Ai-Based Diagnostic Model For Early Keratoconus Integrating Corneal Topography And Biomechanical Parameters

To develop an intelligent diagnostic model based on Pentacam and Corvis ST data, and to assess the efficacy of the model in diagnosing early keratoconus.

This multicentre retrospective study included 1087 eyes of 657 patients from four hospitals in Wuhan, Chongqing, Kunming and Guangzhou between January 2017 and March 2024, of which 365 keratoconus（KC）, 193 subclinical keratoconus (SKC), 129 Forme fruste keratoconus eyes(FFKC), and 400 normal controls were included.

Twenty-five Pentacam parameters and 15 Corvis ST parameters were collected for each eye. 789 eyes were included as a training set,  198 eyes as a validation set and  100 eyes as an independent test set. A deep learning model for tabular networks (RMGNN) was built based on Python 3.10 and Pytorch, and after optimizing and training the model on the training set, the model was evaluated using the hold-out method (hold-out). The overall four-category performance of the model and the diagnostic efficacy results distinguishing each category were demonstrated on the validation set, and all baseline models were compared on the training and validation sets, and with five ophthalmologists on an independent test set.

The overall accuracy of the RMGNN model was 96.97% with an AUC of 0.9954, and the sensitivity for diagnosing normal cornea, FFKC，SKC and KC was 100.00%, 100.00%, 92.11%, and 95.83%, respectively, and the specificity was 100.00%, 98.31%, 98.12%, and 100.00%. Compared to the remaining five baseline models, the RMGNN showed the highest overall classification accuracy and AUC of 96.97% and 0.9954, respectively, exceeding the average accuracy of 18.95% by five ophthalmologists in the independent test set.

(1) The overall diagnostic efficacy of the four classifications of the RMGNN model was excellent, achieving a high overall accuracy rate and proving its excellent ability in the diagnostic task of accurate classification of conical corneas.

(2) Not only can the RMGNN model efficiently distinguish between KC and normal eyes, but the RMGNN model can achieve 100% sensitivity and 98.31% specificity for FFKC screening, and 92.11% and 98.12% for SKC screening, which can accurately, sensitively, and efficiently detect FFKC and SKC，which is far beyond the range of sensitivity of the traditional machine learning models and ophthalmologists.