An Integrated Neural Network Architecture for Convenient At-Home Diagnosis of Parkinson’s Disease

The goal of this project is to build a custom-made neural network architecture that provides a consistently accurate diagnosis of Parkinson’s disease. Speech disorders are one of the earliest indicators of Parkinson’s disease, and so the specific features within a person’s voice are speculated to be useful features in the diagnosis. Conventional Artificial Neural Networks (ANNs) are extremely useful in classification problems where they are able to detect complex non-linear patterns that human observation or other computer programs cannot. However, one of the problems of ANNs is their susceptibility to converging to a local minimum. Hence, a single-network architecture may be at constant risk of high generalization error. To overcome this shortcoming, this research demonstrates a unique architecture of neural networks, one that integrates the hypothesis functions with network weights being multiplied to denote how trustworthy a specific network is. This custom-made network architecture is shown to be consistently accurate in its diagnosis of Parkinson’s disease using vocal measures, despite its small size of data. Further real-life application of the architecture is briefly introduced, which paves the way for an authentic, convenient at-home diagnosis of Parkinson’s disease in the near future.