Edited by: Lingzhong Fan, Institute of Automation (CAS), China
Reviewed by: Zhen Yuan, University of Macau, China; Baxter P. Rogers, Vanderbilt University, United States
†These authors have contributed equally to this work and are co-first authors.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Conduct disorder (CD) is a psychiatric disorder occurred in childhood and adolescence, defined by repetitive and persistent pattern of aggressive and antisocial behaviors (
Like other psychiatric disorders, the diagnosis of CD involves multi-informant such as retrospective review, psychiatric interview and observation (
Recently, supervised machine learning (ML) in neuroimaging studies of psychiatric disorders has attracted increasing attention (
So far, there have been no studies to classify CD from HCs. Previous CD studies of sMRI only conducted group-level analysis and their results provided limited information for individual diagnosis (
A total of 60 male adolescents with CD aging 14–15 years were recruited from outpatient clinics affiliated with the Second Xiangya Hospital of the Central South University (Changsha, Hunan, China). Diagnosis was established by two experienced psychiatrists using the Structural Clinical Interview for DSM-IV-TR Axis I Disorder-Patient Edition (SCID-I/P) (
For all participants, exclusion criteria were as follows: history of ADHD, oppositional defiant disorder (ODD), any psychiatric or emotional disorder, any pervasive developmental or chronic neurological disorder, Tourette’s syndrome, post-traumatic stress disorder, obsessive compulsive disorder, persistent headaches, head trauma, alcohol or substance abuse in the past year; contraindications to MRI; or an IQ ≤ 80 on the C-WISC (
The study was approved by each school’s administration and the Ethics Committee of the Second Xiangya Hospital of Central South University (No. CSMC-2009S167). All subjects and their parents were informed of the study’s purpose and signed the informed consent.
For each participant, high-resolution structural T1-weighted images were acquired using a three-dimensional magnetization-prepared rapid gradient echo (MPRAGE) sequence on a 3T Philips Achieva scanner (Amsterdam, Netherlands) at the Second Xiangya Hospital. The acquisition parameters were: repetition time = 8.5 ms, echo time = 3.7 ms, 180 slices, slice thickness = 1 mm, acquisition matrix = 256 × 256, field of view = 256 mm × 256 mm, flip angle = 8°, image voxel size = 1.0 mm × 1.0 mm × 1.0 mm. A standard head coil was used for radiofrequency transmission and reception.
We used fivefold cross-validation (CV) for the training and testing (
Schematic flowchart of the ML classification model. We used two nested loops to build the classification model. In loop 2, we used fivefold CV method. With fivefold CV, the dataset was randomly split into fivefold, and fourfold were used for training and the remaining onefold for testing. In loop 1, the fourfold training data were divided into five fold, and we performed fivefold CV for calculating the optimized parameters. Note: CV, cross-validation.
The original DICOM images were converted to 3D NIFTI format using MRIcron (University of South Carolina, Columbia, SC, United States
Comparison of GM volumes between the two groups was performed in SPM8 using two-sample
The clusters obtained from the VBM analysis of the training data were used for calculating the input features in the testing data. Then MarsBaR 0.44 toolbox
For classifying CD from HCs, firstly we built a classification model using SVM algorithm, a classifier for two-group classification tasks (
We used two nested loops in building the SVM classification model, as shown in
where
A grid search method was used to determine the two parameters C (regularization) and σ (scaling factor of the RBF kernel) in SVM within a range of 2-8, 2-7, …, 28, respectively. Loop 1 was repeated five times, and we measured the accuracy of all the classifiers for all combinations of C and σ. The parameters that produced the highest accuracy across the fivefold were identified as the optimized combination. The classification of the testing data in Loop 2 was predicted using the optimized parameters C and σ. In this work, SVM was performed using LIBSVM
In SVM with linear kernel, the training result is to find a hyperplane in the original space of features and separate the classes as best as possible. The importance of a feature can be represented by its weight, which is the coefficient in the training model in LIBSVM. The absolute value of the coefficient represents the importance of the feature and the direction of weight represents the predicted class. We could take the dot product of any testing sample with the weights of training hyperplane: if the dot product is positive, the testing sample belongs to the positive class; otherwise the testing sample belongs to the negative class.
In addition to SVM, we established classification models by using logistic regression and random forest algorithm implemented in the scikit-learn Python library to compare the performance of different classifiers (
The performance of the classification models was evaluated by using the receiver operating characteristic (ROC) curve. The ROC curve of each classification method was calculated using the testing results for all subjects (after fivefold CV, the labels for all subjects were predicted). The area under ROC curve (AUC) was calculated and it summarized the classifier performance across all decision thresholds. The accuracy, sensitivity and specificity were calculated from the ROC curve according to the decision threshold with the highest accuracy. To evaluate the difference in classification performance of the different models, the ROC curves of SVM with RBF kernel, logistic regression, and random forest were compared with the ROC curve of SVM with linear kernel, respectively. The comparisons were performed with MedCalc package (version 12.1.4.0, MedCalc Software bvba, Ostend, Belgium).
The demographic and clinical characteristics of the two groups are shown in
Demographic and clinical characteristics of the conduct disorder (CD) group and the healthy controls (HCs) group.
Measure | CD | HCs | ||
---|---|---|---|---|
Age in years | 15.3 (1.0) | 15.5 (0.7) | 1.3 | 0.214 |
IQ | 97.0 (12.3) | 105.4 (8.8) | 4.2 | <0.001 |
BIS-attention impulsivity | 18.5 (3.2) | 18.1 (3.1) | -0.7 | 0.481 |
BIS-motor impulsivity | 26.2 (5.0) | 22.4 (3.8) | -4.4 | <0.001 |
BIS-unplanned impulsivity | 31.1 (4.6) | 28.4 (3.7) | -3.4 | 0.001 |
BIS-total scores | 75.8 (10.9) | 69.0 (8.1) | -3.7 | <0.001 |
The common clusters with significant group differences in GM volumes in the five repetitions were summarized in
Gray matter differences between CD group and HC group by VBM analysis.
Region (hemisphere) | Cluster size (voxels) | MNI coordinates |
Peak |
Regional volume (Mean ± SD) |
Feature weight (Mean ± |
|||
---|---|---|---|---|---|---|---|---|
CD (mm3) | HCs (mm3) | |||||||
Medial frontal gyrus/anterior cingulate(L) | 9392 | -2 | 51 | 7 | 5.1 | 11515.8 ± 1314.4 | 10317.4 ± 1330.0 | 0.5 ± 0.1 |
Precuneus(L) | 1666 | -44 | -75 | 37 | 4.3 | 2093.0 ± 438.4 | 1761.7 ± 486.6 | 0.2 ± 0.1 |
Superior parietal lobule(L) | 478 | -32 | -64 | 63 | 3.8 | 245.8 ± 57.0 | 205.3 ± 60.3 | 0.4 ± 1.8 |
Superior frontal gyrus(R) | 318 | 21 | 62 | -26 | 3.8 | 221.1 ± 56.0 | 187.1 ± 46.6 | 0.6 ± 1.6 |
Subthalamic nucleus(R) | 375 | 6 | -16 | -14 | 4.3 | 155.4 ± 32.5 | 135.3 ± 18.6 | 2.0 ± 1.8 |
Cerebellum posterior lobe(R) | 210 | 15 | -39 | -51 | -3.9 | 166.3 ± 26.0 | 184.6 ± 25.7 | -2.6 ± 1.7 |
Inferior parietal lobule/insula(R) | 1088 | 56 | -18 | 22 | -4.7 | 1618.2 ± 275.3 | 1870.7 ± 388.1 | -2.3 ± 0.4 |
Lingual gyrus(R) | 71 | 5 | -94 | -18 | -3.6 | 32.0 ± 13.1 | 41.4 ± 15.6 | -1.7 ± 1.6 |
Results of VBM analysis presented at
We calculated the features weights in the fivefold CV of SVM with linear model, as shown in
The ROC curve was shown in
The ROC curves of different classification models. SVM, support vector machine.
Performance of the proposed classification model with different classifiers.
SVM (linear) | SVM (RBF) | Logistic regression | Random forest | |
---|---|---|---|---|
Accuracy (%) | 80.4 | 79.6 | 79.4 | 77.9 |
Specificity (%) | 73.3 | 73.8 | 78.8 | 80.4 |
Sensitivity (%) | 87.5 | 85.5 | 80.0 | 75.4 |
AUC | 0.78 | 0.79 ( |
0.76 ( |
0.80 ( |
In the current study, we demonstrated that GM volumes can be used to distinguish CD from HCs by using supervised ML techniques. The regional GM volumes which were significantly different across the groups, combined with ML algorithm, correctly identified CD from HCs with approximately 80.0% accuracy. The performance of our proposed model was comparable, even better than previous similar studies.
Efficient feature extraction may greatly improve the performance of the classification model. In our study, we did not use the BIS scales or other behavioral information as the features because our aim was to explore whether the imaging features are able to classify CD from HCs. Compared with HCs, CD exhibited GM volume alterations in multiple brain regions, which were predominantly in the frontal lobe, parietal lobe, anterior cingulate, cerebellum posterior lobe, lingual gyrus, and insula areas (
In the present study, the highest coefficient in SVM model with linear kernel was found for the right cerebellum posterior lobe, which indicated that the volume of right cerebellum posterior lobe greatly contributed to the classification results. In addition, the weight of right insula was also large and this demonstrated that the volume of right insula played also an important role in the classification task. The absolute value of right lingual gyrus weight was also large, compared with the first four features in the
Our studies also detected higher GM volumes of CD patients in the left precuneus, the anterior cingulate, and the superior frontal gyrus Precuneus is involved in self-referential and self-centered thinking, and plays an important role in self-referent information processing (
Taken together, the findings of GM volumes in our study were supported by the results of previous studies, which might implicate the etiology of CD. Thus, such abnormalities were expected to serve as efficient features in identifying CD.
Besides the feature extraction, efficient classification also requires an appropriate classifier which can learn the decision rules from the given features. In this study, we employed SVM with linear kernel, SVM with RBF kernel, logistic regression and random forest (
Compared with logistic regression, the loss function of SVM do not penalize subjects for which the correct decision is made with sufficient confidence, and this may be good for generalization. However, logistic regression loss function does not become zero even if the subject is classified with sufficiently confidence, and this may lead to reduced accuracy in logistic regression classification (
The classification results may differ among samples. Generally, the variance of the classification is expected to decrease as the sample size increases (
There were several limitations in this study. Firstly, we only analyzed structural MRI images, but previous studies suggested that CD was also associated with brain functional abnormalities (
In this study, we detected regional differences of GM volume between CD and HCs by using VBM, and these regional GM volumes were shown reliable in establishing a ML model to discriminate between CD patients and HCs with high accuracy. Although our classification model was not meant to be a substitute to the current clinical diagnosis of CD, it might be an objective and reliable diagnostic tool that could help reduce the variability in clinical practice, and thus may help to improve the diagnosis of CD.
WL, SY, and BH: study conception and design, funding support. JinZ and YG: acquisition of data. JiaZ, WL, and QW: analysis and interpretation of data. JiaZ and JinZ: drafting of manuscript. WL, YJ, JG, SY, and BH: critical revision.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
The study was funded by the National Natural Science Foundation of China (No. 81471384) and the Seed Funding from Scientific and Technical Innovation Council of Shenzhen Government (No. 000048).