An Implementation of FastAI Tabular Learner Model for Parkinson’s Disease Identification

Abstract

This paper presents the implementation of the tabular learner model by FastAI for Parkinson’s disease (PD) identification. FastAI is one of the deep learning libraries that provide practitioners with high-level components that can quickly, efficiently, and easily. The dataset that we use is generated from the calculation of 29 acoustic parameters of the Multidimensional Voice Program (MDVP) in 3 different types of voice recording groups, namely PD voice, PD control voice, and Healthy voice. Our group of voice recordings was obtained from clinical assessments conducted by Taipei Medical University Hospital on PD patients and PD controls who were patients with suspected PD but not PD after clinical assessment. The healthy voice recording group we collected as a dataset contribution with a total of 55 healthy voice recording samples, 145 PD voice recording samples, and 55 control PD voice recording samples. Samples were recorded with vocal /a/ and vocal /u/ for 3 seconds with a recording distance of 5 cm, 16-bit quantization, and the sampling rate of 44.1 kHz. Our proposed model achieves the best results with an accuracy of 82%, root mean square error (RMSE) of 3.35 and mean absolute error (MAE) of 2.94 using a dataset of PD voice recordings and healthy voice recordings. It was concluded that the tabular learner model by FastAI has considerable potential for PD identification based on voice recordings.