Table 3 Performance of machine learning-based formulas
From: Machine learning adaptation of intraocular lens power calculation for a patient group
Study | Carmona González D et al. (N = 260) [17] | Zhao J et al. (N = 53) [21] | Wei L et al. (N = 1450) [20] | Current | ||
|---|---|---|---|---|---|---|
Calculation method | Hill-RBF* | SVR + MAR | Hill-RBF* | Kane | XGBoost | SRK/T + SVR |
IOL | 10 models | 10 models | SBL-3 | SBL-3 | 8 models | SN60WF |
Country | Spain | Spain | China | China | China | Japan |
Mean prediction error (SD), D | − 0.17 (0.40) | 0.04 (0.30) | − 0.51 (0.61) | − 0.50 (0.60) | N.A. | 0.01 (0.38) |
MedAE, D | 0.28 | 0.18 | 0.55 | 0.45 | 0.29 | 0.21 |
Within ± 0.25 D | 48.1% | 65.4% | 24.5% | 28.3% | 43.9% | 54.4% |
Within ± 0.50 D | 80.8% | 90.4% | 47.2% | 52.8% | 72.8% | 83.5% |
Within ± 1.00 D | 100.0% | 100.0% | 81.1% | 83.0% | 99.1% | 98.5% |