EEG Analytics for Early Detection of Autism Spectrum Disorder: A data-driven approach

Abstract

Autism spectrum disorder (ASD) is a complex and heterogeneous disorder, diagnosed on the basis of behavioral symptoms during the second year of life or later. Finding scalable biomarkers for early detection is challenging because of the variability in presentation of the disorder and the need for simple measurements that could be implemented routinely during well-baby checkups. EEG is a relatively easy-to-use, low cost brain measurement tool that is being increasingly explored as a potential clinical tool for monitoring atypical brain development. EEG measurements were collected from 99 infants with an older sibling diagnosed with ASD, and 89 low risk controls, beginning at 3 months of age and continuing until 36 months of age. Nonlinear features were computed from EEG signals and used as input to statistical learning methods. Prediction of the clinical diagnostic outcome of ASD or not ASD was highly accurate when using EEG measurements from as early as 3 months of age. Specificity, sensitivity and PPV were high, exceeding 95% at some ages. Prediction of ADOS calibrated severity scores for all infants in the study using only EEG data taken as early as 3 months of age was strongly correlated with the actual measured scores. This suggests that useful digital biomarkers might be extracted from EEG measurements.

Figure 1

Time series and complex networks are related, and methods for reconstructing essential elements of one from the other have been developed. See, for example.

Figure 2

A schematic representation of classification with a support vector machine (SVM) method is shown with 3 dimensions representing 3 features. The training set is used to create a model or separating hyperplane in the feature space. Axes of the feature space are the nonlinear EEG features. Test subjects are then classified by determining which side of the plane the subject’s features places them. ASD and LRC- subjects are used for the training set. Leave-one-out cross validation leaves out a single subject from the training set and then makes a prediction for the left-out subject. The distance of a subject from the hyperplane can be used to estimate severity. This distance was scaled to approximate the Calibrated Severity Score (CSS), where 1.0 is the lowest score for children with no autism symptoms and 10 is the highest score for the most severe autism cases.

Figure 3

The coarse-graining procedure introduced by Costa et al. is illustrated. For powers of 2, the resulting scaled time series are identical to Haar wavelet transform approximations.

Figure 4

Measured CSS scores for each outcome group (solid lines), along with predicted scores derived from EEG features at each age are shown (dashed lines). Confidence intervals are shaded for predicted values. The age refers to the age at which EEG data was collected and used for the prediction. Correlation coefficients between predicted scores and the measured CSS values are shown by Xs if the axis values are multiplied by 10⁻¹.

Figure 5

Differences between ASD and LRC− group values for SampE, DFA, and DET are shown using color for every nonlinear value. The vertical axis for each subplot represents six frequency bands, from low (delta) to high (gamma+), as defined in Table 2. The horizontal axis for each subplot is the scalp location. The axis labels are shown in a single large horizontal label across the bottom of the plot. The left or right side of each subplot corresponds to left or right sensors, respectively. Centrally located sensor values are in the center of the subplots. Red colors indicate that the values for this feature are higher in the ASD group. Blue colors indicate that ASD values are lower for that feature. Color saturation (darker red or darker blue) is correlated to the significance of the group differences. The scale on the right of the figure shows p-values of the color saturation. White areas with washed-out color indicate p-values close to 1.0, hence no significant differences.

Figure 6

Similar to Fig. 5: differences between ASD and LRC− group values for RR, L_max, and TT are shown using color for every nonlinear value.

Figure 7

Similar to Fig. 5: differences between ASD and LRC− group values for LAM, L_mean, and L_entr are shown using color for every nonlinear value.

Figure 8

Developmental trajectories for SampE in the left temporal region (T7 sensor) in higher frequencies (beta + gamma) for ASD, LRC−, and HRA−.

Figure 9

Developmental trajectories for SampE in the right temporal-parietal region (T8 + P4 + P8 sensors) in frequencies theta through gamma for ASD, LRC−, and HRA−.

Figure 10

Developmental trajectories for DET in the left lateral-frontal region (F7 sensor) in frequencies theta through gamma for ASD, LRC−, and HRA−.

Figure 11

Developmental trajectories for DET in the entire frontal region (Fp1 + F7 + Fz + F8 + Fp2 sensors) in higher frequencies (beta + gamma) for ASD, LRC−, and HRA−.

Figure 12

Developmental trajectories of SampE for the posterior region (O1 + O2 sensors) in low frequency (delta) for ASD, LRC−, and HRA−.