Performance reserves in brain-imaging-based phenotype prediction.

Area of Science:

• Neuroimaging
• Machine Learning
• Cognitive Neuroscience

Background:

• Machine learning models are increasingly used to predict cognitive and mental health phenotypes from brain imaging data.
• The impact of sample size on the predictive performance of these models is not fully understood.
• Optimizing prediction accuracy is crucial for the clinical translation of neuroimaging findings.

Purpose of the Study:

• To investigate the effect of sample size on the accuracy of predicting cognitive and mental health phenotypes using brain imaging and machine learning.
• To evaluate the contribution of integrating multiple imaging modalities to prediction performance.
• To assess the practical and clinical utility of large-scale neuroimaging datasets for phenotype prediction.

Main Methods:

• Utilized machine learning algorithms to predict phenotypes from brain imaging data across a range of sample sizes (1,000 to 1 million participants).
• Compared prediction performance using single versus multiple imaging modalities.
• Analyzed the relationship between sample size and the informativeness of different imaging modalities.

Main Results:

• Prediction performance improved 3- to 9-fold with sample size increases from 1,000 to 1 million participants.
• Integrating multiple imaging modalities significantly boosted prediction accuracy, comparable to doubling the sample size.
• Despite improvements, prediction accuracy remained low, indicating substantial room for performance gains.
• The most informative imaging modality varied with increasing sample size.

Conclusions:

• While larger sample sizes enhance prediction accuracy in neuroimaging, current performance levels are insufficient for widespread clinical application.
• Combining multiple imaging modalities offers a substantial advantage, but achieving clinically relevant prediction may require impractically large datasets.
• Future research should focus on optimizing data acquisition, feature selection, and model development to improve the utility of machine learning in neuroimaging.