Publication

Abstract : Anatomically inflammation in the anterior chamber of the eye, specifically in the iris and choroid is named as anterior uveitis. For the effective management of the disease it is essential for regular monitoring. Quantifying aqueous flare as a continuous measure of the intensity of light scatter (ILS) assists in accurately evaluating inflammation. Nevertheless, there is a potential for the subject’s blinking to disrupt the ILS data. This leads to increased and misleading levels of aqueous flare when assessing the extent of inflammation. Thus, our objective was to use an EOG-based spot fluorometer to examine the influence of eyeblink artifacts on ILS outcomes. This approach was founded on empirical data collected by quantifying the blink-induced and blink-artifact-free ILS in individuals with good health. A dataset of synthetic uveitis was generated using the LSTM deep learning technique. In addition, unsupervised machine learning techniques including k-means clustering, agglomerative hierarchical clustering, and Gaussian mixture clustering were used to identify blink artifacts in both the healthy and synthetic uveitis data. The optimal choice for minimizing artifacts was found to be the model that demonstrated superior performance. Our study findings indicate that the Gaussian mixture model outperformed other models in predicting blink-induced ILS, resulting in the most substantial decrease in blink artifacts. Furthermore, we successfully resolved the ILS by using our artifact removal technique, resulting in an impressive accuracy rate of 89%. The experiment verifies that our methodology successfully mitigates the occurrence of blinking errors in ILS measurements, thereby allowing a spot fluorometer to precisely grade uveitis.