A high dimensional dataset comprising of 756 samples and 754 initial features were used to evaluate the performance of the proposed framework. The initial analysis revealed the presence of great imbalance in terms of the number of healthy (HC) and parkinson disease (PD) cases, which were 25.4 and 74.6, respectively, with an imbalance ratio of 1:2.94. In an attempt to reduce possible bias of the model on majority class, Synthetic Minority Over sampling Technique (SMOTE) was used to balance the dataset to 1,128 samples with the class ratio of 1:1. The feature engineering was on 26 Mel frequency cepstral coefficients (MFCC) and 7 other vocal biomarkers. The dimensionality was then decreased to the top half of the most relevant attributes (378 features) to optimize the relationship between the cost of computation and prediction. The evaluation of the proposed HTFL-DBC framework was conducted with the help of the 5 fold stratified cross validation on the Parkinson Disease Classification Dataset. The sample has 756 voice samples containing 33 acoustic measures (MFCC coefficients and biomarkers unique to Parkinson, PPE, RPDE, DFA, jitter, shimmer). The framework detects the cases of the Parkinson with 90.62 percent precision and 76.99 percent recall is good to ensure that there are no false negatives which are important in early detection. The AUC value is 0.884, which means that there is a high discriminative ability between healthy and pathological voice patterns.
Hussain et al. (Sat,) studied this question.