Exploration and analysis of On-Surface and In-Air handwriting attributes to improve dysgraphia disorder diagnosis in children based on machine learning methods.

Dysgraphia is a type of learning disorder that affects children's writing skills. Poor writing skills can obstruct students' academic growth if it is undiagnosed and untreated properly in the early stages. The irregularity in the symptoms and varying levels of difficulty at each age level made the dysgraphia diagnosis task quite complex. This work focuses on developing machine learning-based automated methods to build the dysgraphia screening tool for children. The proposed work analyzes the various attributes of online handwritten data recorded by digitizing tablets during On-Surface (when the pen is on the tablet's surface) and In-Air activity (when the pen is away from the tablet's surface). The proposed work has considered feature extraction from the whole handwriting data in a combined manner instead of feature extraction from task-specific (word, letter, sentence, etc.) handwritten data separately to reduce the number of features. This approach has significantly reduced the number of features by about 85%. Extracted features are used to train and evaluate multiple machine learning classifiers such as K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Random forest, and AdaBoost. Evaluation in a publicly available dataset indicates that the AdaBoost classifier achieved a classification accuracy of 80.8%, which is 1.3% more than the state-of-the-art method. Moreover, a deep analysis of different characteristics (kinematic, dynamic, temporal, spatial, etc.) of online handwriting is conducted to examine their significance in distinguishing normal and abnormal handwritten data. The analysis can help psychologists determine what attributes and methods should be considered for effective treatment.