Skip to main content

Table 2 Prediction performance of the stacking machine learning models in the testing set

From: Artificial intelligence-based refractive error prediction and EVO-implantable collamer lens power calculation for myopia correction

 Parameters

SD (D)

MAE (D)

MedAE (D)

Interquartile of AE (D)

Percentage of eyes within the ranges (%)

     

 ± 0.25 D

 ± 0.5 D

 ± 0.75 D

Postoperative SE prediction of NT-ICL cases (n = 130)

 MVF

0.525

0.385

0.310

0.391

43.08

70.77

87.69

 Random forest

0.445

0.339

0.268

0.372

48.46

73.08

89.23

 LASSO

0.448

0.347

0.303

0.399

44.62

74.62

91.54

 SVR

0.477

0.359

0.275

0.415

48.46

73.08

88.46

 XGBoost

0.485

0.385

0.305

0.402

43.08

70.77

87.69

Postoperative sphere prediction of NT-ICL cases (n = 130)

 

 MVF

0.485

0.392

0.335

0.375

36.15

73.85

84.62

 Random forest

0.489

0.386

0.336

0.312

40.77

70.00

86.15

 LASSO

0.487

0.387

0.312

0.347

41.54

64.62

88.46

 SVR

0.486

0.386

0.347

0.328

40.00

67.69

87.69

 XGBoost

0.513

0.403

0.328

0.391

36.92

70.77

86.15

Postoperative SE prediction of TICL cases (n = 205)

 MVF

0.470

0.341

0.269

0.311

47.32

80.48

92.68

 Random forest

0.462

0.333

0.264

0.320

48.78

80.49

90.24

 LASSO

0.457

0.337

0.275

0.333

47.32

78.54

94.15

 SVR

0.461

0.337

0.271

0.336

46.83

79.02

93.66

 XGBoost

0.452

0.325

0.257

0.316

50.24

79.51

92.20

Postoperative sphere prediction of TICL cases (n = 205)

 MVF

0.460

0.323

0.243

0.328

53.66

81.46

90.73

 Random forest

0.447

0.313

0.232

0.351

54.15

81.46

92.20

 LASSO

0.462

0.325

0.231

0.354

51.71

78.05

91.71

 SVR

0.467

0.330

0.238

0.343

52.68

78.05

91.22

 XGBoost

0.443

0.308

0.241

0.344

52.68

80.98

92.20

  1. NT-ICL = non-toric implantable collamer lens; TICL = toric implantable collamer lens; MAE = mean absolute error; MedAE = median absolute error; AE = absolute error; SD = standard deviation; D = diopters; MVF = modified vergence formula; SVR = support vector regression
  2. The mean prediction error was adjusted to zero in each subgroup. The numerically smallest SD, MAE, MedAE, and AE quartile are marked in bold for each subgroup. The P value of the prediction error distribution of postoperative spherical equivalent after TICL implantation was 0.014. The distribution of the prediction error of random forest and XGBoost differed from that of the modified vergence formula P = 0.016 and 0.045, respectively. All the other P values of the Friedman test were > 0.05