Skip to main content
. 2020 Mar 20;27(5):700–708. doi: 10.1093/jamia/ocaa017

Table 4.

Diabetic retinopathy (DR) and Chest X-ray (CXR) mean and standard deviation test set accuracies across 10 runs for the various label distribution splits with centrally hosted and distributed training. Bold entries represent optimizations that resulted in significantly better performance than performance of cyclical weight transfer without optimizations for the same split

Model DR test set accuracyMean ± Std. Dev. CXR test set accuracyMean ± Std. Dev.
Split 6
 Central hosting 78.3 ± 0.9 76.5 ± 0.9
 CWT 77.9 ± 1.0 76.2 ± 0.7
 CWT + LWMS 78.0 ± 1.0 75.9 ± 0.9
 CWT + CWL 77.7 ± 1.2 76.4 ± 0.7
Split 7
 Central hosting 78.0 ± 0.7 76.7 ± 0.7
 CWT 77.0 ± 0.8 75.5 ± 0.5
 CWT + LWMS 77.7 ± 0.7 76.3 ± 0.9
 CWT + CWL 78.0 ± 0.8 76.1 ± 0.7
Split 8
 Central hosting 78.4 ± 0.5 77.0 ± 0.6
 CWT 76.3 ± 0.8 74.9 ± 0.8
 CWT + LWMS 77.3 ± 0.5 75.8 ± 1.0
 CWT + CWL 77.8 ± 0.7 76.2 ± 0.8
Split 9
 Central hosting 78.4 ± 0.7 76.2 ± 0.8
 CWT 75.9 ± 0.8 73.5 ± 0.8
 CWT + LWMS 77.1 ± 0.9 75.6 ± 0.6
 CWT + CWL 77.1 ± 0.6 75.1 ± 0.4
Split 10
 Central hosting 77.9 ± 0.6 76.6 ± 0.4
 CWT 74.4 ± 0.6 73.5 ± 0.7
 CWT + PLTI 76.8 ± 0.8 75.9 ± 0.7
 CWT + CWL 77.2 ± 0.8 75.4 ± 0.6