Mahalanobis Metric-based Oversampling Technique for Parkinson's Disease Severity Assessment using Spatiotemporal Gait Parameters

Parkinson's Disease (PD) severity level detection is crucial for timely and effective medical intervention. Due to the scarcity of PD gait data especially subject samples of higher severity levels, data generation techniques are adopted. This paper proposes the Mahalanobis-Metric-based Oversampling Technique (MMOTE), an algorithm that generates data within the boundaries of the existing data samples while also being diverse to address the problem of class imbalance within a dataset. The proposed technique is evaluated on PD gait data and the results show that MMOTE outperformed alternative oversampling techniques. A hybrid approach of combining ensemble learning with oversampling the minority class for PD severity level assessment is adopted. The minority class recognition is enhanced with an accuracy of 99%, thereby improving the generalizability of the classifier. Statistical analyses such as Levene's test and Wilcoxon signed-rank test are conducted to validate the significance of the findings. Moreover, the importance of optimal sample size determination for obtaining reliable prediction results is also discussed. Overall classification accuracy of ≈98% is reported using sample size estimated by plotting the learning curve for Random Forest.