fMRI-based machine learning analysis of neural 1 substrates of pediatric anxiety: Temporal Pole 2 and emotional face-responses.

15 A prominent cognitive aspect of anxiety is dysregulation of emotional 16 interpretation of facial expressions, associated with neural activity from the 17 amygdala and prefrontal cortex. We report machine learning analysis of 18 fMRI results supporting a key role for a third area, the temporal pole (TP) 19 for childhood anxiety in this context. This ﬁnding is based on diﬀerential 20 fMRI responses to emotional faces ( e.g. , angry versus fearful faces) in 21 children with one or more of generalized anxiety, separation anxiety, and 22 social phobia (n = 22) compared with matched controls (n = 23). In 23 our machine learning model, the right TP distinguished anxious from 24 control children (accuracy = 81%). Involvement of the TP as signiﬁcant 25 for neurocognitive aspects of pediatric anxiety is a novel ﬁnding worthy of 26 further investigation. 27 study to use a data-driven approach to classify versus non-anxious children using emotional facial stimuli. Here, we used a non-linear super learner (AdaBoost with logistic regression as a base estimator) to target the Talairach regions that could best distinguish anxious from non-anxious children. Our 423 model achieved an accuracy above 81% for this task. Subsequently, we examined how 424 diﬀerent negative emotional faces would be processed in both groups. We found that 425 fear and angry faces could clearly be distinguished in the TP, but only after functional 426 alignment was applied to the brain scans of all subjects. This study illustrates the power 427 of task-based fMRI designs to predict disease states and stimulus conditions. It also 428 indicates that the TP is a region that should be further examined in pediatric anxiety. may We have demonstrated that learning of used


30
Clinical anxiety is associated with inability to control or autoregulate one's autonomic . . (959 regions) .

Talairach region parcellation
. .  Fig. 1: Processing pipeline for the primary analysis. Voxels from 959 Talairach regions were individually trained on a super learner (SL) classifier to determine which regions could best distinguish between anxious and non-anxious children. The SL included a nested cross-validation process to hypertune parameters for an AdaBoost model. The region with the highest average accuracy was selected for our analysis. Note: each time produced its own prediction; we labeled each person with the majority vote over the time points for that subject. . .

Region selection
Highest acc % region Preprocessing pipeline for negative stimuli analysis. Voxels from the selected region (from the primary analysis) were used to predict the stimulus label for each timepoint (fear versus anger). We used a probabilistic shared response model (SRM) to transform all functional images into a shared common space. Thereafter, a linear SVM was trained on the functionally aligned data, which were used to predict a facial stimulus for each time point, for each subject. Model metrics include mean accuracy and standard error from 5-fold CV.
key aspects for altered functional connectivity in anxious children in this context. Using a publicly available dataset  rectly classified participants), precision, sensitivity (i.e., recall), and F1-score. Using 162 our SL, we achieved an accuracy of 81% an overall precision of 80% as seen in Figure   163 4A, recall at 80% and an F1-score of 80%. Here, we conducted a high level, between-group ROI analysis for region #41 to examine 167 activation differences between anxious and non-anxious children. Using the beta values 168 from our second level, grouped Bayesian representational similarity analysis (GBRSA), 169 we compared activation in this region by using a mask to confine our analysis. Figure   170 3 (left) shows the region-based beta values for both anxious and non-anxious children.

171
We compared all 685 pairs of voxels in this region using a Mann Whitney U-test 172 (two-tailed). This statistical test revealed a significant difference in voxels for region 173 #41 (U = 215017.00, p < 0.005).

174
Fully connected network with region #41 175 We examined the most correlated regions (top 30%) with our ROI using absolute 176 Pearson correlations for anxious and non-anxious children. Twenty-six brain regions 177 were correlated with Talairach region #41 for anxious children but only 16 brain 178 regions for non-anxious children as seen in Figure 4 (right side) and  and an F1 score of 42.7% as seen in Table S2. Figure   were also 73%. Non-anxious children viewing angry faces revealed the highest recall at 208 76%, and non-anxious children viewing fearful faces revealed the highest precision at 209 75% as seen in Figure 4C and Table S3. No significant differences were found between 210 the 4 classes with respect to their precision, recall, or F1 scores.  children from non-anxious children based on their brain scans as they view negative 217 facial stimuli. We used machine learning to demonstrate that task-based fMRI activity 218 related to this anatomical area is sufficient to achieve a relatively high accuracy. In 219 our primary analysis, we trained a super learner for each individual Talairach brain 220 region. Region #41 (the temporal pole) returned the highest mean accuracy with 81%. we achieved an accuracy of 73%, suggesting that we can accurately identify both neural 239 signatures of anxious children and how they process fear and angry faces.

240
Due to the diverse structure and connectivity to a number of regions, the 241 putative role of the TP has been inconsistent and subject to significant debate 10,34,35 .

242
The TP has been proposed as a social-emotional cognition hub that receives various processed in the brain of anxious and non-anxious children. We may extend our study stimuli.

419
In summary, the goal of this study was to use a data-driven approach to classify anxious The authors declared no competing interests.

442
Data as seen in Table 1. Impairment and emotional symptoms were recorded prior to the 498 start of the fMRI study and were representative of psychiatric symptoms that interfere 499 with daily functioning. Impairment scores were assessed using the World Health (1/60 Hz) 63 . Task blocks were removed from analysis if two volumes were removed 555 from the start of the block or more than 3 volumes in total were removed from the 556 block. Additionally, the entire run was excluded from subsequent analyses if more 557 than one block of emotional stimuli was removed 63 . Next was spatial smoothing.

558
Spatial smoothing is a method used to increase the signal to noise ratio in fMRI brain Using the Talairach atlas (with 2mm voxel size), we segmented the brain into 959 569 regions, then applied a super learner (SL) to examine which areas were closely related 570 to our diagnostic labels. This was a brain-wide regional analysis to test our hypothesis 571 against other brain regions. The SL would make a prediction on every time point      Table S3: Four-class ensemble classification results. This final classification model coupled the predictions from our group and negative stimuli classification models into a four-class performance task. Our models achieved a balanced accuracy of 73% compared to baseline which was 26%.

Classifier Class Precision Recall F1 Score Accuracy
Four-class ensemble model Non-anxious angry Non-anxious fear Anxious angry Anxious fear