Practical Strategies for Extreme Missing Data Imputation in Dementia Diagnosis

Area of Science:

• Computational biology
• Medical informatics
• Data science

Background:

• Clinical decision support systems rely on complete data, but real-world data often has missing values.
• Extreme data missingness (∼50% missing in >50% features) poses challenges for both training and testing datasets.
• Dementia diagnosis is particularly affected by data incompleteness due to long delays and high variability.

Purpose of the Study:

• To evaluate multiple imputation and classification workflows for handling extreme data missingness.
• To assess the impact of missing data on diagnostic accuracy and computational cost.
• To identify optimal strategies for predictive modeling in dementia diagnosis.

Main Methods:

• Replicated extreme missingness structure on a larger open dataset using real-world memory clinic data.
• Evaluated various data imputation techniques and classification models.
• Compared approaches based on diagnostic accuracy and computational expense.

Main Results:

• Computational cost varied significantly across different imputation and classification methods, while accuracy did not.
• Iterative imputation on training data combined with a reduced-feature classification model proved most efficient.
• This combined approach demonstrated a favorable balance between speed and diagnostic accuracy.

Conclusions:

• Data imputation strategies are crucial for developing robust clinical decision support systems, especially in dementia diagnosis.
• The choice of imputation and classification methods significantly impacts computational cost but not necessarily accuracy.
• Iterative imputation and reduced-feature classification offer a promising solution for handling extreme data missingness in predictive diagnostic models.