Background The prediction of anatomical axial intraocular lens position (ALP) is one
of the major challenges in cataract surgery. The purpose of this study was to develop
and test prediction algorithms for ALP based on deep learning strategies. Methods
We evaluated a large data set of 1345 biometric measurements from the IOLMaster 700
before and after cataract surgery. The target parameter was the intraocular lens (IOL)
equator plane at half the distance between anterior and posterior apex. The relevant
input parameters from preoperative biometry were extracted using a principal component
analysis. A selection of neural network algorithms was tested using a 5-fold cross-validation
procedure to avoid overfitting. The results were then compared with a traditional
multilinear regression in terms of root mean squared prediction error (RMSE). Results
Corneal radius of curvature, axial length, anterior chamber depth, corneal thickness,
lens thickness and patient age were identified as effective predictive parameters,
whereas pupil size, horizontal corneal diameter and Chang-Waring chord did not enhance
the model. From the tested algorithms, the Gaussian prediction regression and the
Support Vector Machine algorithms performed best (RMSE = 0.2805 and 0.2731 mm), outperforming
the multilinear prediction model (0.3379 mm). The mean absolute prediction error yielded
0.1998, 0.1948 and 0.2415 mm for the respective models. Conclusion Modern prediction
techniques may have the potential to outperform traditional multilinear regression
techniques as they can deal easily with nonlinearities between input and output parameters.
However, in all cases a cross-validation is mandatory to avoid overfitting and misinterpretation
of the results.
