Qualifying optical biometry data before cataract surgery using autoencoders

Purpose: To develop, implement and demonstrate a data driven strategy to identify ‘suspects’ and ‘outliers’ in datasets of biometric measurements taken before cataract surgery. Methods: A shallow autoencoder with one hidden layer and 3 neurons was trained on a large multicentre dataset (N = 152,397, 9 clinical centres) from the IOLMaster 700 including axial length (AL), central corneal thickness (CCT), anterior chamber depth (ACD), lens thickness (LT) and corneal front surface radius (R). Measurements were identified as ‘suspects’ or ‘outliers’ depending on the mean squared prediction error. Training was performed on 7 of the 9 datasets, and the remaining 2 datasets were used to evaluate the performance of the autoencoder. Results: Crossvalidation proved that the autoencoder with 3 neurons showed no noticeable overfitting. Measurements were marked as ’suspects’ or ’outliers’ based on the 95% and 99% quantiles as derived from the mean squared error with the training data. After piecewise linear correction of the autoencoder prediction error to correct for trend errors probably due to internal calibrations of the biometer, the autoencoder successfully identified potentially faulty measurements for the 2 test datasets. Conclusions: Autoencoders are quite popular for different application fields in engineering, and could also help in ophthalmology to identify potentially faulty biometric measurements which could lead to refractive surprises if used for intraocular lens power calculation. Further studies are required to validate this concept with other parameters or datasets from other biometers.