Meta-matching as a simple framework to translate phenotypic predictive models from big to small data

Area of Science:

• Neuroscience
• Machine Learning
• Biostatistics

Background:

• Predictive models trained on large datasets often perform poorly on small, independent datasets with novel phenotypes.
• Translating models across different datasets and phenotypes is a significant challenge in biomedical research.

Purpose of the Study:

• To introduce meta-matching, a novel framework for transferring predictive models from large-scale datasets to small-scale studies with unseen phenotypes.
• To enhance prediction accuracy in small studies by leveraging correlations between phenotypes in large and small datasets.

Main Methods:

• Meta-matching framework developed to exploit correlations between related phenotypes in large and small datasets.
• Application of meta-matching to predict non-brain-imaging phenotypes using resting-state functional connectivity data.
• Validation using UK Biobank (N=36,848) and Human Connectome Project (HCP) (N=1,019) datasets.

Main Results:

• Meta-matching significantly improved prediction of new phenotypes in small, independent datasets.
• Translating a UK Biobank model to 100 HCP participants showed an eight-fold improvement in variance explained.
• Average absolute gain in prediction accuracy was 4.0% across 35 phenotypes.

Conclusions:

• Meta-matching offers a powerful approach to boost predictive performance in small-scale studies by leveraging large-scale data.
• The framework demonstrates substantial potential for phenotype prediction across diverse datasets and study sizes.
• Results highlight the growing utility of large-scale datasets for advancing precision medicine through model transfer.