Latent Similarity Identifies Important Functional Connections for Phenotype Prediction

Area of Science:

• Neuroimaging
• Machine Learning
• Biomarker Discovery

Background:

• Brain imaging studies often face challenges with small sample sizes and high-dimensional data, limiting the identification of reliable biomarkers like brain age and fluid intelligence.
• Reproducibility in neuroimaging research is hindered by these data limitations.

Purpose of the Study:

• To develop an interpretable, multivariate algorithm (Latent Similarity - LatSim) for classification and regression tasks with small sample sizes and high feature dimensions.
• To improve the identification of neuroimaging biomarkers and understand their predictive power.

Main Methods:

• Latent Similarity (LatSim) algorithm combines metric learning, kernel similarity, and softmax aggregation to identify inter-subject similarities.
• Utilized multi-paradigm fMRI data for three prediction tasks, leveraging computational efficiency for a greedy selection interpretability method.

Main Results:

• LatSim demonstrated significantly higher predictive accuracy on the Philadelphia Neurodevelopmental Cohort (PNC) dataset, especially with small sample sizes.
• Identified 4 functional brain networks enriched in connections crucial for predicting brain age, sex, and intelligence.
• Connections identified by LatSim showed superior discriminative power compared to other methods.

Conclusions:

• A small number of connections (1-5) contain most predictive information for specific tasks.
• The default mode network is consistently over-represented in the top predictive connections across all tasks.
• The proposed LatSim algorithm offers a novel approach for analyzing small-sample, high-dimensional neuroimaging data, advancing both algorithmic design and neuroscience research.