Adaptive Binning Improves Self-Supervised Learning for Medical Tabular Data

Deep learning for tabular data, especially in clinical research, remains an underexplored area despite the abundance of structured clinical variables. The high cost of expert labeling for reliable data makes self-supervised learning (SSL) an attractive alternative. Existing binning-based SSL methods often use a fixed, global quantile discretization and apply feature-agnostic supervision. This paper introduces Adaptive Binning, a novel training-adaptive discretization pretext that couples the discretization process with learning through a feature-wise coarse-to-fine curriculum. The method progressively refines feature discretization and selects representation-aware splits, improving both value-space concentration and representation-space coherence. It unifies categorical reconstruction with ordinal supervision and demonstrates consistent performance gains on public medical tabular datasets, establishing a new benchmark for reproducible progress in this domain.