D'Errico, Lorenzo (2026) Artificial Intelligence for Human Assistance and Understanding: Affective, Cognitive, and Clinical Pathways Toward Human-Centered AI. [Tesi di dottorato]

Documento PDF Thesis-Lorenzo-DErrico-reviewed.pdf Visibile a [TBR] Amministratori dell'archivio Download (24MB) | Richiedi una copia

Abstract

Artificial Intelligence systems are increasingly deployed in domains where success depends not on computational performance alone, but on the ability to support human decision-making under uncertainty, data scarcity, and ethical constraints. In such contexts, AI systems are required to operate as assistive components rather than autonomous decision-makers, adapting to human needs while remaining interpretable, robust, and accountable. This thesis investigates how human-centered Artificial Intelligence principles can be operationalized across two distinct application domains—socially assistive robotics and biomedical image analysis—treated as parallel case studies of assistive AI systems operating under different interaction and data constraints. Rather than proposing a single unified architecture, the work examines how recurring methodological challenges emerge when AI is designed to augment human perception, reasoning, and action. Part I focuses on interactive assistance through human–robot interaction. A multimodal affect recognition framework combining EEG and audio–visual signals is developed, achieving up to 99.2% accuracy in emotion classification and 91.4% in multimodal fusion. These perceptual capabilities are embedded within adaptive interaction loops and validated through emotivational studies involving human participants (n=20), as well as through a clinical-oriented application for mental health screening (RoboTald). Emphasis is placed on interpretability, controllability, and the consequences of classification errors when AI outputs directly influence robot behavior. Part II addresses diagnostic assistance in biomedical imaging, where supervised learning is constrained by class imbalance, limited annotations, and acquisition costs. Generative learning approaches are explored for dataset augmentation in Digital Breast Tomosynthesis, volumetric synthesis in Magnetic Resonance Imaging, and morphology-aware virtual staining in digital pathology. These case studies demonstrate how task-driven representation learning can enhance downstream diagnostic pipelines when data availability is limited. Across both domains, the thesis advances a methodological perspective in which human-centered AI systems are evaluated not solely by predictive accuracy or visual realism, but by task relevance, robustness, and alignment with human workflows. By analyzing how similar design principles manifest in interactive and analytical settings, this work contributes domain-specific solutions and a comparative framework for the development of assistive AI systems in which human expertise remains central.

Downloads

Actions (login required)

Modifica documento