Analysis of Freezing of Gait in Parkinson's Disease Detection Using a Multimodal Prototype Learning Framework

Patients with severe Parkinson's disease (PD) frequently have freezing of gait (FOG), a gait disability. By anticipating FOG before it occurs, pre-emptive cueing can either prevent FOG or lessen its severity and duration. To improve the accuracy of FOG detection, both electroencephalography (EEG) data and other complementary modalities, such as gait-based data, are increasingly being explored. The use of multimodal data is particularly important, as it enhances the robustness and accuracy of the detection models by combining different perspectives of the disease. Deep learning algorithms got a lot of attention in recent years for automated FOG identification; however their usefulness has been restricted due to a shortage of data samples, particularly medical data such as EEG. The scarcity of data can lead to overfitting in deep learning models, making it crucial for researchers to develop robust classification models that can operate effectively with a limited number of samples. Few-shot learning methods, such as prototype learning, have been introduced to mitigate these challenges by enabling models to effectively learn from a modest number of labeled samples. Thus in this research, we propose a prototype learning framework called CSE-ProtoNet, which utilizes CondenseNet with SEBlock for FOG detection in PD patients. Importantly, our study will leverage not only EEG data but also multimodal inputs, which can enhance the robustness and accuracy of PD detection. The method outperforms baseline models such as CSE-ProtoNet-ED, ProtoNet-CS, and ProtoNet-ED in terms of accuracy, F-score, recall, specificity, precision and AUC. The CSE-ProtoNet model also differentiated patients with FOG and Non-FOG with an accuracy of 98.75%. Cross-data validation was conducted to ensure the robustness and generalizability of the proposed method, confirming consistent performance across different folds.