B0783
Title: Confounder control using semi-structured networks for neuroimaging data
Authors: Manuel Pfeuffer - Humboldt-Universitaet zu Berlin (Germany) [presenting]
Roshan Rane - Charite - Universitaetsmedizin Berlin (Germany)
Kerstin Ritter - Charite - Universitaetsmedizin Berlin (Germany)
Sonja Greven - Humboldt University of Berlin (Germany)
Abstract: Deep neural networks are a promising tool in medicine, especially in the neuroimaging domain. However, they have been criticized for their lack of transparency and bias when dealing with confounding variables such as age, sex, or comorbidities. Various methods have been developed to address these challenges, often resulting in a trade-off between model performance and transparency of predictions. Semi-structured networks are proposed to control for confounding variables and achieve more interpretable results without compromising model performance. Semi-structured networks combine structured regression models with deep neural networks, enabling us to explicitly model the effects of confounders while eliminating their influence from learned network representations. The effectiveness of this removal often depends on the batch size used during training. Various methods are compared for removing confounder effects, and the impact of batch and sample size are assessed on their performance through simulation studies. The first results of semi-structured networks applied to tasks within the neuroimaging domain, such as diagnosing Alzheimer's disease from structural MRI and covariates, where sample and batch sizes are typically small. It is anticipated that semi-structured networks will enhance model transparency by explicitly addressing confounder effects and facilitating unconfounded interpretation of learned network representations using explainable AI techniques.
