Adaptive Feature Selection using Fisher-Based Supervised Hill Climbing for Dysgraphia Handwriting Classification

Dysgraphia features selection remains a challenge. Fisher’s criterion excels at highlighting the discriminative features of dysgraphia but lacks guidance for choosing the optimal number of features. Whereas Hill Climbing shows robust feature selection but often gets trapped in local optima. This study aims to avoid the Hill Climbing trap in local optima when selecting the best dysgraphia feature. Thus, the Fisher-Based Supervised Hill Climbing (FSHC) method is introduced. The contribution of this study is an optimized machine-learning-guided hill-climbing method that uses a classifier on a validation set as the objective function. A plateau mechanism also guided Hill Climbing exploration, not by a single Fisher point but by the neighboring subsets. The dataset used contains the graphomotor slant line task from 119 children aged 8-15 years (47.5% diagnosed with dysgraphia), with 10000 to 50000 data points per user. It is organized into kinematic, spatial, dynamic, and temporal features, yielding 117 sub-features. A stratified 5-fold cross-validation is set for training and testing, reaching 21 features. Comparative test—Linear SVM, SVM RBF, Sigmoid SVM, Polynomial SVM, Random Forest, AdaBoost, KNN, Decision Tree, Gradient Boosting, Gaussian Naive Bayes, and Gaussian Classifier—showed that linear SVM achieves the best performance with a weighted average precision, recall, and F1 score of 0.93. Linear SVM also outperformed the three approaches: no feature selection, the traditional Fisher, and machine-learning-based feature selection (weighted KNN and SVM). It can be concluded that the proposed method is more robust than the state of the art by highlighting key points for avoiding overfitting.