Bridging the Gap between Deep Learning and Hypothesis-Driven Analysis via Permutation Testing

Area of Science:

• Neuroscience
• Psychiatry
• Machine Learning

Background:

• Traditional neuroscience research often tests hypotheses using neuropsychological and behavioral measures.
• Deep learning offers a data-driven approach to predict outcomes and identify influential factors.
• Current deep learning methods lack statistical significance, limiting their impact on hypothesis testing.

Purpose of the Study:

• To integrate hypothesis testing into deep learning analyses for neuroscience research.
• To develop a flexible and scalable approach combining deep learning with permutation testing.
• To identify statistically significant risk factors associated with depression symptoms in adolescents.

Main Methods:

• Developed a novel approach integrating permutation testing with deep learning.
• Applied the method to yearly self-reported data from 621 adolescents in the National Consortium of Alcohol and Neurodevelopment in Adolescence (NCANDA) study.
• Used the approach to predict negative valence, a symptom of major depressive disorder (MDD) per NIMH Research Domain Criteria (RDoC).

Main Results:

• The integrated approach successfully predicted negative valence in adolescents.
• Identified categories of risk factors that significantly explain the depression symptom.
• Demonstrated the utility of linking deep learning predictions with statistical hypothesis testing.

Conclusions:

• The proposed method offers a statistically rigorous way to apply deep learning in neuroscience.
• This approach enhances the interpretability and impact of data-driven findings in understanding mental health conditions.
• Facilitates the identification of specific risk factors contributing to adolescent depression.