Classification of depression, anxiety, and stress from handwriting and drawing using stacking models

Handwriting and drawing data are considered a meaningful data source in emotion analysis, as they contain dynamic and statistical features that reflect individuals' emotional states. This study aims to determine the optimal model combination for detecting emotional states using dynamic and statistical features obtained from handwriting and drawings. Feature vectors were derived from physical (kinematic), statistical, and signal processing (spectral, cepstral, and frequency domain) analyses. To address the class imbalance problem, oversampling and undersampling methods were applied together.With this approach, the number of samples in the minority class was increased through oversampling, while the majority class was reduced using undersampling methods.Dimensionality reduction was performed using feature extraction and feature selection methods. After identifying the minimum and most effective features, the samples were classified using a two-level ensemble learning method. In the study, four popular tree-based methods were selected as base and meta models. Model and hyperparameter optimization were carried out within the Optuna framework. The performance of the models was evaluated through experiments conducted on the publicly available EMOTHAW dataset. In this study, the emotional states of depression, anxiety, and stress were aimed to be determined using data obtained from seven different handwriting/drawing tasks of 129 individuals. The results show that the model demonstrates a noteworthy performance.