# Machine Learning‐Based Identification of Children With Intermittent Exotropia Using Multiple Resting‐State Functional Magnetic Resonance Imaging Features

## TL;DR

This study uses machine learning and brain scans to identify children with intermittent exotropia, a type of eye misalignment, with high accuracy.

## Contribution

The study introduces a novel ML-based approach using rs-fMRI parameters, particularly slow-5 fALFF, to distinguish children with intermittent exotropia from healthy controls.

## Key findings

- The slow-5 fALFF parameter achieved the best classification performance with an AUC of 0.957 in training data.
- Brain regions related to eye movement and cognition showed the most discriminative power in distinguishing IXT children from HCs.

## Abstract

To investigate the performance of machine learning (ML) methods based on resting‐state functional magnetic resonance imaging (rs‐fMRI) parameters in distinguishing children with intermittent exotropia (IXT) from healthy controls (HCs).

Forty‐one IXT children and 36 HCs were recruited. The amplitude of low‐frequency fluctuations (ALFF), fractional ALFF (fALFF) in the slow‐4 and slow‐5 bands, and regional homogeneity (ReHo) were calculated. The 360 cortical areas of the Human Connectome Project multimodal parcellation atlas (HCP‐MMP 1.0 atlas) were chosen as 360 regions of interest (ROIs). Each rs‐fMRI parameter value of one ROI was taken as a feature. The Pearson correlation coefficient (PCC) was performed to reduce dimensions. We used four feature selection methods and nine classifiers. The ten‐fold cross‐validation was applied to evaluate the results.

The ML methods combined with rs‐fMRI parameters had good classification performance in distinguishing IXT children from HCs, with the slow‐5 fALFF parameter showing the best classification performance. The linear regression (LR) classifier with analysis of variance (ANOVA) feature selection achieved the highest area under the receiver operator characteristic curve values (0.957, 0.804, and 0.818 for the training, validation, and test datasets, respectively) using five features, including the slow‐5 fALFF values of the right inferior parietal gyrus (IPG), right supplementary motor area (SMA), left primary somatosensory complex, right frontal opercula, and left dorsolateral prefrontal cortex (DLPFC), and the accuracy, sensitivity, and specificity values were 0.759, 0.759, and 0.760, respectively. The brain regions showing the greatest discriminative power included right IPG, right SMA, left primary somatosensory complex, right frontal opercula, left DLPFC, right posterior orbitofrontal cortex (pOFC), left medial superior temporal (MST), left parieto‐occipital sulcus (POS), and right anterior ventral insula.

Based on the slow‐5 fALFF values of the five cortices as the features, LR with ANOVA was the best ML model for distinguishing between IXT children and HCs. The result indicates the slow‐5 fALFF parameter has the potential to serve as a biomarker for distinguishing IXT children from HCs. In addition, brain regions related to stereopsis, eye movement, and higher‐order cognitive functions play an important role in the neuropathologic mechanisms underlying IXT.

The machine learning methods combined with resting‐state functional magnetic resonance imaging parameters (ALFF, slow‐5 fALFF, and ReHo) had good classification performance in distinguishing intermittent exotropia (IXT) children from healthy controls (HCs). Among several parameters, the slow‐5 fALFF showed the best classification performance. The slow‐5 fALFF can be a promising biomarker for distinguishing IXT children from HCs.