Centralized versus Federated Learning for Multimodal Mental-Health Prediction: A Privacy-Focused Comparative Analysis

This paper proposes a privacy-preserving comparative framework to evaluate Centralized Learning (CL) and Federated Learning (FL) paradigms for multimodal mental-health prediction using text, speech, and facial cues. Existing centralized models achieve high diagnostic accuracy but suffer from privacy and compliance limitations under GDPR and HIPAA. To address this, the proposed framework systematically compares CL and FL across four key dimensions — accuracy, privacy, efficiency, and deployment feasibility — using normalized simulation data derived from established benchmarks such as AVEC and recent federated learning studies. Quantitative analysis reveals that FL attains up to 90% of centralized accuracy while reducing privacy exposure by over 80%, demonstrating its suitability for real-world healthcare deployment. The findings validate that while centralized systems remain ideal for research and prototyping, federated frameworks provide a more balanced, ethical, and regulation-compliant approach for implementing intelligent mental-health prediction systems in practice.