Detect and correct bias in multi-site neuroimaging datasets.

Area of Science:

• Neuroimaging
• Machine Learning
• Data Science

Background:

• Increasingly large datasets are crucial for training complex machine learning algorithms and enhancing statistical power in neuroimaging association studies.
• Pooling data from independent studies is a common method to increase sample size, but it risks introducing selection, measurement, and confounding biases, leading to spurious correlations.

Purpose of the Study:

• To investigate and quantify dataset bias in large-scale neuroimaging data.
• To develop and evaluate methods for harmonizing neuroimaging datasets to mitigate bias.
• To examine the impact of harmonization on confounding bias and causal relationships.

Main Methods:

• Combined 35,320 magnetic resonance images from 17 independent studies.
• Conducted a 'Name That Dataset' experiment demonstrating significant dataset identifiability (71.5% accuracy).
• Modeled confounders as latent variables and used Kolmogorov complexity to compare confounded versus causal models.
• Extended the ComBat algorithm for dataset harmonization to control for global variations across image features.

Main Results:

• Empirical evidence confirmed the presence of dataset bias, with scans being identifiable by their source study.
• Harmonization methods successfully reduced dataset-specific information in imaging features.
• Confounding bias was demonstrably reduced, and in some cases, transformed into a causal relationship.
• Harmonization requires careful application to avoid removing essential subject-specific information.

Conclusions:

• Dataset bias is a significant concern in pooled neuroimaging studies.
• Proposed harmonization techniques, including an extended ComBat algorithm, can effectively reduce dataset bias and confounding effects.
• While beneficial, harmonization must be applied cautiously to retain valuable individual-level data and avoid over-correction.