This preprint extends the embedding-based Early Warning System (EWS) framework introduced in prior work (Wirasinghe, 2026a; https://doi.org/10.5281/zenodo.19748474) for detecting unsafe normal predictions in automated OCT screening. The study provides a systematic analysis of distance metrics as representation-space safety signals, with a focus on Mahalanobis distance as a covariance-aware measure of deviation. Euclidean and cosine distances are evaluated as comparative baselines. Experiments conducted on the Kermany OCT2017 dataset demonstrate that Mahalanobis distance provides superior separation between normal and disease samples, with separation remaining consistent and improving across high-dimensional and PCA-reduced embedding spaces. This work establishes distance-based representation modelling as a principled foundation for post-classification safety validation in medical AI systems.
Ajantha Indunil Wirasinghe (Sun,) studied this question.