Parkinson's Disease Detection Using Deep Learning

Parkinsin's disease (PD) is very important to make early diagnoses of such neurodegenerative disorders leading to progressive motor and non-motor symptoms. Traditional methods of diagnosis rely purely on subjective clinical evaluation. Recent studies on machine and deep learning offer new avenues toward voice-based automated PD detection, which is minimally invasive and highly accessible. In this study, we analyze the performance of three deep architectures in the context of classifying PD from vocal features: CNN, RNN, and a combined architecture that uses both RNN combined with LSTM networks. We preprocess the data using the normalization method and perform train-test splitting for training our dataset. Each model is architected and trained with specific architectural choices to try to capture the underlying patterns in data. These metrics include accuracy, precision, recall, and F1-score that are used to evaluate the performance of these models. Preliminary results indicate that RNN + LSTM outperforms other architectures concerning accuracy and robustness, hence indicating its superior suitability for classifying time-series for PD detection purposes. Results of this study contribute to the growing body of evidence for advanced machine learning techniques in the early diagnosis of Parkinson's disease and pave the path for further studies to validate the model on larger datasets and possibly more features.