Sensor-Derived and Deep Features in Binary and Multi-Class in Dysgraphia
Description
363 The dataset consists of processed public dysgraphia data, expanded into 260 sensor-based features, and incorporates deep learning features for binary (DB2) and multi-class (DB1) classification. Deep learning features were extracted using (a) DenseNet-201, (b) ResNet-50, (c) EfficientNet-B7, (d) ViT, (e) Swin Transformer, and (f) ConvNeXt. The dataset comprises three word levels: easy, difficult, and pseudo-words.
Files
Steps to reproduce
The public raw datasets include x, y, altitude, azimuth, time, pen status, and pressure. We generated 260 sensor-derived features by applying statistical measures—specifically the mean, median, standard deviation, percentiles (5th and 95th), maximum, and minimum of time-derivative functions—to each data component. Meanwhile, we developed deep features using two pretrained transformer methods (ViT and Swin Transformer) and four pretrained CNN architectures (ResNet, DenseNet, ConvNeXt, and EfficientNet). The public data used include: https://doi.org/10.1038/s41598-020-78611-9 and 10.1109/TCDS.2025.3597742.
Institutions
- State University of MalangEast Java, Malang