Privacy-preserving federated learning enables collaborative disease prediction while safeguarding sensitive patient data. This study explores the Federated Proximal (FedProx) algorithm within a federated learning framework to address the challenges of heterogeneous healthcare data. A simulated network of healthcare providers trains a shared model for disease prediction, particularly heart disease, using a synthetic multi-feature health dataset. Our methodology integrates data preprocessing techniques (handling missing values, mitigating outliers, normalisation, and anonymisation) and feature engineering (feature extraction, principal component analysis, and feature importance evaluation). FedProx-based training, coupled with split learning, enhances privacy and mitigates data heterogeneity. Experimental results demonstrate that the FedProx federated model achieves high accuracy, F1-score, and ROC-AUC, comparable to a centralised model, while ensuring strict privacy preservation. FedProx improves training stability across non-IID data sources, outperforming standard federated averaging (FedAvg). Feature importance analysis highlights Age, BMI, blood pressure, and sleep duration as key predictors and principal component analysis (PCA) confirms that two components capture 95% of data variance, validating dimensionality reduction techniques. This research confirms the viability of FedProx-enhanced federated learning for privacy-preserving disease prediction. Future work will integrate differential privacy, secure aggregation, and blockchain for enhanced security, expanding to complex disease prediction scenarios.