Subject-independent dyslexia detection using eye movement signals with noise-resilient deep learning

Dyslexia is a neurodevelopmental disorder that affects reading ability. Early identification and screening can support timely intervention. This study presents a deep learning framework for eye-movement-based dyslexia classification using binocular eye-tracking signals. To efficiently capture spatial and temporal gaze dynamics, the proposed framework uses a hybrid architecture that combines a convolutional neural network, bidirectional long short-term memory, and a temporal attention mechanism. A preprocessing pipeline that incorporates a second-order Butterworth IIR filter and a Constant Acceleration Model Kalman-based filter is used to reduce noise in eye-tracking signals. The framework is evaluated on two publicly available eye-tracking datasets with varying sampling rates, using 5-fold subject-wise GroupKFold cross-validation to reduce the risk of subject-level data leakage. On DS1, the proposed model achieved a subject-level accuracy of 93.15 ± 3.04% and an AUC of 98.40 ± 1.50, while on DS2 it achieved a subject-level accuracy of 81.43 ± 7.04% and an AUC of 91.70 ± 3.08. Subject-level aggregation is also evaluated to compare subject-level and window-level performance. A comprehensive feature ablation study further evaluates the contribution of velocity, acceleration, and vergence features to classification performance. The framework is evaluated independently on both datasets, providing evidence of subject-independent performance under the respective dataset-specific evaluation protocols. Overall, the proposed method provides a subject-independent framework for dyslexia-related classification, with potential for future computer-assisted screening.