AI-Powered Biomedical Sensing and Privacy-Enhancing Technologies: From Smart Insulin Pen Systems to Federated and Encrypted Frameworks

Artificial Intelligence (AI) is transforming the biomedical domain by enabling intelligent sensing, data sharing, and decision support across connected environments. To truly integrate AI into healthcare, systems must be not only accurate and efficient but also trustworthy and privacy-preserving. This doctoral thesis proposes a unified framework that brings together three complementary paradigms — Edge Artificial Intelligence (Edge AI), Federated Learning (FL), and Fully Homomorphic Encryption (FHE) — to design distributed biomedical architectures capable of learning from data, collaborating across institutions, and operating securely on encrypted information. These pillars form a continuum: Edge AI enables local intelligence close to the patient, Federated Learning orchestrates knowledge across clinical sites without sharing raw data, and FHE ensures end-to-end confidentiality during computation. Together, they define a new paradigm of privacy-preserving biomedical intelligence. Within this conceptual vision, the thesis introduces the di-Pen, a smart add-on for insulin pens that embodies the principles of edge computing and personalized care. The device integrates multiple sensing modules — optical, proximity, and inertial — to monitor the conditions of insulin administration in real time. It supports patients in optimizing injection technique, angle, and site selection, while simultaneously acquiring physiological data for clinical analysis. Among its functions, the optical subsystem detects subcutaneous tissue alterations such as lipodystrophy, a frequent complication that compromises insulin absorption, while the inertial module classifies injection zones to guide safer and more effective insulin delivery. A clinical trial conducted at the Bambino Gesù Children’s Hospital validated the feasibility of this approach in pediatric subjects. Optical signals captured by the di-Pen were analyzed through lightweight machine learning models, revealing distinctive optical patterns associated with both lipodystrophic tissue and injection sites. The best-performing classifiers (Random Forest and XGBoost) achieved F1-scores between 0.78 and 0.87, confirming the feasibility of performing reliable inference directly on embedded hardware. When extending the dataset to multiple patients, performance showed a moderate decline due to increased inter-subject variability. This variability, however, reflects true physiological heterogeneity and can be overcome by scaling to larger and more diverse cohorts. These findings motivated the exploration of Federated Learning as a natural evolution toward distributed, cross-institutional collaboration without centralizing data. Expanding beyond the device, the thesis explores how distributed intelligence can connect multiple medical institutions through Federated Learning. Instead of aggregating sensitive data, FL enables each clinical node to train locally and share only model updates, preserving patient confidentiality and institutional autonomy. This paradigm was applied to heterogeneous biomedical datasets, including imaging data for prostate segmentation and clinical classification tasks. The experiments highlight the scalability, robustness, and regulatory alignment of federated architectures, achieving comparable accuracy to centralized training while maintaining full data sovereignty. To further enhance privacy during inference and model exchange, the research integrates Fully Homomorphic Encryption as a complementary security layer. FHE allows computations to be performed directly on encrypted data, enabling AI models to operate without ever revealing the underlying information. Experimental validation confirmed that encrypted inference can reproduce plaintext predictions with negligible accuracy loss and manageable latency once hardware acceleration is introduced. In addition to confidentiality, this approach mitigates potential adversarial and data-exfiltration attacks by ensuring that neither raw signals nor model parameters are exposed during computation. By merging Edge AI, Federated Learning, and FHE into a single architectural vision, this thesis contributes to the evolution of secure and explainable digital health ecosystems. It shows how intelligent devices like the di-Pen can act as edge nodes within a federated network, where encrypted models and data exchanges ensure compliance, transparency, and trust. The work illustrates a pathway from individual patient monitoring to collaborative, privacy-preserving analytics across healthcare infrastructures. In conclusion, this research offers both a technological and conceptual contribution: it demonstrates how privacy-enhancing technologies can coexist with clinical usability and real-world validation. The di-Pen stands as a tangible example of human-centric Edge AI, while the integration of FL and FHE projects this intelligence toward distributed, secure, and ethically grounded healthcare systems. The result is a comprehensive vision of AI-powered biomedical sensing — one that protects data, empowers patients, and fosters a trustworthy transition toward the next generation of intelligent healthcare.