DysLSTM: a deep sequential framework for dyslexia prediction with comparative analysis of baseline and advanced models

Abstract

Dyslexia is one of the most common types of learning disorders, affecting both the reading and decoding of words in children and adults. The early detection is of paramount importance for timely intervention and yet the widely-used detection approaches in traditional systems are expensive, time-consuming and subjective. In this work, we propose DysLSTM: a robust deep learning framework that relies on Long Short-Term Memory (LSTM) networks to accurately predict dyslexia. We trained and tested it on an open dataset, collecting 3,644 individuals, with elaborate data pre-processing, feature mapping, scaling and balancing. Experimental results on multiple baseline machine learning (Logistic Regression, Random Forest, SVM) and deep learning (CNN, GRU, RNN) models are shown that DysLSTM outperforms all of them with 97% accuracy and high precision-recall scores as well as a Cohen's Kappa score of 0.91, which suggests that DysLSTM has nearly perfect agreement.