Artificial Intelligence-empowered bio-medical applications

Panagoulias, Dimitrios P.; Παναγούλιας, Δημήτριος Π.

Βιοϊατρικές εφαρμογές ενισχυμένες με τεχνητή νοημοσύνη

Doctoral Thesis

Author

Panagoulias, Dimitrios P.

Παναγούλιας, Δημήτριος Π.

Date

2024-06

Advisor

Tsihrintzis, George
Τσιχριντζής, Γεώργιος

Keywords

Artificial intelligence ; Large language models ; Bio-informatics ; Algorithms ; Machine learning ; Rational unified process ; Microservices ; AI-explainability

Abstract

This doctoral dissertation examines the advancements in personalized medicine, emphasizing the transition from a generalized "one-size-fits-all" approach to a more tailored strategy in disease diagnosis and patient management. This shift is facilitated by the use of biomarkers, which are critical in developing and training prognostic models and neural networks within the fields of machine learning and artificial intelligence. Biomarkers, defined as measurable and reproducible medical signs, serve as objective indicators of a patient’s medical state. The research highlights significant achievements in using machine learning techniques to predict various health indicator. For example, blood exams are utilized to classify body mass index with an average accuracy of 84%, while cascaded SVM- based classifiers are used to determine systolic blood pressure with a 74% accuracy. Furthermore, a new system has been proposed and implemented to predict metabolic syndrome with an 84% accuracy, relying on multiple blood parameters excluding defining factors of it, thus proposing alternative diagnostic pathways. In addition to specific diagnostic tools, this dissertation explores broader applications of AI-empowered systems that use machine learning models, in healthcare through microservices in distributed systems, which simplify management and scaling by segregating functions into distinct services. In addition the regulatory and validation challenges are also considered and investigated to define optimum development paths and digital architecture solutions using the Rational Unified Process and intuitive workflows. It also addresses the necessity of explainability and interpretability in AI systems to enhance usability and trust, particularly in the medical domain. A novel framework (PINXEL) is introduced to define explainability requirements using the Technology Acceptance Model (TAM). Following that, the capabilities and effectiveness of large language models were evaluated, particularly ChatGPT, in the medical domain. The focus was on this novel system’s ability to aid in medical diagnosis through structured evaluations and domain- specific follow-up analysis. Then a methodology was outlined and implemented to further assess correctness and accuracy of GPT-4V (Generative Pre-trained Transformer 4 with vision) using multimodal multiple-choice questions in the domain of General Pathology. In addition to that using our proposed methodology, requirements can be extracted for targeted interventions and fine-tuning of state-of-the art models, for their optimisation in domain specific tasks. Finally, this doctoral dissertation proposes a novel AI-powered system designed to provide diagnostic advice in primary care settings, employing all technologies explored during this research. This system enhances large language models and other machine learning tools with a rule-based approach, offering tailored medical advice through interactive AI assistant named “Med|Primary AI assistant” and “Dermacen Analytica”. This systems leverage natural language processing, image analysis and segmentation, domain-specific knowledge and multi-modal LLMs to create a personalized user interaction, which is evaluated through a rigorous context and dialogue-based methodological framework. This holistic approach underscores the integration of AI into healthcare, aiming to assist medical professionals in patient diagnosis and management with high accuracy and user-specific adaptability.