Dr.Jyoti Matalia, M10661, Dr.Shetty Bhujang K, Dr.Nandini C, Dr.Himanshu Matalia
Purpose: This study was undertaken to determine if we could use artificial intelligence(AI) to predict the association of different grades of myopia with corneal tomographical (TP) indices provided byPentacam HR and biomechanical indices (BP) obtained from the Corvis-ST (OCULUS optikgerätegmbh). The study was looking beyond most of the directly associated parameters such as axial length, age.
Methods:Institutional Ethics board clearance was obtained prior to commencement of the study. All the enrolled subjects and their parents were well informed and consent was taken. 980 eyes from 980 children (one eye was chosen randomly) aged 7-17 years were classified as emmetropia, low myopia if between -0.5D to -3D, moderate myopia if between -3D to – 6D and high myopia if more than -6D after undergoing standard cycloplegic refraction. Patients were then advised to undergoPentacam HR and Corvis-ST scans. All the tomographical and biomechanical indices from the respective instruments were recorded. The deformation data from the Corvis-ST was further analyzed using analytical viscoelastic biomechanical model (J Matalia et. al Cornea 2017) to derive corneal stiffness (kc), extra-corneal stiffness (kg) and extra-corneal viscosity (µg).
All the tomographical and biomechanical indices were analyzed using artificial intelligence (AI)softwareOrange 3.12 (University of Ljubljana). Cross-validated logistical regression algorithm was used for prediction. The prediction was done separately based on
- a) Tomographical indices (27 indices),
- b) Biomechanical indices (61 indices) and
- c) Combination of both (88 indices).
The efficiency of the prediction was assessed using area under the curve (AUC) and classification accuracy (CA) values reported by the AI software. AUC describe acceptance of the AI model, 1.0 represent perfect and 0.5 represent worst. CA describe the ability of the model to classify the cohort correctly. The analysis also calculates prediction score for every index, which represents the importance of a given parameter in making the classification.
Results:The cohort had21% emmetropia, 46% low myopia, 22% moderate myopia, and 11% high myopia.
Tomographicalindices from the Pentacam HR gave AUC =0.7 and CA =0.48. The highest predictive score was given by the following 5 parameters for the Pentacam HR:
- maximum keratometry,
- higher order root mean square (RMS) of aberration,
- lower order RMS of aberration,
- flat keratometry,
- steep keratometry.
Biomechanical indices alone gave an AUC=0.71and CA= 0.52. The highest predictive score was given by the following parameters of Corvis-ST:
- corneal stiffness (kc),
- extra-corneal stiffness (kg)
- a ratio stiffness (kc and kg)
- deformation amplitude ratio max at 1 mm,
- ratio of deformation amplitude to whole eye movement.
Combination of both corneal tomography and biomechanical indices gave an AUC =0.75 and CA = 0.54.
On combining both tomographical and biomechanical indices, the above parameters had the maximum predictive score with biomechanical indices scoring the highest. Hence overall, biomechanical indices were found to have a better influence in AI prediction than tomographical indices in classifying myopic groups.
Conclusion:This study is the first attempt to associate the tomographical and biomechanical indices to myopia. The study showed biomechanical indices could be a better additive factor in predicting pediatric myopia and tracking its progression. This is a novel technique that can change the interface of myopia diagnosis. Also this methodology of assessment could be used in determining best index in predicting a multi-factorial condition from the pool of available indices.One of limitation of the study is that this cohort was derived from patients visiting a tertiary care center, henceequal proportion was not present across the groups of myopia. Also study with larger sample size is required to further understand these associations.
Leave a Comment