Abstract
This paper presents a comparative experimental evaluation of centralized, semi-decentralized, and hierarchical federated learning (FL) architectures combined with FedAvg, FedProx, and FedNova for premature non-communicable disease (NCD) mortality prediction. Using the Ecuadorian mortality registry (89,877 records), we construct a non-IID institutional setting with quantity, label, and feature skew across five hospitals and deploy all configurations on a six-device Raspberry Pi 4 cluster. The evaluation jointly considers predictive performance, convergence, communication overhead, energy estimates, CPU usage, RAM usage, temperature, and straggler delay. Hierarchical FedNova achieves the highest global F1-score (0.808) and accuracy (0.941), with near-zero straggler delay and the lowest cumulative energy among hierarchical configurations (15.35 J). Semi-decentralized FL reduces communication and energy cost but is sensitive to lottery-based aggregator selection under hardware volatility. Centralized FedNova provides the lowest estimated cumulative energy (Ecum=0.90 J) at the cost of lower predictive performance. The findings provide practical guidance for FL deployment in resource-limited, data-heterogeneous public-health environments under the evaluated hardware and data-partition conditions.
Funding Information
  • Convocatoria para el Financiamiento de Inscripción a Eventos para la Difusión de Resultados de Investigación—UITEY 2026

This publication has 28 references indexed in Scilit: