Energy-Regularized Latent Synthetic Exposure for Open-Set Recognition of Clinical Tabular Data
Mohammad Sadman Tahsin, Haitham Y. Adarbah, MD Wahidur Rahman, Afzel Noore
IEEE Access, pp. 1-1
Abstract
Existing clinical tabular classifiers generally operate under a closed-set assumption, requiring every test patient to belong to one of the diagnostic categories observed during training. This assumption is restrictive in practice because previously unseen diseases or disease states may occur after deployment. Although open-set recognition has been extensively studied in computer vision, substantially less attention has been given to structured clinical data. We propose Latent Synthetic Exposure for Open-Set Recognition (LSE-OSR), a neural framework that constructs pseudo-unknown representations exclusively from known-class samples. LSE-OSR generates synthetic exposure representations through different-class latent interpolation and directional extrapolation, and regularizes them using uniform predictive uncertainty and a relative mean-energy objective. The synthetic representations serve as regularization points and are not assumed to lie exactly on a decision boundary or outside the support of the known classes. At inference, unknownness is measured using classifier energy, with the rejection threshold calibrated solely from known validation samples so that no true unknown disease influences training or threshold selection. We evaluated the method using repeated diagnostic-category holdout experiments on three public clinical tabular datasets. HCV was used as a development benchmark to finalize the LSE formulation, whereas Dermatology and ANN-Thyroid were evaluated using the frozen final formulation. Accordingly, the HCV results are reported as development-stage evidence and are not treated as independent confirmatory evidence. Relative to the corresponding cross-entropy energy baseline, the energy-scored LSE-OSR model (LSE-Energy) improved the area under the receiver operating characteristic curve (AUROC) from 0.8342 to 0.8937 on HCV, from 0.8717 to 0.9318 on Dermatology, and from 0.8519 to 0.8668 on ANN-Thyroid, while also improving the area under the open-set c
Cite this
Mohammad Sadman Tahsin, Haitham Y. Adarbah, MD Wahidur Rahman, Afzel Noore, "Energy-Regularized Latent Synthetic Exposure for Open-Set Recognition of Clinical Tabular Data," IEEE Access, pp. 1-1, 2026. doi: 10.1109/ACCESS.2026.3750217.