Parkinson's Disease Detection - An Interpretable Approach to Temporal Audio Classification

Model Interpretability is critical for analyzing the potential risks inherent to the utilization of black-box neural networks in the medical domain. The current methodologies for Parkinson's disease (PD) detection require expertise from medical professionals alongside an expensive and time-consuming procedure. AI-driven Deep Learning (DL) models have generated state-of-the-art performance and have the potential to improve the process of PD detection by leveraging the early stage symptom of vocal deterioration to enable early identification of the disease. However, the potential of DL for PD selection has remained untapped thus far because of the lack of model interpretability and the possibility of even an accurate model failing to capture the correct causal relationship between the input signal and the prediction. This work implements a neural network that contains vowel phonations. The model achieved a classification accuracy of 90.32% with a precision, recall and F1 Score of 0.91, 0.90 and 0.905. A novel model interpretation algorithm is proposed that divides the audio file into logical chunks and provides a category label identifying the productivity of each chunk towards correct inference. This is based on its interactions with every other audio region in the file, and the effective change of model performance and confidence with its inclusion. An audio instance within a clip is significant for desired output only in conjunction with other audio regions that together generate a meaningful pattern indicating the existence of PD. The algorithm also ranks all audio chunks in the order of their importance to the correct prediction by analyzing the degree in which their impact varies when combined with other audio regions. This paper demonstrates that DL can be leveraged to create a reliable, accurate, and efficient method for PD Detection and the model can be extended to provide interpretability to the inferences in a way that is scalable, and free from bias.