HybridNet‐IoMT: Parallel Multi‐Branch Deep Learning Framework for Cybersecurity Attack Detection in Internet of Medical Things

The internet of medical things (IoMT) connects life‐critical clinical devices over resource‐constrained protocols, exposing healthcare networks to cyberattacks with direct patient‐safety consequences. Existing intrusion detection systems (IDS) for IoMT exhibit three recurring limitations: (1) single‐branch or serial hybrid architectures that capture only a partial view of feature structure; (2) uniform temporal pooling that dilutes the signatures of low‐rate attacks; and (3) inadequate handling of extreme class imbalance. This paper presents HybridNet‐IoMT, a parallel tri‐branch architecture in which a one‐dimensional convolutional neural network, a bidirectional gated recurrent unit (Bi‐GRU) with four‐head self‐attention, and a multilayer perceptron (MLP) operate simultaneously on a shared PCA‐reduced input, capturing local co‐occurrence patterns, bidirectional sequential dependencies, and global non‐linear feature interactions in parallel rather than serially. Class imbalance is addressed through a joint strategy of class‐balanced random over‐sampling (ROS) and label‐smoothing cross‐entropy, whose contributions act on orthogonal axes, sample frequency and gradient magnitude, of the optimisation problem. HybridNet‐IoMT is evaluated on three public IoMT datasets, CICIoMT2024 (19 classes), TON_IoT (10 classes), and WUSTL‐EHMS‐2020 (3 classes), achieving macro‐ scores of 99.43%, 99.71%, and 98.27%, respectively, obtaining the highest macro‐ on TON_IoT and WUSTL‐EHMS‐2020 and statistical parity with the strongest of eleven competing methods on CICIoMT2024. Statistical significance is confirmed via Wilcoxon signed‐rank tests, McNemar's test, and bootstrap confidence intervals over 20 independent runs, using a Bonferroni‐corrected significance threshold. Ablation analysis identifies random over‐sampling as the single most decisive component, contributing up to 3.45% macro‐ recovery on TON_IoT and WUSTL‐EHMS‐2020, while attention‐weight and PCA‐ordering analyses confirm that the parallel design and class‐adaptive temporal attention are jointly responsible for the framework's robustness across heterogeneous IoMT environments.