Target leakage, not model class, explains reported accuracy in survey-based cardiovascular screening: a leakage-tiered audit of glass-box and tabular foundation models
The audit says the high screening scores mostly came from leaked post-diagnosis survey fields.
Across ten model types, removing two leakage-prone features cut AUROC by about 0.05 and left the systems nearly tied. A glass-box explainable boosting machine matched the alternatives within the paper’s margin while running far faster than the strongest foundation model. The authors also report threshold-level sex disparity, then show that direct edits to the glass-box model reduced the gap. ArXiv · AI/CL/LG's note
Across ten model types, removing two leakage-prone features cut AUROC by about 0.05 and left the systems nearly tied. A glass-box explainable boosting machine matched the alternatives within the paper’s margin while running far faster than the strongest foundation model. The authors also report threshold-level sex disparity, then show that direct edits to the glass-box model reduced the gap. ArXiv · AI/CL/LG's note
score 4