DeepRepViz: Identifying Confounders in Deep Learning Model Predictions

TL;DR

The DeepRepViz framework provides a systematic approach to test for potential confounders such as age, sex, and imaging artifacts and improves the transparency of DL models for neuroimaging studies.

Abstract

Deep Learning (DL) models have gained popularity in neuroimaging studies for predicting psychological behaviors, cognitive traits, and brain pathologies. However, these models can be biased by confounders such as age, sex, or imaging artifacts from the acquisition process. To address this, we introduce 'DeepRepViz', a two-part framework designed to identify confounders in DL model predictions. The first component is a visualization tool that can be used to qualitatively examine the final latent representation of the DL model. The second component is a metric called 'Con-score' that quantifies the confounder risk associated with a variable, using the final latent representation of the DL model. We demonstrate the effectiveness of the Con-score using a simple simulated setup by iteratively altering the strength of a simulated confounder and observing the corresponding change in the Con-score. Next, we validate the DeepRepViz framework on a large-scale neuroimaging dataset (n=12000) by performing three MRI-phenotype prediction tasks that include (a) predicting chronic alcohol users, (b) classifying participant sex, and (c) predicting performance speed on a cognitive task called 'trail making'. DeepRepViz identifies sex as a significant confounder in the DL model predicting chronic alcohol users (Con-score=0.35) and age as a confounder in the model predicting cognitive task performance (Con-score=0.3). In conclusion, the DeepRepViz framework provides a systematic approach to test for potential confounders such as age, sex, and imaging artifacts and improves the transparency of DL models for neuroimaging studies.

Figures (2)

• Figure 1: A demonstration of DeepRepViz and the Con-score metric: Figure (a) shows a DL model that classifies chronic alcohol users from non-users using the participants' structural MRI data. Con-score is computed for variables in the data using the representation learned in the penultimate layer $H^{(l-1)}$ of the DL model. Con-score is highest for 'sex' when classifying chronic alcohol users. Figure(b) shows the DeepRepViz tool and how it can be used to inspect the learned representation $H^{(l-1)}$. When we select the predicted label $\hat{y}$ in the tool, we can see the linear decision boundary of the model in $H^{(l-1)}$. This decision boundary aligns with the representation of sex in $H^{(l-1)}$ in Figure(a). This is also reflected in the Con-score. This implies that the model could be using the information about the participant's sex encoded in the MRI data as a proxy to identify chronic alcohol users.
• Figure 2: Results of applying Con-score metric to (a) simulated dataset and (b) UK Biobank neuroimaging dataset: (a) shows Con-scores obtained on a simulated binary classification task for 8 different levels of correlations between a confounder $c$, a binary label $y$, and the input features $H=\{h_0, h_1\}$. (b) shows the Con-scores obtained for eight potential confounder variables (see legend) in three brain-phenotype prediction tasks on the UK Biobank dataset.