Classification of Depression, Anxiety and Stress from Handwriting and Drawing with Stacking Models

Recently, emotion analysis from online handwriting and drawings has become an important research topic. This study aims to determine the optimal model combination using dynamic and statistical features obtained from handwriting and drawings for detecting emotional states. The feature vectors are derived from physical (kinematic), statistical and signal processing (spectral, cepstral and frequency domain) analyses. To address the class imbalance problem, ADASYN (Adaptive Synthetic Sampling) and Tomek Links methods were applied together. In this way, the number of samples of the minority class has been increased, while over-sampled classes have been reduced. Dimensionality reduction was realized using PCA, and Gradient Boosting Classifier (GBC) feature selection. After determining the fewest and most effective features, the samples were classified using the two-level stacking ensemble method. eXtreme Gradient Boosting (XGBoost), Light Gradient-Boosting Machine (LightGBM), GBC, Random Forest (RF) methods are selected for base and meta models. Model and hyperparameter optimization was carried out by Optuna frameork. The performance of the models is evaluated through experiments in publicly available EMOTion recognition from HAndWriting and draWing (EMOTHAW) database. The aim was to determine the emotional states of depression, anxiety and stress with the data obtained from 7 different drawing/writing tasks of 129 people. The results obtained from accuracy, f1-score, precision and recall metrics show that the model provides remarkable performance.