Breaking interprovincial data silos: how federated learning can unlock Canada's public health potential

Abstract

Abstract This study introduces the first dual-enabler framework that jointly integrates federated learning (FL) with artificial intelligence (AI) governance to enable privacy-preserving, interoperable collaboration across jurisdictions. Unlike traditional FL approaches, the proposed framework embeds governance mechanisms such as differential privacy (DP), consent enforcement and fairness auditing into the model lifecycle, ensuring secure, compliant and equitable deployment. These conditions result in siloed datasets, impeding national-level advancements in diagnostics, public health and personalized medicine. The framework’s novelty lies in its co-evolutionary design, where technical and governance layers iteratively inform one another to align with Canada’s federated healthcare architecture. We evaluate this framework via three simulations: cancer detection, pandemic prediction, and rare disease analytics and a real-world test using the PathMNIST dataset. Each case incorporates embedded governance mechanisms such as DP, consent enforcement and fairness auditing. Results show that FL combined with governance structures improves equity, deployment speed and compliance. For instance, in synthetic-data simulations, FL improved recall for underrepresented phenotypes by 18%, reduced compliance-violation flags by 35% and accelerated deployment. In the PathMNIST benchmark with 107 000 labelled medical images, our federated model achieved 91.3% accuracy under a strict privacy budget. Together, these results demonstrate a clear and scalable pathway towards ethical, privacy-preserving and governance-aligned AI for Canada’s healthcare ecosystem.

Ähnliche Arbeiten