The quick digitalization of the Healthcare Information System (HIS) has facilitated the efficiency of clinical activities and the accessibility of data, but has also heightened the risks to the advanced cyber-attacks like ransomware, DDoS attacks, and data breaches. The conventional intrusion detectors are not sufficient in responding to these emerging threats because they are static-rule based and centralized in design. In this paper, the researcher suggests a hybrid, safe, and interpretable intrusion detection model that combines Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), Autoencoders, and XGBoost into a weighted ensemble model. In order to save data privacy, a type of federated learning is used that allows the inclusive training of models without providing sensitive healthcare data to other participants. Moreover, explainability by SHAP is also included in order to increase transparency and trust in model decisions. CICIoT2023 and CICIDs2018 datasets were used to test the proposed model. The results of the experiments are better, as they are up to 99.99% accurate on binary classification and 99.76% accurate on multi-class IoT intrusion detection. Random Forest (99.48%), XGBoost (99.33%), and LSTM (98.51) were compared to have lower performance as baseline models. The findings also prove the claim that federated learning is as accurate as centralized training with only slight degradation. In general, the suggested framework offers a scalable, strong, privacy-sensitive solution to the security of the current healthcare infrastructures.