Abstract
Purpose :
To investigate machine-learning procedures to differentiate biomechanical related properties between different topographical severities of keratoconus (KC) by dynamic Scheimpflug tonometry.
Methods :
This prospective and monocentric study included 288 keratoconic and 104 healthy eyes of 288 KC patients and 104 healthy participants. Patient's age had to be between 18 and 45 years. The biomechanical assessment of theses eyes were performed with the dynamic Scheimpflug analyzer (Corvis ST, Oculus, Wetzlar, Germany). Dynamic corneal response parameters (DCR) were chosen by their relevance and checked for multicollinearity. The severity of KC was determined by the topographical keratoconus classification system (TKC) by Pentacam (Oculus). Patients were matched by biomechanical corrected intraocular pressure (bIOP) and age. Advanced cases (TKC = 4) were excluded from this study. Random Forest (RF) and linear discriminant analysis (LDA) were used to develop the models by using R statistics. Database were divided into training dataset (70% of cases) and test dataset (30%)
Results :
The mean bIOP was 15.9±1.84, 14.5±2.0, 14.3 ± 2.1 and 14.2±2.4 for controls (N=104), TKC 1 (N=83), TKC 2 (N=102) and TKC 3 (N=103), respectively. There were significant differences in bIOP between controls and TKC 1 as well TKC 2 (P < 0.05). In RF model, test dataset showed a sensitivity (sn) and specificity (sp) of 91%/85%, 50%/90%, 60%/92% and 84%/95% for healthy controls, TKC 1, TKC 2 and TKC 3, respectively. The overall accuracy was 72.3%. In LDA model, sn and sp was 97%/95%, 52%/87%, 56%/87% and 89%/89% for healthy, TKC 1, TKC 2 and TKC 3, respectively. The overall accuracy was 69.0%. Corneal thickness parameters (pachymetry and ARTh) had the strongest impact on both models followed by integrated radius and DA ratio. The RF model was significant more precise than LDA (P<0.001).
Conclusions :
It can be concluded that RF model showed a higher accuracy than LDA model detecting the stage of keratoconus. The developed models are currently limited in detecting TKC 1 and TKC2. Main influencing DCRs are reserved for pachymetric parameters.
This is a 2020 ARVO Annual Meeting abstract.