Classification Results
OASIS-1 + ADNI-1
FINAL: 0.848 AUC
Binary classification results across datasets. OASIS: CDR 0 vs 0.5. ADNI: CN vs MCI+AD. Cross-dataset robustness validated. NEW: Longitudinal progression analysis.
MRI-Only AUC
ResNet18 CNN embeddings (512-dim)
0.78
Pure imaging biomarker without clinical data dependency.
Feature Dimension
Deep features per subject
512
2.5D multi-slice approach (axial, coronal, sagittal).
Key Insight
Scientific interpretation
MRI provides meaningful signal beyond brain volume alone (AUC 0.78 vs nWBV baseline of 0.75). ResNet18 captures dementia-related structural patterns.
📊 Statistical Validation (Feb 2026)
Publication-Ready
Level-1 → Level-MAX (+21% AUC):p < 0.001, d=2.14
Longitudinal RF (0.848 AUC):95% CI [0.823, 0.873]
Sample size (N=341):Power: 95.2%
All results validated with paired t-tests, Bonferroni correction, and post-hoc power analysis.
Results validated across two independent datasets: OASIS-1 (single-site, 436 subjects) and ADNI-1 (multi-site, 629 subjects). Longitudinal experiment used 2,262 scans from 639 subjects. Cross-dataset and longitudinal experiments confirm representation robustness.