Patient-specific game-based transfer method for Parkinson's disease severity prediction

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Data Science

Background:

• Dysphonia is an early indicator of Parkinson's disease (PD), necessitating accurate severity prediction.
• Current PD prediction models often overlook patient heterogeneity, leading to suboptimal performance.
• Small sample sizes in patient-specific models hinder generalization, highlighting the need for effective data transfer techniques.

Purpose of the Study:

• To develop a patient-specific game-based transfer (PSGT) method for improved Parkinson's disease severity prediction.
• To address the challenge of small sample sizes in personalized PD prediction models.
• To enhance the interpretability and effectiveness of PD severity prediction through instance transfer.

Main Methods:

• A selection mechanism identifies similar PD patients from a source domain to mitigate negative transfer risks.
• Shapley values are employed to evaluate the contribution of transferred data, enhancing model interpretability.
• A game-based transfer approach refines instance selection based on contribution and relevance to the target patient.
• Transferred instances are integrated into a random forest model for enhanced PD severity prediction.

Main Results:

• The PSGT method demonstrated superior performance in predicting motor-UPDRS and total-UPDRS compared to existing methods.
• Achieved mean absolute error of 1.59 (motor-UPDRS) and 1.98 (total-UPDRS).
• Achieved root mean square error of 1.95 (motor-UPDRS) and 2.54 (total-UPDRS).
• Achieved volatility of 1.56 (motor-UPDRS) and 1.94 (total-UPDRS).

Conclusions:

• The PSGT method effectively improves the accuracy and stability of Parkinson's disease severity prediction.
• Patient-specific modeling with instance transfer offers a promising approach for personalized PD monitoring.
• The proposed method enhances interpretability and reduces prediction errors in PD telemonitoring.