A Machine Learning Driven Approach for Forecasting Parkinson’s Disease Progression Using Temporal Data

Parkinson’s disease (PD) is a prevalent neurodegenerative disease that hurts millions globally. For prompt intervention and successful therapy, Parkinson’s disease (PD) must be identified early. However, a significant obstacle to creating progression-predicting models for early detection is the absence of extensive and readily available data. This research aims to present a robust method to forecast Parkinson’s disease progression using minimal information. The Movement Disorder Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) is a commonly used metric for evaluating both motor and non-motor symptoms to determine the severity and course of Parkinson’s disease. The ML models in this research were evaluated based on their ability to predict the MDS-UPDRS values with minimum error with respect to the actual measurements. The preprocessing of this data includes clustering based on patient ID and linear interpolation to fill values between contiguous readings. To create a baseline, MDS-UPDRS [1-4] are treated as unsupervised time series data by setting the time stamp as the input variable and the corresponding MDS-UPDRS value as the output variable. Then, the data is re-framed into a supervised format and tested on the same set of models to compare performance. The data for this research is sourced from the Accelerating Medicines Partnership-Parkinson’s Disease (AMP-PD) Kaggle dataset. Among the machine learning algorithms assessed, the Multi Layer Perceptron (MLP) exhibited the most favorable predictive performance with an MAE of 0.52 and an MSE of 0.57 on average across the four clusters (U1, U2, U3, U4) for supervised format. This performance surpasses the second-best-performing algorithm, Random Forest, substantially. The MLP outperformed Random Forest by 25.46% in terms of MAE and 28.22% in MSE. Additionally, using a supervised format reduced MAE by 17.08% on average for all models and 41.73% for MLP compared to an unsupervised format. This considerable improvement suggests the robustness and efficacy of the MLP model and supervised re-framing in capturing the intricate progression trends of Parkinson’s disease.