Table 4.
Testing performance of the metric learning for the bi-class task.
| a (CI) | τHC(CI) | τMCI(CI) | τAD(CI) | stdτ (%) | |
|---|---|---|---|---|---|
| HC vs. AD | |||||
| k-nn | 80.7 (77.5–83.7) | 94.8 (91.7–96.6) | NA | 61.8 (55.5–67.5) | 23.3 |
| SVM | 90.6 (88.0–92.7) | 88.9 (84.8–92.0) | NA | 92.9 (88.6–95.5) | 2.8 |
| 1-layer NN | 91.2 (88.3–93.1) | 95.7 (93.3–97.6) | NA | 85.1 (79.7–89.0) | 7.5 |
| 2-layer NN | 91.9 (89.2–93.6) | 94.4 (91.6–96.5) | NA | 88.4 (83.9–92.1) | 4.2 |
| Perf. before CKA | 88.6 | 93.5 | NA | 82.1 | |
| ML + k-nn | 90.4 (87.4–92.6)* | 96.9 (94.6–98.4)* | NA | 81.7 (75.8–86.3)* | 10.7 |
| ML + SVM | 91.5 (88.7–93.5) | 93.2 (90.1–95.6)* | NA | 89.2 (84.0–92.7)* | 2.8 |
| Pretrained 1-layer NN | 92.0 (89.4–93.8) | 95.1 (92.4–97.2) | NA | 88.0 (83.0–91.6)* | 5.0 |
| Pretrained 2-layer NN | 92.6 (89.7–94.2) | 95.7 (92.9–97.5) | NA | 88.4 (83.5–91.8) | 5.2 |
| Perf. after CKA | 91.6 | 95.2 | NA | 86.8 | |
| HC vs. MCI | |||||
| k-nn | 59.5 (55.7–62.7) | 51.1 (45.8–56.3) | 65.9 (61.2–70.0) | NA | 10.5 |
| SVM | 64.9 (61.3–68.3) | 40.6 (35.0–45.9) | 83.5 (79.6–86.8) | NA | 30.3 |
| 1-layer NN | 70.4 (67.3–73.3) | 59.7 (54.1–64.9) | 78.6 (74.3–81.9)* | NA | 13.4 |
| 2-layer NN | 71.9 (68.5–74.8) | 59.1 (53.6–64.0) | 81.6 (77.6–84.7) | NA | 15.9 |
| Perf. before CKA | 66.7 | 52.6 | 77.4 | NA | |
| ML + k-nn | 64.7 (60.8–67.9)* | 66.5 (61.0–71.4)* | 63.3 (58.4–67.8)* | NA | 2.3 |
| ML + SVM | 66.0 (62.3–69.1) | 59.4 (53.7–64.7)* | 71.1 (66.3–75.1)* | NA | 8.3 |
| Pretrained 1-layer NN | 72.5 (69.1–75.6)* | 61.2 (56.1–66.5) | 81.2 (77.3–89.8)* | NA | 14.1 |
| Pretrained 2-layer NN | 71.2 (68.0–74.0) | 57.2 (51.5–62.0) | 86.9 (78.1–85.3) | NA | 21.0 |
| Perf. after CKA | 68.6 | 61.1 | 74.4 | NA | |
| MCI vs. AD | |||||
| k-nn | 67.0 (63.3–70.8) | NA | 84.4 (80.6–87.5) | 39.5 (33.7–46.2) | 31.7 |
| SVM | 68.5 (64.7–72.0) | NA | 76.0 (71.8–80.3) | 56.6 (50.8–63.0) | 13.7 |
| 1-layer NN | 67.9 (63.8–71.1) | NA | 74.3 (70.0–78.4) | 57.8 (51.6–63.6) | 11.7 |
| 2-layer NN | 71.1 (67.3–74.1) | NA | 79.0 (75.2–82.8) | 58.6 (52.0–64.7) | 14.4 |
| Perf. before CKA | 68.6 | NA | 78.4 | 53.1 | |
| ML + k-nn | 68.5 (65.0–71.8)* | NA | 78.7 (74.4–82.5)* | 52.3 (46.8–58.8)* | 18.7 |
| ML + SVM | 69.4 (65.3–72.6) | NA | 78.2 (74.1–82.2) | 55.5 (49.5–61.3) | 16.1 |
| Pretrained 1-layer NN | 69.2 (65.5–72.6) | NA | 78.7 (74.7–82.4)* | 54.3 (48.5–60.3)* | 17.3 |
| Pretrained 2-layer NN | 72.4 (69.1–75.8) | NA | 85.4 (81.6–88.5)* | 52.0 (46.5–58.8)* | 23.6 |
| Perf. after CKA | 69.9 | NA | 80.3 | 53.5 | |
Average and confidence interval (CI) for 1,000 resamples bootstrapping are reported.
*
Indicates whether the performance after CKA is larger than before CKA for each classifier at a significance level of 1%.
The gray shade denotes the best result for each performance measure.