Ensemble classifiers for an objective prediction of severity of uveitis based on measurement of aqueous flare and first-order patient characteristics

Abstract

Purpose : SUN (Standard Uveitis Nomenclature) scoring is subjective, hence its use to characterize the severity of inflammation in mild-to-moderate uveitis is clinically challenging. By leveraging objective measurements of the intensity of light scatter (ILS; as a measure of aqueous flare), we intend to achieve a robust prediction of the severity of inflammation by machine learning (ML) for enhanced clinical management of uveitis through consistent grading and granular indexing of the inflammation.

Methods : Patients diagnosed with uveitis were graded by SUN scores, and their ILS were recorded using a laser flaremeter (Kowa FM700). Normal subjects with no pathology served as controls. Ensemble method classifiers were used to predict the severity of inflammation based on ILS and first-order patient characteristics, which included the type of uveitis, status of iris/lens/pupil/cornea, and comorbidities such as diabetes.

Results : Fig. 1 is the summary of ILS in uveitis patients and normals. Regression showed a weak correlation between ILS and the SUN grade (p<0.0001, r2 = 0.41 on log2 of ILS). To train supervised ML classifiers, SUN scoring by one expert clinician was treated as the ground truth. Since our challenge is to distinguish mild-to-moderate uveitis, we focused only on data of patients with SUN scores of 0 and 1+. The dataset was imbalanced as the # of eyes w/SUN grade of 1+ was smaller compared to the # of eyes w/grade 0. Therefore, we explored data balancing approaches. The approach of the balanced bagging classifier, which incorporates under-sampling of the data, produced an F1 score of 0.92 with an accuracy of 90%. Another approach relied on SMOTE for data oversampling. The resulting balanced dataset was used to train the Random Forrest classifier, resulting in an F1 score of 0.88, with an accuracy of 83%. The classifier also identified the mean and SD of ILS as the top features of importance, thereby validating the significance of ILS measurements for grading the severity.

Conclusions : The aqueous flare is reliable for objective quantification of the intraocular inflammation. Our analysis has shown that ILS and its SD can reliably train ML models to grade uveitis on par with expert clinicians. Further optimization of the ML models and additional data on patients with Grade 1+ can be expected to enhance the performance of the ensemble methods.

This is a 2021 ARVO Annual Meeting abstract.