Speech Data for Predicting Parkinson's Disease Progression: A Machine Learning Solution for Remote Monitoring

This project is based on the UCI Parkinson's Disease Dataset and builds a complete machine learning pipeline. It predicts the severity and progression of the disease using speech acoustic features, providing a practical technical solution for remote medical monitoring of Parkinson's disease and addressing the cost and inconvenience of frequent hospital visits in traditional monitoring.

Parkinson's disease is a chronic progressive neurological disorder. Early symptoms are hard to detect, and frequent hospital visits increase patients' time and economic costs. Studies show that the disease affects vocal cord and respiratory muscles, leading to measurable changes in speech such as Jitter (fundamental frequency perturbation), Shimmer (amplitude perturbation), HNR (Harmonic-to-Noise Ratio), and voice disorders. These features can be collected via smartphone/computer microphones, providing a basis for remote monitoring.

The project uses the UCI dataset, and the pipeline includes: 1. Data cleaning and feature engineering (creating baseline and change features); 2. Preprocessing (using GroupShuffleSplit to split training and test sets to prevent data leakage, filtering for multicollinearity, SMOTE oversampling to handle imbalance, standardization, and polynomial features); 3. Exploratory Data Analysis (EDA); 4. Model training and evaluation (comparing algorithms like SVM, logistic regression, and decision trees).

EDA reveals the correlation patterns between disease severity and speech features (such as baseline distribution, degradation trends, and feature correlation heatmaps). Model evaluation uses accuracy, F1 score (macro average), ROC-AUC, and confusion matrix to comprehensively measure the performance of various algorithms and ensure the effectiveness of the solution.

Potential applications of the project include home monitoring apps (regular speech assessment of the disease), early warning systems (alerting for medical attention when abnormal trends are detected), treatment effect evaluation (tracking changes after intervention), and scientific research data collection (standardized solutions), providing a low-cost, non-invasive solution for telemedicine.

The current solution has limitations: speech features are easily affected by colds, emotions, and environmental noise; the prediction ability of a single modality is limited and needs to be combined with multi-modal data such as gait and tremor; the robustness of the model across different devices/environments needs to be verified. Future optimizations can be made in these directions.

This project demonstrates the practical value of machine learning in the healthcare field. By combining speech biomarkers with a rigorous pipeline, it provides a feasible solution for remote monitoring of Parkinson's disease, which is of reference significance to medical AI practitioners and health technology developers.