X-CRSAA: An Explainable, Clinically-Regularized Self-Attention Autoencoder for Unsupervised Thyroid Anomaly Detection

The core contributions of this work are as follows:
1) A Feature-Tokenized Multi-Head Self-Attention (FTMHA) encoder paired with dual-decoder reconstruction
heads, optimized via a composite loss enforcing biological TSH/FTI constraints (Sections IV and V).
2) A Peaks-Over-Threshold EVT threshold calibration
pipeline that replaces heuristic percentile cut-offs with
extreme-quantile estimation (Section VI).
3) A dependency-free Kernel-SHAP explainability module
providing global and local clinical feature attributions
validated against synthetic ground truth (Section VIII).
4) An extensive empirical ablation demonstrating that
reconstruction-only attention optimization outperforms
masked-feature pretraining and latent-density fusion on
tabular medical data (Section XII).