Neuro-environmental interactions: a time sensitive matter

Abstract

Introduction: The assessment of resting state (rs) neurophysiological dynamics relies on the control of sensory, perceptual, and behavioral environments to minimize variability and rule-out confounding sources of activation during testing conditions. Here, we investigated how temporally-distal environmental inputs, specifically metal exposures experienced up to several months prior to scanning, affect functional dynamics measured using rs functional magnetic resonance imaging (rs-fMRI).

Methods: We implemented an interpretable XGBoost-shapley additive explanation (SHAP) model that integrated information from multiple exposure biomarkers to predict rs dynamics in typically developing adolescents. In 124 participants (53% females, ages, 13-25 years) enrolled in the public health impact of metals exposure (PHIME) study, we measured concentrations of six metals (manganese, lead, chromium, copper, nickel, and zinc) in biological matrices (saliva, hair, fingernails, toenails, blood, and urine) and acquired rs-fMRI scans. Using graph theory metrics, we computed global efficiency (GE) in 111 brain areas (Harvard Oxford atlas). We used a predictive model based on ensemble gradient boosting to predict GE from metal biomarkers, adjusting for age and biological sex.

Results: Model performance was evaluated by comparing predicted versus measured GE. SHAP scores were used to evaluate feature importance. Measured versus predicted rs dynamics from our model utilizing chemical exposures as inputs were significantly correlated (p < 0.001, r = 0.36). Lead, chromium, and copper contributed most to the prediction of GE metrics.

Discussion: Our results indicate that a significant component of rs dynamics, comprising approximately 13% of observed variability in GE, is driven by recent metal exposures. These findings emphasize the need to estimate and control for the influence of past and current chemical exposures in the assessment and analysis of rs functional connectivity.

Keywords: XGB classifier; exposome analysis; fMRI; machine learning; resting state.

Figure 1

Overview of ML predictive framework. (A) Resting-state fMRI data were processed, and the averaged time-series were extracted using the Harvard-Oxford atlas. Then, the global efficiency (GE) metric was computed for each participant. Small solid gray circles represent nodes of the graphs (brain regions), while gray connecting lines are the edges of the graph (functional connections). Larger dotted circles represent segregated sub-graphs/networks at the whole brain level. For the exposure, six biological samples (blood, saliva, hair, urine, fingernails, and toenails) were collected and processed for six metal concentrations (manganese, lead, chromium, copper, nickel and zinc). (B) XBoost model was used to predict the GE metric using all exposure biomarkers data (“features”). (C) For model training, 500 iterations of leave-one-out cross validation were used, and all features were utilized in the model training. (D) The performance of the XGBoost model was evaluated by computing the correlation between predicted GE metric versus scaled GE metric (adjusted for age and biological sex) in the hold out subjects obtaining a p < 0.001, r = 0.36 and an explained variance (VE) of 13% (Panel a). Then, SHAP scores were computed for all features (metals exposures) used in the model and the average absolute SHAP value was used to quantify the feature importance. In Panel b, each bar shows the mean absolute SHAP value of each feature, sorted in decreasing order. The most impactful features are displayed and higher SHAP score indicates a more significant contribution in the model prediction, in this case in the prediction of GE metric. Panel c shows the beeswarm plot of SHAP values distribution for the highest ranking features of our model. Feature values associated with a single GE prediction are color-coded, yellow/purple corresponding to low/high metals exposure values, respectively. On the x-axis, the SHAP values are shown representing the impact of a feature with respect to the prediction of the GE metric. Features are sorted using the mean absolute SHAP value in descending order with most important features at the top. Each dot corresponds to one subject. Plots are based on the XBoost model with all features included and leave-one-out-cross validation. BOLD: blood oxygen-level dependent. Features abbreviations: the first letter represents the medium (S, saliva; B, blood; U, urine; H, hair; F, fingernails; T, toenails) and the second and third letters represent the metals (Mn, manganese; Pb, lead; Cr, chromium; Cu, copper; Ni, nickel; Zn, zinc).