NATIONAL PHD IN ARTIFICIAL INTELLIGENCE

Thesis title: The long and winding path towards a trustworthy AI

In recent years, Artificial Intelligence (AI) has contributed to profound transformations across various sectors, due to its ability to analyze large datasets, detect intricate patterns, and support complex decision-making processes. However, its integration into critical fields and applications brings significant technical and ethical challenges, notably regarding transparency, interpretability, and fairness. This thesis addresses these challenges through four complementary investigations, aiming to foster responsible and ethical AI applications. The first study explores the application of machine learning models enhanced with explainable AI (XAI) techniques to predict the efficacy of immunotherapy in non-small cell lung cancer (NSCLC) patients. By integrating clinical and imaging data, the proposed framework not only forecasts treatment outcomes but also provides interpretable insights to support clinical decision-making. The second investigation examines how demographic and socioeconomic factors are encoded in medical imaging data. Using vector embeddings -- numerical representations that capture complex and nuanced patterns within data -- derived from chest radiographs, this work quantifies biases that could undermine fairness in clinical applications and discusses strategies to mitigate these effects and promote more equitable AI models. The third area addresses the challenge of Continual Learning (CL), proposing methods that allow AI systems to assimilate new information (e.g., from evolving data) without forgetting previously acquired knowledge. This is particularly critical in rapidly evolving healthcare contexts, where sustaining both accuracy and fairness over time is essential. Eventually, the thesis explores neuro-symbolic approachs for combining the pattern-recognition capabilities of neural networks with the logical rigor of symbolic reasoning. Applied to compliance verification in manufacturing, specifically in electrical control panels, this hybrid framework is assessed over effectiveness and potential advancements.