Bridging the Brain‐Age Gap: Quantifying regional contributions using Shapley‐Owen values

Abstract

AbstractBackgroundThe concept of brain age gap (BAG) is increasingly being used to express the deviation of brain from the normative trajectory. However, each brain region is subject to independent aging transformations due to differences in neurobiological processes. So, reducing high‐dimensional imaging features down to a single number for simultaneously evaluating brain health, capacity, and residual lifespan is an error‐prone task. Furthermore, brain age is observed to be correlated with chronological age. Therefore, there is need for a systematic empirical evaluation of BAG to understand contributions of different regions when considering BAG as a potential aging biomarker for age‐related neuroanatomical abnormalities. Although methods such as saliency maps have been proposed, there is need for a model‐agnostic technique that quantifies consistent feature contributions across various predictive models.MethodWe used morphological features estimated by Freesurfer from T1w (OASIS3 dataset) to train predictive models for BAG. The bias‐correction was performed using two strategies (i.e., scale‐down and scale up) on each of linear regression and XGBoost models as shown in Figure 1. We present feature importance for top‐10 features (based on median feature importance scores) across 100 bootstrap runs (Figures 2 and 3 respectively), along with correlations with chronological age. We use Shapley‐Owen feature attribution technique as a way of assigning importance to features.ResultAs shown in Figures 2 and 3, we found sub‐cortical, anterior cingulate, and ventricle volumes among the most relevant features. We also observe that SHAP determines an almost identical set of most relevant features for predicting corrected BAG irrespective of the predictive model (linear regression or XGboost) or the bias‐correction approach. Furthermore, we found the results to be consistent with prior literature e.g., subcortical brain volumes esp. gray matter. We also empirically observe that bias‐correction mitigates dependence on chronological age.ConclusionWe present a systematic empirical evaluation to identify the most relevant morphological features for predicting corrected BAG using SHAP. A consistent overlap of most relevant features irrespective of predictive model and the bias‐correction approach provides evidence for the robustness of using Shapley‐Owen values in explaining the contribution of morphological features for predicting corrected BAG.