Performance Analysis of Eye Movement Event Detection Neural Network Models with Different Feature Combinations

Abstract

Event detection is the most important element of eye movement analysis. Deep learning approaches have recently demonstrated superior performance across various fields, so researchers have also used them to identify eye movement events. In this study, a combination of two-dimensional convolutional neural networks (2D-CNN) and long short-term memory (LSTM) layers is proposed to simultaneously classify input data into fixations, saccades, post-saccadic oscillations (PSOs), and smooth pursuits (SPs). The first step involves calculating features (i.e., velocity, acceleration, jerk, and direction) from positional points. Various combinations of these features have been used as input to the networks. The performance of the proposed method was evaluated across all feature combinations and compared to state-of-the-art feature sets. Combining velocity and direction with acceleration and/or jerk demonstrated significant performance improvement compared to other feature combinations. The results show that the proposed method, using a combination of velocity and direction with acceleration and/or jerk, improves PSO identification performance, which has been difficult to distinguish from short saccades, fixations, and SPs using classic algorithms. Finally, heuristic event measures were applied, and performance was compared across different feature combinations. The results indicate that the model combining velocity, acceleration, jerk, and direction achieved the highest accuracy and most closely matched the ground truth. It correctly classified 82% of fixations, 90% of saccades, and 88% of smooth pursuits. However, the PSO detection rate was only 73%, highlighting the need for further research.