EyeTrackDL: A Robust Deep Learning Framework for Saccade Detection via Simulated Data Augmentation

Abstract

Saccade detection is a fundamental task in visual behavior analysis and vestibular diagnostics. However, video Head Impulse Test (vHIT) recordings are often noisy, heterogeneous, and affected by class imbalance, particularly for covert saccades. In this paper, we propose EyeTrackDL, a lightweight yet effective deep learning framework based on a multilayer perceptron (MLP) architecture for classifying three types of eye movements: non-saccades, overt saccades, and covert saccades. Input signals are preprocessed using a fourth-order Butterworth filter, and two high-level features (onset time and duration) are extracted per saccade. To address data scarcity and imbalance, we apply SMOTE resampling and incorporate synthetic data generated from a kinematic vestibulo-ocular reflex (VOR) model. The model is evaluated using K-fold cross-validation (K = 2 to 10) on both real and simulated datasets. Results show that EyeTrackDL achieves an average accuracy of up to 96.5% on simulated data and approximately 83% in the real data, with significant improvements in the sensitivity of the covert saccades. Our findings demonstrate the potential of integrating simulation-based augmentation and class balancing for robust saccade detection in clinical environments.