Integrating Sandplay Analysis and CNN–SVM Modeling for Predicting English Learning Difficulties among Vocational College Students: A Computational Psychology Approach
Year: 2026
Type: article
Abstract: This study integrated sandplay analysis with a hybrid convolutional neural network–support vector machine (CNN–SVM) framework to identify English learning difficulties among vocational college students, with the aim of improving the objectivity and reproducibility of projective assessment in educational psychology. A total of 190 students were included, of whom 134 were classified as having English learning difficulties and 56 as non-difficulty cases. Group assignment followed a dual-criterion approach: students with semester English scores below 60, low academic self-efficacy, and high foreign language anxiety were assigned to the difficulty group, whereas those with scores above 70, high self-efficacy, and low anxiety were assigned to the non-difficulty group. Under standardized and controlled conditions, all participants completed a sandplay task, and the resulting images were preprocessed through cropping, normalization, and resizing, with data augmentation applied to mitigate class imbalance. A shallow CNN was used to extract symbolic and spatial features from the sandplay images, and the resulting feature embeddings were then classified using an SVM with a radial basis function kernel. Model hyperparameters were optimized through nested cross-validation, and robustness was further evaluated using repeated hold-out validation. The CNN–SVM model achieved a mean validation accuracy of 91.5% and an area under the curve (AUC) of 0.95. On the independent test set, the model reached an accuracy of 83.9%, and recall for the difficulty group was 96.2%, indicating strong sensitivity and a reduced likelihood of missing at-risk students. Grad-CAM visualizations further showed that figurine clustering and central spatial arrangements were among the most important discriminative cues, consistent with established interpretations of sandplay symbolism. Overall, the findings suggest that quantifiable symbolic cues in sandplay can support accurate prediction of learning difficulties, and that this reproducible framework offers a promising bridge between projective techniques and artificial intelligence for early screening. Future research should include larger and more diverse samples, as well as broader assessment indicators, to strengthen the external validity of the model.
Source: Applied & Educational Psychology
Language: English
Subfield: Education
Field: Social Sciences
Domain: Social Sciences
SDG: Quality education
Open Access status: hybrid
Awards: G143213130, G8564267554
OpenAlex ID: W7154830449
