Leveraging Bidirectional Long Short-Term Memory with Bayesian Optimization for Accurate Eye Movement Classification

The Covid-19 outbreak has accelerated scientific growth, especially in touchless technology to reduce disease transmission. Despite its 30-year history, the gaze-based user interface remains underused in touchless technology. This is due to the intricate challenges associated with eye movement classification, particularly for complex time series data. Despite prior efforts, hyperparameter optimization (HPO) for enhancing deep learning in eye movement classification has been largely overlooked. To address this scientific gap, we leveraged the Bayesian Optimization for Bidirectional Long Short-Term Memory (Bi-LSTM) to enhance the accuracy of eye movement classification. We compared the proposed method with several state-of-the-art deep learning models for time series classification. We also implemented other pervasive HPO techniques, such as Grid Search and Randomized Search. We focused on several key hyperparameters, including initialization mode, activation function, learning rate, and optimizer. Experimental results show that the Bi-LSTM with Bayesian Optimization achieved F-1 scores of 0.86 and 0.81 for fixation and smooth pursuit eye movement, respectively. The overall accuracy of the proposed method is 0.79. The experimental results indicate that the proposed method is promising for future development of eye movement-based touchless technology.