Federated learning enables collaborative intrusion detection across Internet of Medical Things (IoMT) deployments without centralizing sensitive patient data. In practice, institutions deploy heterogeneous sensing modalities that produce incompatible feature schemas and institution-specifc attack taxonomies with non-overlapping label spaces. A federated model
may therefore transfer well for some feature types and attack classes while failing silently for others, with no mechanism to identify which features remain semantically meaningful before
deployment. To close this gap, this paper presents a semantic feature transferability framework that organizes heterogeneous IoMT telemetry into canonical semantic categories and evaluates the cross-domain stability of each category using a dual node-edge criterion, which combines the consistency of SHAPbased feature importance rankings with the similarity of feature
correlation graphs across domains. Evaluated on three public IoMT datasets using leave-one-domain-out federated training with linear probing, the framework reveals substantial crossdomain degradation, with minority attack classes collapsing while majority classes are preserved. Restricting federated training to stable feature categories improves cross-domain F1 by 0.044
points on CICIoMT2024 and 0.043 points on MedSec-25, validating the criterion as a practical pre-deployment selection tool.

Federated learning,intrusion detection,Internet of Medical Things,feature transferability,SHAP,heterogeneous,cross-domain generalization,IoMT