MCI-to-AD Prediction
The 16-feature GeoSAE representation achieved the highest AUC among the tested single-scan MCI-to-AD prediction methods. Removing non-specific features improved AUC, suggesting they added noise rather than useful conversion signal. GeoSAE…
1 sources - 4 claims
The 16-feature GeoSAE representation achieved the highest AUC among the tested single-scan MCI-to-AD prediction methods. Removing non-specific features improved AUC, suggesting they added noise rather than useful conversion signal. GeoSAE outperformed raw foundation model embeddings and several dimensionality-reduction or autoencoder baselines. Randomly selected alive GeoSAE features performed near chance, so prediction gains were not caused by arbitrary sparse subsampling.