Abstract
dc:descriptionThe reduced quality and the increased structural and conceptual heterogeneity of the clinical databases combined with the presence of data silos obscure the sharing and analysis of medical data. These open issues in healthcare leverage the development and secure deployment of robust and unbiased AI (Artificial Intelligence) workflows to address clinical unmet needs, including: (i) the development of robust disease classification and risk stratification models, (ii) the detection of new biomarkers, and (iii) the discovery of targeted therapies, among others. In this thesis, we aim to address the open issues and unmet needs in healthcare through the development of beyond the state of the art methods which are built on top of four main innovation areas: (i) Innovation Area 1 - data curation, where we propose a fully automated, efficient and scalable medical data curation workflow to enhance the quality of the diverse medical data including clinical and genetic data across multiple time-points, (ii) Innovation Area 2 - data harmonization, where we propose a hybrid, fully automated data harmonization workflow combining lexical and semantic analysis based on word embeddings which is built on top of external knowledge bases to overcome structural heterogeneities across clinical databases, (iii) Innovation Area 3 - synthetic data generation, where we propose a large-scale synthetic data generator to significantly enhance the statistical power of clinical databases with insufficient population size in order to enable the simulation of clinical trials, as well as, to enhance the classification performance of the existing AI models through data augmentation, and (iv) Innovation Area 4 – federated/distributed learning, where we propose a federated AI deployment framework which removes the need for the installation of local servers or any type of software in each site through the adoption of a federated AI modeling engine supporting a large family of federated AI algorithms yielding interpretable and explainable AI models. The proposed four stage workflow was evaluated across six different clinical domains, including autoimmune diseases (AD) and particularly in primary Sjogren’s Syndrome (pSS), hypertrophic cardiomyopathy (HCM), cardiovascular diseases (CVD), mental disorders (MD), systemic autoinflammatory diseases (SAIDs), and particularly Kawasaki disease (KD), and Coronavirus disease (COVID-19). The applicability of the proposed workflow was successfully demonstrated by: (i) enhancing the quality of the clinical and laboratory data in pSS, HCM, COVID-19, CVD, MD, KD, (ii) reducing the levels of structural and conceptual heterogeneity among the clinical and laboratory data in pSS, CVD, MD and at the same time enabling the evaluation of cross-domain data harmonization, (iii) producing high quality and large scale synthetic data for in silico clinical trials in HCM, (iv) augmenting the existing lymphoma classification models in pSS and HCM risk stratification models, and (v) producing robust AI models for lymphoma classification in pSS, the detection of biomarkers for lymphomagenesis, the detection of biomarkers for Kawasaki disease, HCM risk stratification, ICU admission and mortality classification in COVID-19.
Degree
thesis:*- Grantor dc:publisher
- University of Ioannina
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Pezoulas, Vasileios
- Πεζούλας, Βασίλειος
Subjects
dc:subject × 20- Τεχνητή νοημοσύνη
- Μηχανική μάθηση
- Εξυγίανση δεδομένων
- Εναρμόνιση δεδομένων
- Δημιουργία συνθετικών δεδομένων
- Εκμάθηση εντός του νέφους
- Βιοϊατρική μηχανική
- Artificial intelligence
- Machine learning
- Data curation
- Data harmonization
- Synthetic data generation
- Federated Learning
- Biomedical engineering
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/52264