Enhanced multilevel autism classification for children using eye-tracking and hybrid CNN-RNN deep learning models
Suresh Cheekaty,G. Muneeswari
DOI: https://doi.org/10.1007/s00521-024-10633-0
2024-12-06
Neural Computing and Applications
Abstract:This article aims to improve the detection of ASD in children using ET data and advanced ML techniques. ASD, a neurodevelopmental disorder characterized by impairments in social communication, interaction, and repetitive behaviors, typically manifests before the age of three. Early diagnosis is crucial for timely and effective interventions. ET data have revealed distinct gaze patterns in children with ASD, such as diminished attention to social cues and increased fixation on repetitive stimuli. However, current methods of ET data analysis are largely manual and lack scalability for widespread clinical screening. To address these limitations, we propose a novel, scanpath-based ASD detection method that identifies atypical gaze behaviors through dynamic changes in visual attention. We extract four sequential features from scanpaths and analyze variations in feature space across different ASD severity levels low, mild, moderate, and severe using the MultiMatch and DTW similarity metrics. Our analysis reveals that children with ASD exhibit unique and highly individualized gaze patterns when compared to TD children. Notable differences are observed in attention duration and vertical gaze distribution, providing key insights into ASD-related visual behaviors. For classification, we employ a hybrid CNN-RNN model, which significantly outperforms traditional ML methods. The CNN-RNN model achieves an accuracy of 97%, recall of 98.24%, and an F1-score of 97.04% using the feature set ( x , d , y ). In comparison, models based on GRU and 2-LSTM networks show competitive accuracies of 92% and 90%, respectively. However, RFC and XGBoost models underperform, with accuracies ranging between 70.25 and 80.80%. These findings demonstrate the efficacy of DL approaches, particularly the CNN-RNN hybrid model, in accurately classifying ASD based on ET data, emphasizing their potential to enhance diagnostic accuracy. The proposed scalable method for ASD detection holds promise for improving early diagnosis and intervention, which is critical for better long-term outcomes in children with ASD. Further evaluation on an independent dataset will assess the generalizability and precision of the model.
computer science, artificial intelligence