Trustworthy Edge Intelligence for Continuous Cardiovascular Monitoring: Combining PPG Foundation Models with Adversarially Secure Medical AI Agents

Abstract

Continuous cardiovascular monitoring at the edge of healthcare networks promises to transform the detection and management of arrhythmias, hypertension, and heart failure by shifting computation onto wearable and near-body devices. Recent advances in photoplethysmography (PPG) foundation models have demonstrated remarkable generalization across diverse populations and sensor modalities, yet the deployment of such models within medical AI agent architectures raises profound questions of trustworthiness. This paper examines the system-level integration of PPG foundation models with adversarially secure medical AI agents operating at the network edge. We analyze architectural trade-offs among on-device inference, collaborative edge-cloud splitting, and federated learning topologies, highlighting the tension between model expressivity and resource constraints. A central contribution is a conceptual adversarial security framework that rethinks medical agent hardening in light of emerging models, where input perturbations, model inversion, and prompt-level attacks threaten both diagnostic accuracy and patient data privacy. We explore how statistical-prior informed generative masking strategies within PPG pretext tasks can serve as implicit regularization against distributional shifts and adversarial noise, while structured uncertainty quantification and certified robustness methods bolster agent reliability. The discussion extends into fairness auditing across demographic strata, governance mechanisms for federated model updates, and sustainability considerations for edge hardware lifecycles. By synthesizing cross-domain insights from embedded systems, foundation model training, and adversarial machine learning, we outline a trustworthy edge intelligence paradigm for cardiovascular care that balances clinical safety, data protection, and equitable access.