Smart decision support system for keratoconus severity staging using corneal curvature and thinnest pachymetry indices

Abstract

Background: This study proposes a decision support system created in collaboration with machine learning experts and ophthalmologists for detecting keratoconus (KC) severity. The system employs an ensemble machine model and minimal corneal measurements.

Methods: A clinical dataset is initially obtained from Pentacam corneal tomography imaging devices, which undergoes pre-processing and addresses imbalanced sampling through the application of an oversampling technique for minority classes. Subsequently, a combination of statistical methods, visual analysis, and expert input is employed to identify Pentacam indices most correlated with severity class labels. These selected features are then utilized to develop and validate three distinct machine learning models. The model exhibiting the most effective classification performance is integrated into a real-world web-based application and deployed on a web application server. This deployment facilitates evaluation of the proposed system, incorporating new data and considering relevant human factors related to the user experience.

Results: The performance of the developed system is experimentally evaluated, and the results revealed an overall accuracy of 98.62%, precision of 98.70%, recall of 98.62%, F1-score of 98.66%, and F2-score of 98.64%. The application's deployment also demonstrated precise and smooth end-to-end functionality.

Conclusion: The developed decision support system establishes a robust basis for subsequent assessment by ophthalmologists before potential deployment as a screening tool for keratoconus severity detection in a clinical setting.

Keywords: Corneal tomography; Feature selection; Keratoconus; Machine learning; Severity staging; Smart web.

Fig. 1

Workflow of the user-system interaction

Fig. 2

Development stages of the proposed staging predictor

Fig. 3

Sampling distribution of the collected dataset (n = 644)

Fig. 4

Pre-processing procedures applied to the study dataset

Fig. 5

Comparison between real samples (left columns) and augmented samples (right columns) in each stage

Fig. 6

The relative importance of features within the dataset for predicting the severity class labels based on the Gini method (n = 40). Asph_QB, asphericity coefficient (Q value) of the corneal back surface (posterior), asphericity Q value refers to the variation in the curvature of the cornea from its center to the periphery; Asph_QF, asphericity coefficient (Q value) of the corneal front surface (anterior); Astig_B (D), central corneal astigmatism (posterior corneal values measured in diopters); Astig_F (D), central corneal astigmatism (anterior corneal values measured in diopters); Axis_B (flat), corneal meridian of the least astigmatic power (posterior); Axis_F (flat), corneal meridian of the least astigmatic power (anterior); CKI, central keratoconus index; D0mm_Patchy – D10mm_Pachy, average pachymetry on concentric rings with radii (0–10 mm) around corneal thinnest point, respectively; IHA, index of height asymmetry; IHD, index of height decentration; ISV, index of surface variance; IVA, index of vertical asymmetry; KI, keratoconus index; KMax_Seg_Front (D), keratometry of the steepest point (anterior); Num_Ecc_B and Num_Ecc_F, Fourier-based posterior and anterior eccentricity in central 30 degrees, respectively; Pachy_Apex, corneal thickness in apex; Patchy_Min, thinnest pachymetry (µm); Pachy_Min_Pos_X and Pachy_Min_Pos_Y, x- and y-coordinates of the thinnest location, respectively; Pupil_Pos_X and Pupil_Pos_Y, x- and y-coordinates of the pupil position relative to the corneal apex, respectively; Pachy_Pupil, corneal thickness at the pupil center; Rh_F (mm), central radius in horizontal direction (anterior); Rm_B (mm), curvature radius of the back surface of the cornea (posterior); Rm_F (mm), curvature radius of the front surface of the cornea (anterior); Rs_F (mm), steepest radius (anterior); R_Per_F (mm), average anterior radius of curvature between 6 mm and 9 mm zone; R_Per_B (mm), average posterior radius of curvature between the 6 mm and 9 mm zone; Rv_B (mm), central radius in vertical direction (posterior); Rv_F (mm), central radius in vertical direction (anterior)

Fig. 7

Pairwise bivariate distributions of the selected features. Rm_B (mm), curvature of the back surface of the cornea (posterior), measured in mm; Rm_F (mm), curvature of the front surface of the cornea (anterior), measured in mm; Patchy_Min, thinnest pachymetry measured in µm

Fig. 8

The development process method

Fig. 9

Six-fold cross validation

Fig. 10

Out-of-bag error versus number of trees

Fig. 11

Confusion matrixes of the developed classifier models. a Logistic regression; b Support vector machine; c Random forest

Fig. 12

Structure overview of the flask framework

Fig. 13

Example test results for corneas at various KC severity stages. a Stage 0; b Stage 1; c Stage 2; d Stage 3; e Stage 4