Rescuing missing data in connectome-based predictive modeling.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Data Science

Background:

• Brain-phenotype prediction models often require large sample sizes.
• Increasing sample sizes exacerbates the issue of missing data.
• Conventional methods like complete-case analysis reduce effective sample size and discard valuable information.

Purpose of the Study:

• To integrate and evaluate imputation methods within the Connectome-based Predictive Modeling (CPM) framework.
• To compare the performance of imputation against complete-case analysis for missing connectome and phenotypic data.
• To assess the utility of imputation accuracy as a guide for method selection.

Main Methods:

• Integrated various imputation techniques into the CPM framework.
• Validated the approach using four large, open-source datasets (HCP, PNC, CNP, HBN).
• Compared imputation strategies against complete-case analysis for missing connectomes and phenotypic measures.

Main Results:

• Imputing missing connectomes significantly improved prediction performance over complete-case analysis.
• Imputation accuracy was a reliable indicator for selecting methods for missing phenotypic data, but not for connectomes.
• In a real-world cognition prediction task, imputation doubled the sample size and increased explained variance by 45%.

Conclusions:

• Imputation is a powerful strategy to address missing data in connectome and phenotypic datasets for predictive modeling.
• Effective imputation methods enhance prediction performance by retaining participants, unlike complete-case analysis.
• This study provides a benchmark for imputation techniques in neuroimaging predictive modeling.