Integrating Genetic, Environmental, Cognitive, and Temperament Data for ADHD Prediction in Explainable Deep Learning...

Area of Science:

• Neuroscience
• Computational Psychiatry
• Genetics

Background:

• Attention-deficit/hyperactivity disorder (ADHD) is a complex neurodevelopmental disorder with significant clinical and etiological heterogeneity.
• Accurate ADHD diagnosis can be challenging, often requiring integration of various data sources for improved classification.
• Existing diagnostic approaches may benefit from advanced computational methods to synthesize multi-modal data.

Purpose of the Study:

• To develop and evaluate an explainable deep learning model for ADHD classification.
• To determine if integrating genetic, environmental, cognitive, demographic, and temperament data can improve ADHD diagnosis.
• To identify key features contributing to ADHD classification and individual predictions.

Main Methods:

• Utilized the Oregon ADHD-1000 cohort, split into training, validation, and test sets.
• Trained modular neural networks using diverse data including genotype, polygenic scores, demographics, environmental factors, cognitive measures, and temperament scores.
• Employed integrated gradients for quantifying feature importance at multiple levels and evaluated model performance using AUC, precision-recall, and calibration analyses.

Main Results:

• The best model incorporating temperament features achieved an AUC of 0.97, demonstrating high accuracy, sensitivity, and specificity.
• Models excluding temperament features showed a lower AUC of 0.75, underscoring the importance of temperament.
• Feature importance analysis highlighted temperament, demographic, and cognitive domains, with individualized explanations revealing domain-specific prediction drivers.

Conclusions:

• Explainable, multi-modal classification models effectively integrate heterogeneous ADHD-relevant information.
• These models can identify critical features for individual ADHD predictions, supporting clinical decision-making.
• This approach holds promise for advancing ADHD risk modeling and aiding clinicians in complex diagnostic cases.