Pushing the boundaries of healthcare: AI-IoT solutions for a secure precise and explainable healthcare system

The integration of Artificial Intelligence (AI) with the Internet of Things (IoT) has revolutionized numerous domains, and healthcare stands as a critical beneficiary of these advancements. The promise of improved diagnostic precision, personalized treatment, and enhanced operational efficiency is shaping the future of Healthcare 5.0. However, this transformation is not without challenges. Complex computational demands, the need for explainable and interpretable models, and stringent security and privacy requirements hinder the seamless adoption of AI-IoT solutions in clinical and real-world scenarios. These issues provided the impetus for this research, which aims to develop robust methodologies addressing these pressing challenges. Edge devices, a cornerstone of IoT ecosystems, are constrained by limited computational resources, energy efficiency, and memory capacity. Yet, they must perform complex inference tasks in real time to enable actionable insights in critical scenarios like continuous glucose monitoring or arrhythmia detection. Current AI solutions often rely on cloud computing, introducing latency and risking data breaches. The need to optimize AI models for resource-constrained environments drives this research's focus on model compression techniques such as pruning and quantization, ensuring that cutting-edge Neural Networks (NNs) can operate effectively within the constraints of edge devices. While AI systems have achieved remarkable performance in various applications, their black-box nature poses significant challenges for adoption in healthcare. Clinicians and stakeholders require transparent and interpretable solutions to trust AI-driven recommendations, especially in life-critical situations. This research emphasizes the development of explainable AI techniques, such as feature-based knowledge distillation to embed transparency into predictive models. These efforts aim to enhance user trust and support clinical decision-making by offering clear insights into the reasoning behind AI predictions. Healthcare data are inherently sensitive, and ensuring their security is paramount. Conventional centralized learning approaches require data to be collected to train predictive models, often compromising patient confidentiality and trust. This research prioritizes privacy-preserving methodologies by integrating Federated Learning frameworks with cryptographic enhancements such as Zero-Knowledge Proofs and Distributed Ledger Technology. These innovations enable secure, scalable AI deployments that uphold data privacy without compromising performance, particularly in decentralized IoT environments. The motivation for this thesis is rooted in the potential of AI-IoT solutions to transform healthcare delivery. From enabling real-time monitoring and diagnosis to facilitating predictive and preventive healthcare, the applications of this technology are vast. This work addresses the interdisciplinary challenges required to bridge the gap between theoretical advancements and practical deployments, ensuring that AI-IoT systems are not only effective but also trustworthy, interpretable, and secure. In summary, this research is driven by the pressing need to overcome the technical, ethical, and operational barriers to AI-IoT integration in healthcare. By focusing on computational efficiency, interpretability, and security, this work contributes to the realization of intelligent, impactful, and sustainable solutions for Healthcare 5.0.