Corvis St Biomechanical Indices In The Diagnosis Of Corneal Stromal Disorders. A Machine Learning-Based Comparative Study

To analyze the value of the Corvis ST biomechanical indices in the diagnosis of corneal stromal disorders (CSDs).

Institutional (Hôpital National des 15-20, Paris, France).

Design: Retrospective case-control study. Study population: 409 eyes with a CSD and 303 normal eyes (controls). Main outcome measures: Corvis ST indices. The collected data were divided into a training set (70%) and a test set (30%). Machine learning frameworks (logistic regression, ROC curves, and random forests) were used to distinguish each stromal disorder from controls and to classify corneas into six groups: keratoconus corneas (untreated or after cross-linking, ring segment implantation, or keratoplasty), high-risk corneas (forme frustre keratoconus, family history of keratoconus), laser corneal refractive surgery (PRK, LASIK, and Smile) corneas, Fuchs dystrophy corneas, glaucoma corneas, and normal corneas.

SSI significantly increased with age in the control group. Compared with controls matched for age, keratoconus corneas featured higher CBI (AUC in the test set, 0.92), lower SSI (0.78), and lower pachymetry (0.91); high-risk corneas featured higher CBI (0.84) and lower pachymetry (0.85); refractive surgery corneas featured higher CBI (0.90) and lower pachymetry (0.87); Fuchs dystrophy corneas featured higher SSI (1.00) and pachymetry (0.87); glaucoma corneas featured higher CBI (0.69) and lower pachymetry (0.79). The accuracy (percentage of corneas correctly classified in the test set) of the random forest classifier using CBI, SSI, and pachymetry as inputs was 76%.

Corvis ST indices are relevant for the diagnosis of CSDs and for distinguishing various CSDs from each other.