This thesis focuses on two main principal topics, regarding respectively the pre-processing performance analysis and the ECG classification process. The pre-processing performance analysis was achieved with the combined use of two indexes including the SNR percentual increment and the diagnostic distortion measure (DDM) that were used to optimize filters parameters for respectively maximizing denoising effect and minimizing filtering signal distortion. The ECG classification process includes a new algorithm for AF detection from ultra-short (10 seconds) single lead ECG records. The AF detection algorithm is composed by two successive classification stages. Firstly, HRV signal is extracted from ECG record and it is then decomposed in 5 beats ROI from which a set of HRV features are extracted and used in the first ROI classification stage through MLP NN. Then, the sequence of classified ROI extracted from each ECG record is transformed into a grey levels image where each ROI corresponds to a pixel. A set of features are extracted from grey levels image and are used in the second image classification stage through MLP NN. AF detection algorithm was validated with 5-fold cross-validation technique and average performances show a sensibility, specificity, and accuracy of 92.62%, 91.44% and 92.21% respectively.
AUTHOR: Giovanni Di Martino
ADVISORS: Eros Gian Alessandro Pasero, Jacopo Ferretti
DEGREE COURSE: Master’s Degree in Biomedical Engineering
ACADEMIC YEAR: 2019/2020