Multivariate prediction of long COVID headache in adolescents using gray matter structural MRI features

Abstract

Objective: Headache is among the most frequent symptoms after coronavirus disease 2019 (COVID-19), so-called long COVID syndrome. Although distinct brain changes have been reported in patients with long COVID, such reported brain changes have not been used for predictions and interpretations in a multivariate manner. In this study, we applied machine learning to assess whether individual adolescents with long COVID can be accurately distinguished from those with primary headaches.

Methods: Twenty-three adolescents with long COVID headaches with the persistence of headache for at least 3 months and 23 age- and sex-matched adolescents with primary headaches (migraine, new daily persistent headache, and tension-type headache) were enrolled. Multivoxel pattern analysis (MVPA) was applied for disorder-specific predictions of headache etiology based on individual brain structural MRI. In addition, connectome-based predictive modeling (CPM) was also performed using a structural covariance network.

Results: MVPA correctly classified long COVID patients from primary headache patients, with an area under the curve of 0.73 (accuracy = 63.4%; permutation p = 0.001). The discriminating GM patterns exhibited lower classification weights for long COVID in the orbitofrontal and medial temporal lobes. The CPM using the structural covariance network achieved an area under the curve of 0.81 (accuracy = 69.5%; permutation p = 0.005). The edges that classified long COVID patients from primary headache were mainly comprising thalamic connections.

Conclusion: The results suggest the potential value of structural MRI-based features for classifying long COVID headaches from primary headaches. The identified features suggest that the distinct gray matter changes in the orbitofrontal and medial temporal lobes occurring after COVID, as well as altered thalamic connectivity, are predictive of headache etiology.

Keywords: adolescents; connectome-based predictive modeling; long COVID headache; multivoxel pattern analysis; structural MRI.

FIGURE 1

Patient selection procedure for multivariate classification of long COVID headache.

FIGURE 2

Schematic diagram for multivariate classification of long COVID headache. Two multivariate classifications were conducted using the gray matter volume (A,B) and the structural covariance matrix (C,D).

FIGURE 3

(A) Shows receiver operating characteristic (ROC) curves and (B) shows confusion matrix generated for classification of long COVID headache from primary headache (cyan line, AUC = 0.73, permutation p = 0.001, accuracy = 63.4%; CH class accuracy = 69.5%, PH class accuracy = 56.5%) based on MVPA using SVM. The shaded area represents the 95% confidence interval. (C) Shows a non-thresholded multivariate (SVM) weight map overlaid on a T1-weighted MRI image (raw image available at https://neurovault.org/images/795018/). The colors represent relative positive weight distributions (orange) and negative weight distributions (cyan). Note that the gray matter volume decrease in the orbitofrontal (ellipse) and parahippocampal (arrows) areas classifies long COVID headaches.

FIGURE 4

(A) Shows receiver operating characteristic (ROC) curves and (B) shows confusion matrix generated for classification of long COVID headache from primary headache using a structural covariance matrix (orange line, AUC = 0.81, permutation p = 0.005, accuracy = 69.5%; CH class accuracy = 73.9%, PH class accuracy = 65.2%) based on CPM-SVM. The shaded area represents the 95% confidence interval. (C) Shows the identified “consensus edges”; the edges colored in red and blue are those increased and decreased in long COVID headache, respectively.