Brain-phenotype predictions can survive across diverse real-world data

Area of Science:

• Neuroscience
• Machine Learning
• Psychiatry

Background:

• Clinical data-based machine learning models for psychiatric treatment prediction often fail with unharmonized samples.
• Neuroimaging offers a potentially more robust alternative due to inherent neurobiological information.
• Existing neuroimaging studies often lack rigorous external validation across diverse, unharmonized datasets.

Approach:

• Evaluated the generalizability of predictive models across three diverse, unharmonized neuroimaging datasets: Philadelphia Neurodevelopmental Cohort (PNC), Healthy Brain Network (HBN), and Human Connectome Project in Development (HCPD).
• Assessed model performance considering substantial inter-dataset heterogeneity in demographics, clinical factors, and neuroimaging acquisition parameters.
• Investigated functional connectivity-based predictive models for their robustness and generalizability.

Key Points:

• Reproducible and generalizable brain-behavior associations were achieved across diverse datasets with hundreds of participants.
• Functional connectivity-based predictive models demonstrated robustness despite significant inter-dataset variability.
• For HCPD and HBN, cross-dataset prediction outperformed within-dataset cross-validation, suggesting benefits of training on diverse data.

Conclusions:

• Neuroimaging predictive models hold potential for robust psychiatric treatment outcome prediction across diverse, real-world clinical settings.
• Rigorous external validation across unharmonized datasets is crucial for evaluating the true generalizability of these models.
• Training on diverse datasets may enhance predictive accuracy in specific scenarios, paving the way for clinical translation.