The rise of IoMT (Internet of Medical Things) has transformed healthcare by enabling instant health monitoring and tailored medical support, significantly improving patient outcomes. However, this progress also introduces serious security concerns. IoMT devices handle critical and privacy-sensitive medical data and operate in high-stakes environments where any security lapse can lead to data leaks, system failures, or Potential harm to human life. Safeguarding these systems is vital to ensure the healthcare system’s credibility and stability. Traditional binary classification methods often do not detect the wide range of attacks that can compromise IoMT systems. This research focuses on developing an efficient preprocessing pipeline to address data imbalance and reduce model complexity through mapped class-wise feature selection. It proposes a stacking ensemble machine learning model, integrating XGBoost and Random Forest with a Decision Tree meta-learner, to detect various protocol-based attacks. The model is evaluated using the realistic CICIoMT2024 dataset, achieving an impressive 99% accuracy and a 5% improvement in F1 score compared to baseline models. Furthermore, a detailed comparison with standard machine learning models shows enhanced classification performance, ensuring precise detection of malicious attack classes within the IoMT ecosystem.