Recurrent Neural Network Optimized by Grasshopper for Accurate Audio Data-Based Diagnosis of Parkinson's Disease

Proposed here is a speech-based diagnostic framework for detecting Parkinson's disease that utilizes a Long Short-Term Memory neural network and the Grasshopper Optimization Algorithm. The framework aims to improve the detection of PD while ensuring accurate and efficient classification of speech-based signals. Incorporated within the system is a dimensionality reduction technique called Linear Discriminant Analysis. This approach allows the system to optimally utilize the features extracted from the given speech signal while ensuring that all the conditions necessary for successful classification are met. The LSTM neural network serves as the core deep learning model for this framework. It is trained on the speech signals produced by both healthy individuals and those with PD, allowing it to detect the class of a given speech signal based on the conditions necessary for successful classification. We assess the suggested technique on two speech datasets that are publicly available: NeuroVoz, a Castilian-Spanish body of work, and EWA-DB, a database in the Slovak language. The classification accuracy of our model on NeuroVoz is 99.45%, while on EWA-DB it is 99.71%. Moreover, these results exceed those from several contemporary approaches that were also tested on these two datasets. We interpret our results as strong evidence that uniting deep learning with nature-inspired optimization does yield an effective tool for providing, in a non-invasive way, a diagnosis of early-stage Parkinson's across diverse linguistic contexts.