of the signal behaviour and how they are made in real life context as explainable AI could trigger this perception and link the inner states of a network with the biological traits.
|
|---|