|
|
Machine Learning-Aided Monitoring and Prediction of Respiratory and Neurodegenerative Diseases Using Wearables
Skibińska, Justyna ; Esposito, Anna (oponent) ; Faundez-Zanuy, Marcos (oponent) ; Hošek, Jiří (vedoucí práce)
This thesis focuses on wearables for health status monitoring, covering applications aimed at emergency solutions to the COVID-19 pandemic and aging society. The methods of ambient assisted living (AAL) are presented for the neurodegenerative disease Parkinson's disease (PD), facilitating 'aging in place' thanks to machine learning and around wearables - solutions of mHealth. Furthermore, the approaches using machine learning and wearables are discussed for early-stage COVID-19 detection. Firstly, a publicly available dataset containing COVID-19, influenza, and healthy control data was reused for research purposes. The solution presented in this thesis is considering the classification problem and outperformed the state-of-the-art methods, whereas the original paper introduced just anomaly detection. The proposed model in the thesis for early detection of COVID-19 achieved 78 % for the k-NN classifier. Moreover, a second dataset available on request was utilized for recognition between COVID-19 cases and two types of influenza. The scrutinisation in the form of the classification between the COVID-19 and Influensa groups is proposed as the extension to the research presented in the original paper. The accuracy of the distinction between COVID-19 cases and influenza in the middle of the pandemic was equal to 73 % thanks to the k-NN. Furthermore, the contribution as the classification model of two aforementioned combined datasets was provided, and COVID-19 cases were able to be distinguished from healthy controls with 73 % accuracy thanks to XGBoost algorithm. The undeniable advantage of the illustrated approaches is taking into consideration the incubation period and contagiousness of the disease. In addition, some solutions for the detection of the aforementioned aging society phenomenon are presented. This study explores the possibility of fusing computerised analysis of hypomimia and hypokinetic dysarthria for the spectrum of Czech speech exercises. The introduced dataset is unique in this field because of its diversity and myriad of speech exercises. The aim is to introduce a new techniques of PD diagnosis that could be easily integrated into mHealth systems. A classifier based on XGBoost was used, and SHAP values were used to ensure interpretability. The presented interpretability allows for the identification of clinically valuable biomarkers. Moreover, the fusion of video and audio modalities increased the balanced accuracy to 83 %. This methodology pointed out the most indicative speech exercise – tongue twister from the clinical point of view. Furthermore, this work belongs to just a few studies which tackle the subject of utilising multimodality for PD and this approach was profitable in contrast with a single modality. Another study, presented in this thesis, investigated the possibility of detecting Parkinson's disease by observing changes in emotion expression during difficult-to-pronounce speech exercises. The obtained model with XGBoost achieved 69 % accuracy for a tongue twister. The usage of facial features, emotion recognition, and computational analysis of tongue twister was proved to be successful in PD detection, which is the key novelty and contribution of this study. Additionally, the unique overview of potential methodologies suitable for the detection of PD based on sleep disorders was depicted.
|
host ::
RSS.