Comparison of machine learning-based algorithms using corneal asymmetry vs. single-metric parameters for keratoconus detection

Purpose: To evaluate the diagnostic performance of three different parameter sets relevant to corneal asymmetry in comparison to conventional parameters including maximum anterior corneal curvature (K_max) and thinnest corneal thickness for diagnosis of keratoconus.

Methods: In this retrospective case control study, 290 eyes with keratoconus and 847 eyes of normal patients were included in the analyses. Corneal tomography data were acquired from Scheimpflug tomography. The sklearn and FastAI libraries were used in a Python 3 environment to create all machine learning models. The original topography metrics and derived metrics together with the clinical diagnoses were used as the dataset for model training. The data were first split to assign 20% of the data to an isolated test set. The remaining data were then split 80/20 to a training and validation group for model training. Sensitivity and specificity outcomes with standard parameters (K_max, central curvature, and thinnest pachymetry) and ratio of asymmetry across horizontal, apex centered, and flat axis-centered axis of reflection were studied via various machine learning models.

Results: Thinnest corneal pachymetry and K_max were 549.8 ± 34.3 µm and 45.3 ± 1.7 D in normal eyes and 460.5 ± 62.6 µm and 59.3 ± 11.3 D in keratoconic eyes. Use of only corneal asymmetry ratios across all 4 meridians had mean sensitivity of 99.0% and mean specificity of 94.0%, better than utilizing K_max alone or traditional measures combined (K_max, thinnest cornea and inferior-superior asymmetry).

Conclusions: By using the ratio of asymmetry between corneal axes alone, a machine learning model could identify patients with keratoconus in our dataset with satisfactory sensitivity and specificity. Further studies on pooled/larger datasets or more borderline population can help validate or refine these parameters.