SNAKE optimised deep neural networks for Parkinson’s disease detection using multimodal handwriting as a digital biomarker

Abstract

Early identification of Parkinson’s disease (PD) remains a major clinical challenge, largely due to the lack of definitive biological biomarkers and the continued dependence on subjective neurological evaluations. Subtle motor abnormalities often manifest in handwriting behaviour, making handwriting analysis an increasingly recognised non-invasive digital biomarker for detecting early neuromotor dysfunction associated with PD. In this work, a multimodal deep learning framework is developed that combines handwritten image patterns with sensor-derived handwriting signals to improve automated detection of PD. For analysing spatial handwriting patterns, several deep learning architectures are evaluated, like conventional convolutional models, residual and efficient networks, and recent transformer-based architectures such as Vision Transformer, Swin Transformer, DaViT, and ConvNeXtV2. Temporal handwriting dynamics are analysed using sequential learning models designed to capture motor execution patterns, including Bidirectional Long Short-Term Memory (BiLSTM), Bidirectional Gated Recurrent Unit (BiGRU), and a hybrid Conv1D–BiGRU architecture. To improve learning stability and model generalisation, hyperparameters are optimised using the SNAKE metaheuristic optimisation algorithm. In addition, a decision-level ensemble fusion strategy using accuracy-based weighting is applied to combine predictions from image and signal models, enabling effective utilisation of complementary spatial and temporal information. The experimental evaluation on the NewHandPD dataset shows that transformer-based vision architectures and hybrid sequential models achieve superior performance for individual modalities. The proposed multimodal fusion framework achieves classification accuracies of 98.95% for meander patterns and 97.74% for spiral patterns, which represent statistically significant improvements compared with unimodal approaches. These results show that multimodal handwriting analysis can serve as a robust and scalable digital biomarker for early PD detection. The proposed framework could provide the foundation for future clinical decision-support systems and remote neurological screening platforms after further clinical validation. Not applicable.