Table 3.
Results achieved by different methods in SCD conversion prediction (i.e., pSCD vs. sSCD classification), with models trained on ADNI and tested on CLAS. Methods marked as “-M” denote that only subjects with MRI in ADNI are used for training, while the remaining methods employ all subjects with complete MRI and PET in ADNI for training. All 76 SCD subjects in CLAS are used as test data.
| Method | AUC | BAC | SPE | SEN | F1S |
|---|---|---|---|---|---|
| ROI-M | 0.621 | 0.642 | 0.784 | 0.500 | 0.511 |
| PBM-M | 0.592 | 0.516 | 0.490 | 0.542 | 0.413 |
| LLEP-M | 0.570 | 0.598 | 0.529 | 0.667 | 0.500 |
| HGAN+SCFR-M | 0.669 | 0.632 | 0.596 | 0.667 | 0.525 |
| ROI | 0.652 | 0.614 | 0.686 | 0.542 | 0.490 |
| PBM | 0.546 | 0.554 | 0.608 | 0.500 | 0.429 |
| LLEP | 0.611 | 0.536 | 0.529 | 0.542 | 0.426 |
| Zero | 0.648 | 0.647 | 0.711 | 0.583 | 0.528 |
| KNN | 0.652 | 0.646 | 0.750 | 0.542 | 0.520 |
| HGAN+SCFR | 0.632 | 0.553 | 0.539 | 0.583 | 0.452 |
| JSRL (Ours) | 0.747 | 0.721 | 0.692 | 0.750 | 0.621 |