June 2021
Volume 62, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2021
Accurate machine-learning estimation of axial length distribution in human populations from refraction, corneal curvature, age and gender
Author Affiliations & Notes
  • Brian Fitzpatrick
    Centre for Eye Research Ireland, Technological University Dublin, Dublin, Dublin, Ireland
  • Ian Flitcroft
    Centre for Eye Research Ireland, Technological University Dublin, Dublin, Dublin, Ireland
  • James Loughman
    Centre for Eye Research Ireland, Technological University Dublin, Dublin, Dublin, Ireland
  • Footnotes
    Commercial Relationships   Brian Fitzpatrick, None; Ian Flitcroft, Ocumetra Ltd (I); James Loughman, Ocumetra Ltd (I)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2021, Vol.62, 2339. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Brian Fitzpatrick, Ian Flitcroft, James Loughman; Accurate machine-learning estimation of axial length distribution in human populations from refraction, corneal curvature, age and gender. Invest. Ophthalmol. Vis. Sci. 2021;62(8):2339.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose : Human axial length (AL) is one of the most important ocular biometric parameters when it comes to understanding, tracking, and treating many ocular diseases. Ophthalmic datasets featuring AL are relatively small since AL measurements require specialized costly equipment. Datasets featuring simpler biometric parameters such as refraction but without AL can be very large in comparison since these variables are more straightforward to measure. We introduce a new machine learning technique that leverages refraction, corneal curvature, age and gender data to obtain estimates of the AL distribution in a population that can be at least as accurate as those given by direct measurement and kernel density estimation.

Methods : A regression-based machine-learning model was generated to predict AL from refraction, corneal curvature, age and gender (auxiliary data) based on a training dataset (n=383 records). Kernel density estimation (KDE) which incorporated this imputed AL information was then performed on a testing dataset (n=3809 records). On big datasets this approach is computationally intensive, so methods to accelerate the computations were explored.

Results : Comparison of the imputed AL distribution with conventional kernel density estimation of measured AL in the testing dataset showed no significant difference on the KS test at the 0.05 level. A Fast Gauss Transform (FGT) based algorithm significantly shortened computational time. The FGT algorithm accelerated the process 321-fold for a dataset of 800 eyes and 2087-fold for 4000 eyes (figure 1). Mathematical modelling demonstrated that the bias and variance of the estimator depend primarily on the number of input AL observations, the number of auxiliary data observations, and the properties of the auxiliary data. As a sample of AL measurements only estimates the true population AL distribution, a large sample of auxiliary data using this algorithm can match or exceed the accuracy of a smaller sample of AL measurements (figure 2).

Conclusions : The machine learning technique introduced in this work can accurately estimate the population distribution of AL by using refraction, corneal curvature, age and gender auxiliary data. This allows for the development of improved population centile data for AL from a wide range of epidemiological sources, which may be applied clinically in areas such as myopia control.

This is a 2021 ARVO Annual Meeting abstract.

 

 

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×