Purchase this article with an account.
Gareth Lingham, James Loughman, Eoin Kerin, Samantha Sze-Yee Lee, David A Mackey, Siofra Harrington, Kathryn Jill Saunders, Daniel Ian Flitcroft; Machine learning-estimated axial length is better than spherical equivalent for identifying higher-risk myopic eyes. Invest. Ophthalmol. Vis. Sci. 2022;63(7):4336 – A0041.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To assess the performance of a machine learning-based algorithm in estimating axial length (AL) based on refraction and demographic data.
A machine learning-based algorithm (AL estimator) was trained using age, sex, spherical refractive error, astigmatism and corneal radius of curvature data derived from 4403 participants (aged 6-22 years) of Irish and Korean epidemiological studies. AL estimator performance was tested using right eye AL data from participants involved in myopia treatment trials (three Irish, one Australian). Bland-Altman statistics and linear regression were used to compare estimated and actual AL. Receiver operator characteristic analysis was used to assess the ability of the AL estimator to identify children with a high AL (≥26mm) and fast progressors (≥0.3mm axial elongation in 12 months), compared to spherical equivalent (SE).
The AL estimator was tested on 507 participants (n=354, 69.8% Irish), of whom 437 (86%) had 12-month follow-up data [mean age: 11.3 years, range: 6-17; female: 304 (60%); mean AL: 24.8mm, range: 22.1-28.9]. Using baseline visit data, the mean difference in AL (actual – estimated) was -0.07mm (95% limits of agreement [LOA]: -0.97, 0.83; absolute mean error=0.36; R2=0.80). When assessing 12-month AL progression, the mean difference between actual and estimated AL change was -0.001mm (LOA: -0.31, 0.31; absolute mean error=0.12; R2=0.57). The AL estimator performed worse for Australian compared to Irish participants (respective mean errors: AL estimate=-0.16 vs -0.03mm, p=0.003; AL change=-0.04 vs 0.02, p=0.001). Compared to SE alone, the AL estimator was better at identifying eyes with an AL ≥26mm (area under the curve [AUC]: 0.97 vs 0.83) and eyes that progressed ≥0.3mm over 12 months (AUC: 0.90 vs 0.85; Figure 1).
In this cohort of myopic children, the machine learning model was able to provide a reasonably accurate estimation of actual AL. Importantly, the AL estimator demonstrated high diagnostic performance in identifying individuals with long AL and those who exhibited excessive axial elongation. Where biometry is unavailable, the AL estimator may represent a useful clinical tool for identifying children at risk of axial growth-related complications of myopia.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
Figure 1: ROC curves of (a) estimated axial length [AL] and (b) AL change for AL estimator vs spherical equivalent AUC: area under curve
This PDF is available to Subscribers Only