Efficient feature fusion model withmodified bidirectional LSTM for automatic Parkinson's disease classification

The majority of people affected by Parkinson’s disease (PD) are middle-aged and older. The condition causes a variety of severe symptoms, including tremors, limited flexibility, and slow movements. As Parkinson’s disease develops with changing symptoms and growing severity, the importance of computer-aided diagnosis based on algorithms cannot be highlighted. Gait recognition technology appears to be a potential path for Parkinson's disease identification since it captures unique properties of a person’s walking pattern without requiring active participation, providing stability and non-intrusiveness. To begin, the median filter is used to remove noise from the input images received during data collection. This paper describes a new method for finding local and global features in gait images to assess the severity of Parkinson’s disease.Local features are extracted using a stacked autoencoder, and global features are obtained using an Improved Convolutional Neural Network (ICNN). The Enhanced Sunflower Optimisation (ESO) technique is used to improve the CNN model's performance by optimizing hyperparameters such as batch size, learning rate, and number of convolutional layers. To classify PD severity, a modified bidirectional LSTM (MBi-LSTM) classifier receives input in the form of a combination of local and global features. The proposed model's performance is completely evaluated with the GAIT-IT and GAIT-IST datasets, which include key measures such as accuracy, precision, recall, and the F-measure. This study improves the diagnosis of Parkinson’s disease by introducing a non-intrusive real-time monitoring system capable of early detection and prevention.