Abstract Background The corneal back surface is known to add some against the rule
astigmatism, with implications in cataract surgery with toric lens implantation. This
study aimed to set up and validate a deep learning algorithm to predict corneal back
surface power from the corneal front surface power and biometric measures. Methods
This study was based on a large dataset of IOLMaster 700 measurements from two clinical
centres. N = 19,553 measurements of 19,553 eyes with valid corneal front (CFSPM) and
back surface power (CBSPM) data and other biometric measures. After a vector decomposition
of CFSPM and CBSPM into equivalent power and projections of astigmatism to the 0°/90°
and 45°/135° axes, a multi-output feedforward neural network was derived to predict
vector components of CBSPM from CFSPM and other measurements. The predictions were
compared with a multivariate linear regression model based on CFSPM components only.
Results After pre-conditioning, a network with two hidden layers each having 12 neurons
was derived. The dataset was split into training (70%), validation (15%) and test
(15%) subsets. The prediction error (predicted corneal back surface power CBSPP ?
CBSPM) of the network after training and crossvalidation showed no systematic offset,
narrower distributions for CBSPP ? CBSPM and no trend error of CBSPP ? CBSPM vs. CBSPM
for any of the vector components. The multivariate linear model also showed no systematic
offset, but broader distributions of the prediction error components and a systematic
trend of all vector components vs. CFSPM components. Conclusion The neural network
approach based on CFSPM vector components and other biometric measures outperforms
the multivariate linear model in predicting corneal back surface power vector components.
Modern biometers can supply all parameters required for this algorithm, enabling reliable
predictions for corneal back surface data where direct corneal back surface data are
unavailable.
