Abstract
Identifying reliable links between individual differences in neurobiological features and differences in symptom profiles or treatment outcomes is a primary goal of precision psychiatry. In this context, brain-behavior predictive modeling has emerged as a powerful approach for elucidating the neural mechanisms underlying both basic cognitive functions and complex clinical phenomena. However, the widespread adoption of these methods in clinical settings is often hindered by the limited amount of neuroimaging and clinical data available for individual patient populations. Transfer learning-a widely adopted strategy in machine learning and deep learning that extracts generalizable and transferable associations from complex, high-dimensional datasets-offers a promising solution. By leveraging large-scale neuroimaging datasets from consortia, transfer learning enables the fine-tuning of models to generate accurate predictions in smaller, clinically specific datasets. In this review, we provide a conceptual and practical overview of transfer learning approaches applied to brain-behavior modeling, with a focus on their utility in predicting clinical outcomes. We discuss recent studies demonstrating that models pretrained on large population datasets can be adapted to reliably predict clinical features from previously unseen neuroimaging data, thereby enhancing model generalizability and interpretability. Additionally, we address practical and theoretical considerations for the adoption of these methods, underscoring their potential to advance mechanistic understanding and bolster clinical utility in precision psychiatry.