Early (3–4 years) Brain Markers of Autism and Application of Machine Learning to Predict Future Diagnosis: A Feasibility Study Using Clinical Data

Background: Most previous brain imaging studies in autism spectrum disorder (ASD) are based on children older than 6 years, well after the median age of ASD diagnosis (~52 months)[1]. The few studies that have used subjects less than age four years have investigated only a small subset of brain features[2]-[5]. As such, there is a significant knowledge gap of early ASD brain alterations. A comprehensive investigation of early brain markers is critical to identify the neuroanatomical underpinnings of ASD and may ultimately aid earlier diagnosis. Carrying out a prospective study to characterize early brain markers of ASD would be inordinately expensive. However, clinical brain images of patients who are later diagnosed with ASD are available in hospitals.

Objectives: We aim to identify pre-diagnosis brain alterations of ASD from images acquired as part of clinical care. We apply machine learning on brain morphometric features to predict future ASD diagnosis.

Methods: Brain MRI of 15 ASD male subjects of age three to four years and 18 age- and sex-matched non-ASD subjects were obtained from Geisinger Health System. ASD diagnosis status was based on ICD codes found on electronic health records. A comprehensive set of 687 brain morphometric features were extracted using Freesurfer[6]. ASD versus non-ASD prediction was performed using Random Forest[7]. Prediction accuracy and feature importance were evaluated using 5-fold cross-validation.

Results: Significant (p<0.05, uncorrected) ASD versus non-ASD brain differences are presented in Figure 1. We find that although total intracranial volume (TIV) in ASD was 5.5% larger, volumes of many brain areas (as a percentage of TIV) were smaller in ASD and can be partly attributed to larger (>10%) ventricles in ASD. TIV in ASD was correlated to surface area and cortical folding but not to cortical thickness. The correlation between total ventricular CSF and average cortical folding was 0.56 (p=0.02) for ASD, but was -0.13 (p=0.6) for non-ASD. Folding indices of 58 out of 68 cortices were higher in ASD in the frontal, temporal, cingulate, postcentral, and precuneus regions. White matter regions in ASD had less image intensity (predominantly in the frontal and temporal regions) suggesting myelination deficit.

We achieved 95% AUC for ASD vs. non-ASD prediction using all brain features. When prediction was performed separately for each brain feature type, image intensity yielded the highest predictive power (95% AUC), followed by cortical folding index (69%). The important prediction features for each feature types that yielded high AUCs are presented in Figure 2. The most important feature for prediction was white matter intensity surrounding the rostral middle frontal gyrus and was lower in ASD.

Conclusions: In addition to replicating previous findings, we report several novel brain morphometry differences in early ASD. The high degree of prediction success indicates that the application of machine learning methods on brain features holds promise for earlier identification of ASD, but this pilot study result needs to be replicated with a larger sample. To our knowledge this is the first study to leverage a clinical imaging archive to investigate early brain markers in ASD.