Table 1 Machine learning prediction of functional and morphological target variables from local and global features.

From: Unbiased identification of novel subclinical imaging biomarkers using unsupervised deep learning

 Visual functionOptical coherence tomographyFluorescein angiography
BCVA (letter score)LLVA (letter score)RT (\(\mu \hbox {m}\))IRC (nl)SRF (nl)PED (nl)Lesion area (\(\hbox {mm}^{2}\))Leakage area (\(\hbox {mm}^{2}\))
Local features
\(R^{2}\)0.260.440.650.090.440.200.270.22
MAE\(9.3\pm 7.1\)\(10.3\pm 6.5\)\(10.6\pm 11.0\)\(62\pm 48\)\(333\pm 3.3\hbox {e}6\)\(300\pm 248\)\(1.2\pm 1.0\)\(1.3\pm 1.0\)
Global features
\(R^{2}\)0.290.460.640.190.270.280.210.15
MAE\(8.9\pm 7.3\)\(9.7\pm 7.0\)\(10.9\pm 11.0\)\(54\pm 50\)\(342\pm 412\)\(286\pm 237\)\(1.4\pm 1.0\)\(1.3\pm 0.8\)
  1. For each outcome variable, the coefficient of determination (R\(^2\)) and mean absolute error (MAE) are shown. BCVA, best-corrected visual acuity; IRC, intraretinal cystoid fluid; LLVA, low luminance visual acuity; nl, nanoliter; PED, pigment epithelial detachment; RT, retinal thickness; SRF, subretinal fluid.