PD-GLCST: integrating graph learning and sparse attention for accurate Parkinson’s disease diagnosis

• Published in: International journal of machine learning and cybernetics Vol. 16; no. 10; pp. 7637 - 7658
• Main Authors: Shermy, R. P., Santhi, S.
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.10.2025; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Artificial Intelligence; Classification; Complex Systems; Computational Intelligence; Control; Datasets; Deep learning; Diagnosis; Engineering; Handwriting; Identification; Literature reviews; Mechatronics; Medical imaging; Nervous system; Neurological diseases; Optimization; Original Article; Parkinson's disease; Pattern Recognition; Regularization; Robotics; Sensors; Support vector machines; Systems Biology; Handwritten images; Transformer; Sparse Attention; Parkinson’s disease; Deep Learning; Sensor signals
• ISSN: 1868-8071; 1868-808X
• DOI: 10.1007/s13042-025-02677-y

Parkinson’s disease is a fatal incurable neurological disorder that affects the nervous system of the brain and causes several health problems including rigidity, tremors, Bradykinesia, and so on. Timely and accurate Parkinson’s disease detection is mandatory for proper treatment planning, improving patient outcomes, and slowing down neurodegeneration. Various deep learning-based techniques are developed for early and precise diagnosis using handwritten records of healthy and Parkinson’s disease patients. However, traditional methods often struggle to identify whether the person is affected by Parkinson’s disease or not due to poor diagnostic accuracy and high computational complexity. This research paper designed a novel Graph Learning-based Convolutional Sparse Transformer methodology for early and accurate Parkinson’s disease detection and classification using handwritten images and their corresponding signals from NewHandPD, HandPD, and PaHaW datasets. The proposed method utilized a Convolutional SparseTransformer block that integrates a Dual Branch Local Perception unit, Sparse Attention mechanism, and Gated Depth-wise feed-forward network to derive the contextual information from a handwritten image and their sensor signals for enhancing diagnosis accuracy. Furthermore, the Intermediate Graph encoder is developed for learning the temporal interactions among the handwritten images and their respective signals and it performs feature fusion for modeling long-range context information. Moreover, the regularization term is introduced for learning the discriminative features and enhancing the classification performance. The effectiveness of the proposed methodology is evaluated using three different datasets with standard metrics. The simulation outcomes show that the proposed study achieves outstanding performance results in Parkinson’s disease detection and outperforms various previous models. The results of these experiments demonstrate that the proposed framework makes a valuable contribution to the precise detection of Parkinson’s disease, thereby improving patient treatment outcomes and quality of life.