Federated Learning (FL) has emerged as a transformative paradigm for privacy-preserving healthcare analytics, enabling collaborative model training across decentralized medical institutions without raw data exchange [1, 7]. This study introduces a hybrid FL framework (ELSA+Prio+), combining CKKS homomorphic encryption (ELSA) [8, 11]. Fernet symmetric encryption (Prio+) [4] to enhance both security and performance in Parkinson’s disease classification. The proposed dual-path neural architecture (HybridComplexANN) integrates complex-valued and real-valued learning streams, achieving 93.1% accuracy—a 15% improvement over standalone ELSA [8]—while maintaining HIPAA-compliant data privacy [9]. Evaluated against FedAvg [1], SecAgg [2], Prio+ [4], and ELSA [8] over ten communication rounds, the hybrid model demonstrated superior convergence (loss = 0.42), stability (38% lower inter-round variance), and efficiency (35% faster encryption than ELSA alone) [5]. Fine-tuning optimized hyperparameters, including adaptive learning rates (0.001) and batch normalization, further reduced overfitting by 18% [15]. The framework’s robustness was validated through resistance to model inversion attacks (98% success rate) [3, 6] and SHA-256 integrity checks. These results establish the ELSA+Prio+ hybrid as a benchmark for secure, collaborative healthcare analytics, offering an unprecedented balance between diagnostic accuracy and privacy preservation.