Two neuroanatomical subtypes of males with autism spectrum disorder revealed using semi-supervised machine learning

Publicly-available MRI data and phenotype data were downloaded from the Autism Brain Image Data Exchange database (ABIDE) (http://fcon_1000.projects.nitrc.org/indi/abide/). In order to improve the reliability and repeatability of the results, a strict data exclusion scheme was adopted in this study. Data curation was conducted using the following exclusion criteria: (1) female subjects; (2) left or mixed handedness or subjects with no handedness information; (3) subjects with no full-scale IQ information or full-scale IQ below 80; (4) to reduce ageing effects, subjects older than 45 years of age were excluded; (5) an initial quality check was performed by an experienced radiologist, then the Image Quality Rate (IQR) in CAT12 software (Computational Anatomy Toolbox 12, CAT12, http://​www.​neuro.​uni-jena.​de/​cat/​) was used in the quality control of structural images to avoid the subjectivity of manual check. As low image quality can lead to underestimation of the gray matter in most preprocessings [23], therefore, we adopted strict image quality control procedures, and subjects with IQR lower than B- were excluded (The detailed image quality rating criterion and quality control steps are shown in Additional file 1: Fig. S1 and Supplementary Text. In addition, Additional file 1: Table S1 demonstrates the quality control metrics for the enrolled subjects, no significant differences were found in terms of image quality between health controls and ASD patients). (6) Study sites with less than 20 ASD patients were excluded. In addition, the New York University Medical Center (NYU) sample 2 in ABIDE-II databases were excluded because different scanning parameters compared with NYU sample in ABIDE-I and NYU sample 1 in ABIDE-II. The individuals scanned using a head coil with 32 channels in Kennedy Krieger Institute (KKI) sample of ABIDE-II were excluded because the rest KKI data from ABIDE-I and ABIDE-II were scanned using a head coil with 8 channels. After applying these exclusion criteria, data of 478 subjects from ABIDE-I and ABIDE-II from 7 sites remained. Demographics for the resultant sample (221 ASD patients and 257 healthy controls) are presented in Table 1.

At first, CAT12 was applied for VBM preprocessing to structural MRI data from each subject. The main steps applied to the structural MRI data were as follows: (1) normalization of T1 image into the Montreal Neurological Institute (MNI) space and the voxel size was resampled into 1.5 × 1.5 × 1.5 mm³; (2) segmentation of the normalized images into gray matter, white matter and CSF; (3) modulation to convert the voxel values of tissue concentration (density) to volume; (4) calculation of the volume value based on the ROI from gray matter, white matter and CSF; (5) smoothing with an 8-mm full width at half maximum (FWHM) isotropic Gaussian kernel.

For each subject, we obtained volumes of 142 ROIs (The ROIs were derived from the Neuromorphometrics atlas, http://​www.​neuromorphometri​cs.​com/​, detailed information of the ROIs is listed in Additional file 1: Table S3) from structural MRI data as features for further cluster analysis. Prior to subtyping, manual check was carried out to check the alignment between the atlas and the normalized volumetric maps, and to eliminate any absurd values of volumes (Details are shown in the Supplementary Text). Then the effect of age and site-specific factors on the ROI volumes were estimated using a linear model and were regressed out [17]. The covariate regression strategy was also applied to further voxel-wise volumetric and dynamic R-fMRI measures analysis of ASD subtypes, as well as classification between ASD subtypes and healthy controls.

HYDRA was utilized based on the volumetric measures of the ROIs to identify ASD subtypes [21]. HYDRA consists of the following steps: Firstly, ASD subjects are given negative labels and healthy controls are given positive labels. HYDRA will determine the number of hyperplanes by the K value of the clusters number to generate convex polyhedron for separating ASD patients from the healthy controls. An extending standard linear maximum margin classifiers is introduced to calculate the distance from each ASD subject to each hyperplane, and ASD subjects will be assigned to the hyperplane closest to themselves, so that all ASD subjects are divided into K clusters. Following parameters were used to ensure convergent and stable clustering results, and to alleviate computational burden: 50 iterations between estimating hyperplanes and cluster estimation, 20 clustering consensus steps, regularization parameter of 0.25, 10 cross-validation folds and clustering range from 2 to 8.

In this experiment, the above process runs within the framework of tenfold cross-validation. For each time, 9 folds subjects are selected for the above clustering process. The process of 50 times iterations is adopted to find the optimal convex plane for clustering estimation. Finally, for each ASD subject, it participated in 9 clustering processes and obtained 9 same or different clustering labels. The 20 clustering consensus steps will determine the final cluster label of each ASD by a cooccurrence matrix generated from the labels of 9 clustering processes. Meanwhile, the algorithm quantifies the similarity between clustering results in a 10-folds cross-validated fashion by the adjusted rand index (ARI) [24] to assess the clustering stability. ARI evaluates the contingency of grouping and provides a more conservative overlap estimation. The values of ARI range from 0 to 1, and the ARI value of 1 represents a perfect clustering. A schematic illustration of the HYDRA method is shown in Additional file 1: Fig. S3.

To assess the reproducibility of subtypes, we took a series of analyses, including split-sample tests [25, 26] and leave-one-site-out validation [27]. In order to evaluate the reproducibility of ASD subtypes, we conducted a split-sample test analysis. The healthy controls and patients were randomly divided into two parts and then HYDRA was applied in these two parts, respectively. Age was matched between healthy controls and ASD in the two splits (Additional file 1: Table S4). Voxel-wise volumetric maps were further compared between healthy controls and each ASD subtype in the two splits.

In addition, the subtypes were further validated using a leave-one-site-out strategy. In this strategy, the data of 6 sites were used to train HYDRA models and subtype labels of the last remaining sites were identified by the trained models. This procedure was repeated 7 times to complete possible combinations for all sites. In other words, in each leave-one-site-out process, one site was regarded as an independent site, and the predicted labels of this site was determined by the trained model from the other 6 sites. Finally, the predicted labels of leave-one-site-out strategy from all 7 sites were compared with the original labels obtained by taking all 7 sites together. The voxel-wise gray matter volumetric maps were estimated using leave-one-site-out-predicted results between each subtype and healthy controls.

We conducted voxel-wise volumetric analyses using regionally linear multivariate discriminative statistical mapping (MIDAS) [28] to explore the alterations of gray and white volume in ASD subtypes with age and site as covariates. Compared with other information-mapping methods, MIDAS effectively determines the regionally varying, anisotropic filtering of any image data that optimally captures group differences [28]. Voxel-wise group (ASD subtypes and healthy control) comparisons of the smoothed gray and white matter volume maps were performed using two sample t-tests within SPM12. Similarly, in this step age and site were included in a general linear model as covariates. Then, the t-statistic maps were converted to the effect size (Cohen’s d) maps. Finally, the voxel-wise statistical significance values (p-values) were corrected by false discovery rate (FDR) (FDR-p < 0.05) and were then used as a mask to show the effect size maps between the groups via MIDAS.

First, we verified that subtypes were composed of similar proportions of ASD patients per site based on the chi-square test. To assess whether severity of ASD differed between subtypes, a two-sample t-test was utilized on the full-scale IQ, verbal IQ, performance IQ, Autism Diagnostic Observation Schedule (ADOS) total, social, communication, RRB, severity scores, Autism Diagnostic Interview Revised (ADI-R) Social, Verbal, RRB, Social Responsiveness Scale (SRS) total, awareness, cognition, communication, motivation, and mannerisms respectively. Given a widely hypothesized characteristic on verbal and non-verbal IQ discrepancy in ASD [29, 30], the verbal IQ and performance IQ was included. Besides, within each subtype, the relationship between total intracranial volume and above demographic and clinical data was assessed using Pearson’s correlation. In this procedure, subjects with missing ADOS measures were excluded.

In order to verify whether the two neuroanatomical subtypes of ASD were differed in brain functions, we analyzed R-fMRI data from the same cohorts included in the structural MRI dataset. In this step, subjects with head movements greater 2.0 mm of translation or 2.0 degrees of rotation in any direction were excluded. All dynamic R-fMRI measures were calculated via Data Processing & Analysis for Brain Imaging (DPABI, http://​rfmri.​org/​DPABI). At first, we obtained 4 dynamic R-fMRI measures including regional homogeneity (ReHo), voxel-mirrored homotopic connectivity (VMHC), network degree centrality (DC) and global signal correlation (GSCorr). To characterize the dynamic R-fMRI measures, we computed the standard deviation (SD) map across time windows of each measure for subsequent statistical analysis. Besides, we calculated Kendall’s coefficient of concordance (KCC) for the 4 dynamic R-fMRI measures across time windows as the dynamic volume-wise concordance. The mean and SD of the time series of subjects’ dynamic volume-wise concordance index (one for each subject) were performed using two-sample t-tests, age, site, and head motion (mean framewise-displacement, FD) were treated as covariates by the general linear model. Finally, we calculated the voxel-to-atlas KCC mapping, which characterizes the stability of dynamic functional architecture. The detailed preprocessing and dynamic R-fMRI measures calculation steps are shown in Supplementary Text. Voxel-wise group comparisons (between Subtype 1 and Subtype 2) of the above-mentioned R-fMRI measures mapping were performed using two sample t-tests within SPM 12, age, site, and head motion (mean FD) were included in the general linear model as covariates. Gaussian Random Field (GRF) correction was performed to control false positives (voxel-level p < 0.001, to cluster-level p < 0.01, two-tailed). In addition, we also performed comparisons in dynamic R-fMRI measures between healthy controls and ASD subtypes.

To test whether subtyping could improve diagnostic label accuracy beyond group average comparisons, and to identify possible biomarkers from brain regions with volumetric differences, the support vector machine (SVM) model was applied to classify ASD and healthy controls based on the volumetric measures of the ROIs. Classification was performed between different subtypes of ASD patients and healthy controls respectively. In addition, classification was also performed between all ASD patients and healthy controls to test the effect of subtyping on improving classification accuracy between each subtype and the healthy control group. Notably, we ignored the feature selection step and chose a linear kernel to obtain feature weighting values. Since the weighting values indirectly represent the importance of each feature in the classification process, we applied weighting values of each feature to further explore meaningful ROIs in subtypes, respectively.

Since there were more healthy controls than ASD and subtypes of ASD, which would lead to the problem of unbalanced classification. We randomly matched the numbers of healthy controls and ASD subtypes and the process of number matching was repeated 10 times to eliminate random bias. Finally, we performed tenfold cross-validation in each number matching and the hyperparameters were selected in train setting for each fold by a tenfold cross-validation. The final accuracy was defined as the average accuracy of the 10-time matching.

In this study, we evaluated the consistency of clustering assignment by adjusting the number of clusters from 2 to 8 using ARI. The maximum ARI value was found at K = 2 (ARI = 0.82), which indicated that the ASD samples were optimally partitioned by 2 subtypes based on the volumes of anatomical ROIs. The clustering results of HYDRA are shown in Additional file 1: Fig. S4. There were 115 ASD patients assigned into Subtype 1 and 106 ASD patients assigned into Subtype 2.

Compared with healthy controls, ASD patients showed both increased and decreased gray matter and white matter volume by standard case–control test (Additional file 1: Fig. S5). HYDRA effectively subdivides the above effects and further revealed the possible neuroanatomical subtypes behind the complex brain volume changes. Subtypes showed marked differences in their voxel-wise patterns of neuroanatomical deficits. Subtype 1 showed an extensive gray and white matter volume increase compared with healthy controls (Fig. 1). In contrast, Subtype 2 showed an extensive gray and white matter volume decrease compared with healthy controls (Fig. 1).

There are no significant differences in the site composition of two subtypes based on chi-square test (χ² = 3.141, df = 6, p = 0.791). Using two-sample t test, we discovered that full-scale IQ (p < 0.001, T = 3.672), performance IQ (p < 0.001, T = 3.90), ADOS total score (p = 0.041, T = 2.064) and ADOS severity scores (p = 0.004, T = 2.959) were lower in Subtype 2, but the two subtypes did not differ in age, verbal IQ, communication score and RRB score (Table 2 and Additional file 1: Table S5). The total intracranial volume was positively correlated with performance IQ in Subtype 1. Besides, the total intracranial volume was positively correlated with ADOS communication score (r = 0.233; p = 0.047) and ADOS social score (r = 0.236, p = 0.045) in Subtype 2 (Fig. 2).

The superior reproducibility of ASD subtypes were shown in the split-sample test. (Fig. 3). The voxel-wise volumetric patterns were also reproducible between the two halves in the split-half test at 2 subtypes (Additional file 1: Fig. S6 and Additional file 1: Fig. S7). Furthermore, the reproducibility analyses of the subtypes were carried out using the leave-one-site-out cross-validation. When the number of clusters was set to 2, the predicted labels of 2 subtypes from all 7 sites using leave-one-site-out were compared with the original assignments obtained by taking all the sites together. The percentage overlap of patients that were assigned to the same subtype was 87.78% (92% in Site 1, 82.14% in Site 2, 97.14% in Site 3, 75% in Site 4, 95.46% in Site 5, 86.84% in Site 6 and 84.91% in Site 7, Additional file 1: Fig. S8). The analysis of voxel-wise gray and white matter volumetric maps was consistent with the original experiment (Fig. 4).

After exclusion, 87 Subtype 1 and 81 Subtype 2 of ASD were remained. By two-sample t-test between ASD subtypes, we discovered that Subtype 1 showed lower dynamic R-fMRI measures including ReHo, DC and GSCorr mainly in the cerebellum (Additional file 1: Fig. S9). We tested the difference between ASD subtypes for mean and SD of concordance time series. The results of two sample t-test are shown in the Additional file 1: Fig. S10. The mean of concordance index of ASD subtype 1 is significantly higher than Subtype 2 (p < 0.001, T = 4.071), but SD is significantly lower than that of Subtype 2 (p < 0.001, T = 2.654) in the cerebellum. In addition, compared with Subtype2, Subtype 1 showed lower stability of dynamic functional architecture (Additional file 1: Fig. S11).

In terms of two sample t-test between healthy controls and ASD subtypes, Subtype 1 showed lower dynamic R-fMRI measures including DC, GSCorr and ReHo (Additional file 1: Fig. S12). In addition, compared with healthy controls, Subtype 1 showed lower stability of dynamic functional architecture, in contrast, Subtype 2 experienced higher stability of dynamic functional architecture (Additional file 1: Fig. S13).

In the classification of healthy controls and all ASD patients, the SVM model obtained a mean classification accuracy of 51.37%. By using HYDRA to divide the ASD into 2 subtypes, the classification accuracies were significantly improved, with a classification accuracy of 68.17% between healthy controls and Subtype 1 and a classification accuracy of 68.69% between healthy controls and Subtype 2. In addition, we calculated the average weight of each ROI in subtype-healthy control classification by linear SVM. We found that the weight ranking of the two subtype classification experiments had a great difference (Additional file 1: Fig. S14).

There are several limitations in the present study. First, the present study excluded female subjects and left-handed subjects, which limited the ability to investigate the associations between these factors and the heterogeneity of ASD. To overcome this issue, a larger dataset which contains female and left-handed subjects are needed. Furthermore, other distinctions within these subtypes may be discovered, resulting in more fine-grained parsing of heterogeneity by larger samples. Second, while the large and multisite sample in this study is an advantage, it also reduces the depth of clinical phenotyping due to different clinical cognitive scales used across sites. Third, while cross-sectional profiles have clear subtyping value, they only capture a single point in time. Enriched profiles can be derived from longitudinal data, which allows to explore changes of ASD subtypes over time. Finally, despite the advantages of reporting effect sizes and p-value for comparisons of independently collected samples, and the usage of strict multiple comparison correction methods, it is worth mentioning that they are not without limitations, and therefore, the present results should be carefully interpreted.