Purchase this article with an account.
Luis Alfonso Hernandez, Valeria Sanchez-Huerta, Manuel Ramirez-Fernandez, Everardo Hernandez-Quintela; Combinatorial approach to determine top performing keratometric features and machine learning algorithms for keratoconus detection.. Invest. Ophthalmol. Vis. Sci. 2020;61(7):4750.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To determine the sets of 4 keratometric parameters, that best classify patients as keratoconus vs normal, as well as, which of 4 machine learning classification algorithms performs best at this task.
Keratometric data (Pentacam), age, gender and visual acuity from 20 patients (40 eyes) with keratoconus and 20 patients (40 eyes) without keratoconus was collected for statistical analysis. Twenty four female and 16 male patients, with a mean age of 28.8 years. Twenty six data points were included (shown in table 1). Data was analyzed using the python programing language. Out of the 26 data points, all combinations of 4 data points were registered and these were then input into 4 distinct classification algorithms: naive Bayes, decision-Tree, logistic regression and Support Vector Machine. Data was randomly split into training and testing sets. R2 data per classification algorithm was analyzed with an ANOVA test and post-hoc multiple comparison tests Bonferroni and Tukey. These 4 algorithms were also ranked based on R2 performance. Additionally, the 26 data points were also ranked based on the R2 performance of the top 100 combinations.
The ANOVA test on the R2 performance of the 4 classification algorithms showed a significant p of < 0.001. The results of Bonferroni and Tukey tests showed a significant p for the 6 comparisons between the 4 algorithms (p < 0.001 for all combinations). The best performing classification algorithm proved to be naive Bayes with an R2 of 0.9401 (95%Cl 0.93 to 0.94). The results for the other algorithms are shown in table 2. When it comes to the top performing data points, maximum progression index proved to be the most prominent feature, showing in 56 of the 100 classification algorithms with the highest R2.
In conclusion, R2 performance was significantly different among the 4 models. Naive Bayes showed that conditional probability worked best to train classification models. When it comes to the top performing data points, the first 3 were linearly dependent indexes (maximum progression index, Belin-Ambrosio D and medium progression index, respectively), while the 4th was the posterior curvature of the cornea, commonly cited as the first keratometric parameter to be altered in early keratoconus.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only