Abstract
Early accurate drug prediction is crucial in clinical decision support, where privacy of the patient data is a paramount importance. In this study, we introduce a fused weighted adaptive federated learning (FWAFL) framework to achieve joint training among distributed healthcare institutions without requiring raw data sharing. The method employs local model updates and client-level adaptive weighting to enhance generalization and performance while preserving data privacy. A multilayer perceptron is fitted on tabular drug datasets in a decentralized manner, and an ensemble model is created by weighted averaging of the fitted local parameters. Validation results show that our approach outperforms the baseline federated and centralized approaches in both accuracy and robustness. The proposed approach demonstrates its promise for ensuring secure and privacy-preserving early drug prediction in real healthcare environments. An adaptive Federated Learning-based drug prediction approach is used to identify treatment early in the healthcare industry. The proposed model achieves an accuracy of 0.927 and a miss rate of 0.073, which is more accurate than the previously proposed approaches.