Table 2.
ARIES Algorithm results for Experiments 1, 2, and 3
| EXPERIMENT 1a | ||||||
| AUC | Accuracy | Sensitivity | Specificity |
Patients in Group Ab (< 3 extension decisions taken at Visits 4, 5, 6, and 7) |
Patients in Group Bb (3 or 4 extension decisions taken at Visits 4, 5, 6, and 7) |
|
| Training from scratch | 0.77 | 64% | 91% (66/72) | 38% (14/36) | 72 | 36 |
| Following transfer learning | 0.87 | 77% | 83% (60/72) | 71% (26/36) | 72 | 36 |
| EXPERIMENT 2a | ||||||
| AUC | Accuracy | Sensitivity | Specificity | Patients with high IVT-AFL needc | Patients with low IVT-AFL needc | |
| YEAR 1 | ||||||
| Training from scratch | 0.77 | 63% | 84% (54/64) | 41% (18/44) | 64 | 44 |
| Following transfer learning | 0.84 | 75% | 81% (52/64) | 70% (31/44) | 64 | 44 |
| YEAR 2 | ||||||
| Training from scratch | 0.78 | 71% | 82% (86/105) | 59% (68/116) | 105 | 116 |
| Following transfer learning | 0.79 | 73% | 75% (79/105) | 71% (82/116) | 105 | 116 |
| EXPERIMENT 3a | ||||||
| AUC | Accuracy | Sensitivity | Specificity | Patients with short intervals (< 12 weeks) at end of Year 2 | Patients with long intervals (≥ 12 weeks) at end of Year 2 | |
| Training from scratch | 0.74 | 67% | 83% (96/116) | 52% (55/105) | 116 | 105 |
| Following transfer learning | 0.74 | 71% | 80% (93/116) | 63% (66/105) | 116 | 105 |
aThe numerator and the denominator of the fractions for sensitivity and specificity are mean values of the fivefold cross-validation; as such, the result of division may not correspond exactly to the percentage value specified
bGroup A: predicted ‘short’ interval at Visit 7; Group B: predicted ‘long’ interval at Visit 7
cHigh IVT-AFL need was defined as ≥ 8 and ≥ 5 IVT-AFL injections per year for Year 1 and 2, respectively; low IVT-AFL need was defined as < 8 and < 5 IVT-AFL injections per year for Year 1 and 2, respectively
AUC, area under the curve; IVT-AFL, intravitreal aflibercept