Generalizable and replicable brain-based predictions of cognitive functioning across common psychiatric illness.

Area of Science:

• Neuroscience
• Computational Psychiatry
• Machine Learning

Background:

• Cognitive impairments are transdiagnostic, treatment-resistant, and linked to poor patient outcomes.
• Accurate prediction of cognition typically requires large datasets, posing challenges for clinical research.
• Computational psychiatry aims to link brain function and symptoms through predictive modeling.

Purpose of the Study:

• To develop and validate a transfer learning framework for predicting cognitive functioning.
• To assess the generalizability of predictive models across independent clinical datasets.
• To demonstrate that large-scale population data can enhance prediction in smaller clinical samples.

Main Methods:

• Training a predictive model on functional neuroimaging data from the UK Biobank.
• Utilizing a transfer learning approach to apply models to smaller transdiagnostic samples.
• Evaluating model performance and generalization across independent clinical datasets.

Main Results:

• Achieved prediction performance comparable to larger studies in three independent clinical samples.
• Demonstrated a significant boost in prediction accuracy (up to 116%) compared to classical models.
• Confirmed model generalizability, maintaining performance when trained and tested across different datasets.

Conclusions:

• Transfer learning enables accurate prediction of cognitive functioning in clinical samples using models from large population datasets.
• This approach overcomes sample size limitations in clinical research for cognitive prediction.
• The findings support the utility of leveraging large-scale data for advancing computational psychiatry and clinical insights.