Continuous, Learned Imputation of Missing Values in Parkinson's Disease.

Area of Science:

• Neurology
• Machine Learning
• Data Science

Background:

• Accurate clinical scores are crucial for Parkinson's disease (PD) management.
• Missing data in clinical assessments poses a significant challenge to reliable PD scoring.
• Existing imputation methods may not generalize well across diverse patient populations.

Purpose of the Study:

• To evaluate the efficacy of self-supervised learning (SSL) for imputing missing clinical scores in Parkinson's disease.
• To compare the generalization capabilities of SSL against established imputation techniques.

Main Methods:

• Utilized self-supervised learning models for data imputation.
• Compared SSL performance against Multiple Imputation by Chained Equations (MICE), MissForest, and Multiple Imputation With Empirical Weights (MIWAE).
• Assessed generalization across different patient populations.

Main Results:

• Self-supervised learning demonstrated superior generalization capabilities compared to MIWAE, MissForest, and MICE.
• SSL effectively handles missing data, leading to more robust clinical scores.
• The method can be integrated during data collection for enhanced data integrity.

Conclusions:

• Self-supervised learning offers a powerful and generalizable approach to managing missing data in Parkinson's disease clinical assessments.
• This technology enhances the reliability of clinical data collection in real-world settings.
• SSL represents a significant advancement for accurate Parkinson's disease management.