An Embedding-Based Post-Classification Safety Layer for Detecting Unsafe Normal Predictions in OCT Screening

This preprint introduces an embedding-based Early Warning System (EWS) for detecting unsafe normal predictions in automated OCT screening. The framework operates as a post-classification safety layer, analysing representation-space geometry rather than relying solely on softmax confidence. It combines Mahalanobis distance-based atypicality detection, disease-direction alignment, and workflow-aware prioritisation. Evaluated on the Kermany OCT2017 dataset under leakage-corrected conditions, the system identifies structurally atypical normal cases that are not detectable using conventional confidence or entropy-based methods. The results demonstrate that embedding-space geometry provides a complementary safety signal for improving reliability in medical AI screening systems. This work forms the foundation for further research into representation-space safety and embedding-based reliability in clinical AI.