Table 2.
Subgroup analysis for diagnostic accuracy of mtmDR retinopathy on color fundus photographs.
Features of subgroup | Dataa, n | Senb,c | Sped,c | LR+e,c | LR–f,c | AUROCg,c | I2 statisticc | Publication bias (P value) |
Overall mtmDRh | 40 | 0.95 (0.93, 0.97) | 0.90 (0.87, 0.93) | 9.7 (7.4, 12.7) | 0.05 (0.04, 0.08) | 0.97 (0.96, 0.99) | 29 (18, 40) | .11 |
Mean age> 50 years | 32 | 0.95 (0.93, 0.97) | 0.89 (0.85, 0.92) | 8.8 (6.4, 12.0) | 0.05 (0.03, 0.08) | 0.97 (0.95, 0.98) | 33 (20, 46) | .22 |
NNi algorithms | 38 | 0.95 (0.93, 0.97) | 0.91 (0.88, 0.93) | 10.1 (7.7, 13.2) | 0.05 (0.03, 0.07) | 0.98 (0.96, 0.99) | 30 (19, 41) | .14 |
Benchmark test sets | 15 | 0.92 (0.87, 0.95) | 0.90 (0.82, 0.94) | 9.0 (4.8, 16.6) | 0.09 (0.05, 0.16) | 0.96 (0.94, 0.98) | 25 (10, 39) | .22 |
Clinical Test sets |
25 | 0.97 (0.95, 0.98) | 0.90 (0.88, 0.92) | 10.0 (7.9, 12.6) | 0.04 (0.02, 0.06) | 0.97 (0.96, 0.98) | 30 (15, 45) | .06 |
External validation | 31 | 0.96 (0.93, 0.97) | 0.90 (0.87, 0.93) | 9.7 (7.2, 13.0) | 0.05 (0.03, 0.07) | 0.98 (0.96, 0.99) | 29 (16, 42) | .08 |
aMachine learning data derived from each data set was viewed as individual data, and we used “data” as the unit for calculation.
bSen: sensitivity.
cValues in this column are as follows: mean (95% confidence interval).
dSpe: specificity.
eLR+: positive likelihood ratio.
fLR–: negative likelihood ratio.
gAUROC: area under the receiver operating characteristic.
hmtmDR: more-than-mild diabetic retinopathy.
iNN: neural network.