A Cognitive-Driven Stacking Ensemble Approach for Dyslexia Handwriting Classification using LeNet-Based Deep Learning Models

Abstract

Dyslexia, a specific learning disability, affects cognitive processing in approximately 5-20% of children worldwide. In India, the prevalence of dyslexia is alarmingly high, affecting about 10% of the population. Despite advancements in technology, there is a concerning lack of attention towards screening children with Dyslexia. People with dyslexia often struggle with interpreting words and visual stimuli, yet with timely intervention through right education and training, can effectively improve their learning outcomes. Handwriting analysis has emerged as a promising approach for dyslexia detection by using the cognitive motor patterns linked with dyslexia. Although researchers have developed different ways to analyse handwriting patterns in dyslexia, there is still   need for developing more accurate and efficient methods for classifying dyslexic handwriting using advanced machine learning techniques. This research introduces a novel approach utilizing a Stacking ensemble deep neural network for classifying dyslexic handwriting. Our method employs four variants of the LeeNet-5-based Convolutional Neural Network (CNN), distinguished by different numbers of triple convolution layers used to extract cognitive-motor features from handwriting patterns. These models were trained and validated on a dataset using standard performance metrics. Experimental results revealed that increasing the number of feature extraction layers enhances model performance. Ensembling has been performed to combine the strengths of individual models and achieve better accuracy. Significantly, by stacking the four variants of the models, our Stacking ensemble approach attained an impressive accuracy of 96.86% in classifying the three classes of dyslexic handwriting. Notably, the Receiver Operating Characteristic (ROC) curves demonstrate perfect classification for the Corrected and Reversed classes with an Area Under the Curve (AUC) of 1.00, and an AUC of 0.96 for the Normal class, indicating the robustness and reliability of our proposed model. The promising results underscore the potential of deep learning and cognitive-driven handwriting analysis in advancing dyslexia screening, emphasizing the effectiveness of the Stacking ensemble approach in addressing this critical issue.