Stacked Deep Ensemble Model for Learning Disability Detection

Objective: Many students face learning difficulties and recognizing them early is vital for timely support. This study set out to explore handwriting analysis as a way to identify such difficulties with greater accuracy. Method: Four well-known convolutional neural networks—AlexNet, VGG16, ResNet50, and InceptionV3—were tested on a dataset of handwritten samples. After comparing their baseline results, features from ResNet50 and InceptionV3 were combined and passed through a fully connected neural network to form a stacked ensemble model. Findings: ResNet50 gave the strongest single-model performance, with an accuracy of 93.6%, followed by InceptionV3 at 91.8%. VGG16 and AlexNet reached 88.4% and 86.7%. The ensemble improved on all of these, reaching 95.4% accuracy, with precision of 95.1%, recall of 94.7%, and an F1-score of 94.9%. Novelty: The study goes beyond using a single CNN by building a stacked deep ensemble that draws on the strengths of ResNet50 and InceptionV3. This combined approach improves reliability and accuracy, showing clear potential for practical use in early screening of learning disabilities from handwriting. Keywords: Tranfer learning, CNN architecture, AlexNet, VGG16, ResNet50, Inception, Learning Disability