Detection Of Keratoconus Using Data Derived Purely From Corneal And Stromal Pachymetry

Published 2024 - 42nd Congress of the ESCRS

Reference: PP26.08 | Type: Free paper | DOI: 10.82333/ben8-q743

Authors: Shady Awwad* ¹ , Anthony Abou Mrad ¹ , Jad Assaf ² , Bassel Hammoud ³ , Francesco Versaci ⁴ , Renato Ambrosio ⁵

¹American University of Beirut Medical Center,Beirut,Lebanon, ²Casey Eye Institute, Oregon Health and Science University,Portland, Oregon,United States, ³Cole Eye Institute, Cleveland Clinic,Cleveland, Ohio,United States, ⁴Costruzione Strumenti Oftalmici,Florence,Italy, ⁵Universidade Federal do Estado do Rio de Janeiro,Rio de Janeiro,Brazil

Purpose

To develop and evaluate a machine learning algorithm (MLA) for discrimination of keratoconus (KC), keratoconus susceptible (KCS), and normal (NL) corneas using only pachymetry maps.

Setting

Retrospective study.

Methods

133 preoperative eyes of 133 post-LASIK patients, 123 eyes of 123 frank KC patients, and 68 KCS of topographically/tomographically normal and borderline fellow eyes of KC patients (35 TNF; 33 TBF) were evaluated. OCT was performed on all eyes. Zernike decomposition of stromal and corneal pachymetry maps, centered on the thinnest point, was used to train the model with 10-fold cross-validation using a random search over 7 different MLA. The best model was used in 2-way (NL/KC), 3-way (NL/KCS/KC), and 4-way (NL/TBF/TNF/KC) analyses.

Results

10 Zernike polynomials achieved AUROCs of 0.7-0.95. 2-way analysis (NL and KC) yielded 100% sensitivity and specificity in detecting KC. 4-way analysis yielded sensitivities/specificities of 98%/86%, 40%/100%, 85%/99%, and 98%/100% in discriminating between NL, TNF, TBF, and KC corneas respectively.

Conclusions

Data solely based on pachymetry can be used to discriminate NL from KC and KCS with relatively high accuracy. Combining this methodology with other parameters such as curvatures and elevations could further improve diagnostic accuracy in KCS.