Leveraging hybrid model of ConvNextBase and LightGBM for early ASD detection via eye-gaze analysis
ASD is a mental developmental disorder that significantly impacts the behavioural and communicational abilities of the child. ASD is affecting the world hard, and its global presence continuously increases. One of the reasons for this trend may be a pandemic, which increases screen time for children...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-06-01
|
| Series: | MethodsX |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2215016125000147 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | ASD is a mental developmental disorder that significantly impacts the behavioural and communicational abilities of the child. ASD is affecting the world hard, and its global presence continuously increases. One of the reasons for this trend may be a pandemic, which increases screen time for children and decreases communication with peers or family. A lengthy and subjective non-clinical procedure is currently placed for detecting ASD, which is followed by a series of therapy sessions to cure it. This research introduces a novel method for eye gaze analysis to identify autistic traits in children. This proposed work offers • A novel method of ConvNextBase and LightGBM leveraging eye position as a feature for early detection of autistic traits. • A new ConvNextBase architecture proposed with few unfreezed layers and extra dense layers with units of 512 and 128, respectively, and dropout layers with a rate of 0.5 that extract rich, high-level, and more complex features from the images to improve generalization and mitigate overfitting. • A LightGBM model performed classification using 3-fold cross-validation and found the best parameters for bagging_function, feature_fraction, max_depth, Number_of_leaves and learning_rate with values of 0.8, 0.8, −1, 31 and 0.1 respectively, to improve the model's robustness on unseen data.This proposed method is trained and tested on the publicly available Kaggle dataset, and results are benchmarked with other state-of-the-art methods. The experimentation finding shows that the proposed systems outperform other cutting-edge techniques in accuracy and specificity by 95 % and 98 %, respectively. Furthermore, the model achieved a precision of 93 %, showing that the model effectively reduces false positives and identifies false positives correctly. The classification process yielded 91 % under the AUC-ROC curve, showing the model's strong classification capability. |
|---|---|
| ISSN: | 2215-0161 |