Pancorneal Symmetry Analysis of Fellow Eyes: A Machine Learning Proof of Concept Study

Abstract

The goal of this proof of concept project was to test the feasibility of a data mining and machine learning approach to use pancorneal symmetry data in diagnostic algorithms which are commonly used in topographic examinations for distinguishing normal corneas from disease conditions. The data files used in this experiment were derived from a population-based sample of 5190 middle-aged adults who had undergone anterior segment imaging with the Pentacam computerized corneal topographer. Python packages such as Pandas, NumPy, Matplotlib, and Seaborn as well as various modules and codes were used to process the raw elevation data of the entire anterior corneal surface, compute elevation difference matrices, and create colormaps. The steps included data extraction, matching fellow-eye files, rotating the left eye matrix 180° around its Y axis, subtracting data on corresponding corneal point to create elevation differences matrices, exploratory analysis, data visualization, masking the matrix to access data points in five concentric central circles between 2.0 mm and 6.0 mm in diameter, and engineering features for clustering-based unsupervised machine learning. In data visualization, some of the common discernible patterns of interocular difference colormaps were “flat”, “tilt”, “cone”, and “4-leaf”. Clustering was performed with the Simple k Means algorithm in WEKA (Waikato Environment for knowledge analysis) using data from 4613 cases with bilateral data. The number of datapoints in the 2.0, 3.0, 4.0, 5.0, and 6.0 mm circles were 317, 709, 1257, 1961, and 2821, respectively. Mean elevation difference in the 6.0 mm data of the entire sample was 0.20 ±15.6 µm, and 99% of the data points were in the ±40 µm range. In 99% of individuals, the elevation differences in the central 3.0 mm were in the ±40 µm range. For each individual, the difference data in each circle was summarized into various statistics descriptive of central tendency and variability. The attributes using for clustering comprised of the central 95% range of the 6.0 mm circle data as well as the absolute mean, the standard deviation of the mean, and the volume difference attributes derived from the central 3.0 mm circle. Three clusters were generated: 3839 (83.2%) in Cluster 1 as the normal group, 618 (13.4%) in Cluster 2 as the intermediate group, and 156 (3.4%) in Cluster 3 as the abnormal group. Mean interocular differences for measures of corneal thickness and keratometry in these clusters were in agreement with their corresponding groups reported in the literature. Adding features derived from posterior elevation and thickness symmetry and previously developed diagnostic indices along with a larger (or automated) number of clusters could help improve the accuracy of the model and facilitate grouping interocular symmetry patterns.