Analyzing Handwriting Features for ADHD Detection in Children with Autism via Machine Learning

Attention deficit hyperactivity disorder (ADHD) develops in children's brains, portrayed by symptoms of inattentiveness, hyperactivity, and impulsivity. Early identification is very much essential, but traditional diagnostic methods might be subjective in practice and consuming time. Hence this work intends to introduce an end-to-end hybrid deep-learning system which fuses the techniques CNN and Bi-LSTM Networks to develop a system using handwriting to detect ADHD in children. The Bi-LSTM layers capture the sequential dependencies that exist in the handwriting data, capturing the temporal irregularities, whereas the CNN layers extract spatial features such as stroke patterns and pressure variations. The training and evaluation of the proposed model was done by synthetically balancing the dataset with accuracy, Confusion matrix, ROC curve and AUC for assessing performance. A method that has surpassed an accuracy of 90% in classification, this proposed system has been able to properly point out ADHD-related traits by hand. This method is scalable as it approaches to do fewer computations for the detection, thus making it a real data-driven alternative to traditional methods. Future work: Add attention mechanisms to increase the model's ability toward generalization, expand available datasets, and improve generalized capabilities.