Continuous Model Adaptation in Distributed Healthcare Systems: A MLOps Framework for Federated Learning in Multi-Institutional Cardiovascular Risk Assessment

Abstract

When deploying machine learning models across multiple hospitals, we face a critical challenge: patient data is scattered everywhere, models degrade with changing populations, and no one wants to centralize sensitive health information. This paper explores how Machine Learning Operations combined with federated learning can help by letting each institution train models on its own data while sharing only algorithmic improvements, not patient records. We reviewed literature on operational model maintenance and privacy-preserving data sharing in healthcare organizations. Our main finding is that this MLOps-federated approach addresses three essential needs simultaneously: maintaining model performance through automatic retraining when it drifts, protecting patient privacy since raw data never leaves hospitals, and catching bias across diverse patient populations. The challenge lies in handling messy, non-independent clinical data across institutions with different computing resources. The real operational headaches emerge when managing distributed computational resources and when hospitals disagree on measuring model performance fairly. Effective governance requires clear protocols for model versioning, regular performance checks with intervention thresholds, and data quality standards respecting each institution's circumstances. Healthcare systems must invest in infrastructure for automated model management before attempting federated approaches, and regulators need frameworks treating these systems appropriately. This work bridges the gap between theoretical federated learning and actual hospital implementation, providing practical guidance for deploying responsible AI across multiple institutions while maintaining patient privacy.