MetaboGuard-SSL v1 may be shared for self-supervised representation and
anomaly-scoring research.
It must not be published as a cancer/diabetes diagnostic or future-risk model. The former supervised DiaPan-XGB artifact remains invalidated.
| Task | Metric |
|---|---|
| Masked-feature reconstruction | Validation MSE |
| Representation stability | Embedding similarity across repeated corruption |
| Anomaly ranking | AUROC/AUPRC only against independently defined anomaly labels |
| Clustering | Silhouette/stability plus clinical interpretability |
| Linear probe | Cross-validated AUROC/AUPRC, explicitly cross-sectional |
| Missingness robustness | Score/embedding change after controlled masking |
Do not compare:
| Candidate | Use |
|---|---|
| AutoGluon TabPFN Mix | Frozen-representation or linear-probe comparator |
| Prior-Labs TabPFN | Generic tabular representation/classification comparator |
| Variational autoencoder | Probabilistic reconstruction baseline |
| Masked tabular transformer | Self-supervised representation comparator |
| Isolation Forest | Non-deep anomaly baseline |
| PCA | Linear representation baseline |
Every comparator must use identical folds, features and preprocessing boundaries.
A cancer/diabetes development benchmark becomes valid only when a dataset has:
Supported horizons are selected from 1, 3 and 5 years based on event availability. Each horizon requires at least 50 events and 50 eligible non-events for initial evaluation.