Background
Psychiatric disorder is among the most important causes of mortality in humans, which affects the quality of life and increases the social burden [
1‐
3]. Psychotic (such as schizophrenia [SCZ]) and affective disorders (such as major depressive disorder [MDD]), as two typical psychiatric disorders, have extensive comorbidities with each other [
4,
5]. Approximately 80% of patients with SCZ experience depressive episode in the early disorder stages [
6]. The depression prevalence among patients with SCZ can be as high as 40% [
7,
8]. Moreover, patients with MDD have been shown to have a higher risk of developing a psychotic disorder. In addition, depression often precedes psychotic symptoms in people with a high risk of SCZ [
9,
10]. The presence of psychotic symptoms in patients with depression is considered a clinical depression subtype known as psychotic depression, which is associated with increased depressive symptom severity [
11,
12]. Further, both MDD and SCZ significantly impair working memory, planning, shifting and so on [
13,
14]. In addition, MDD and SCZ have significant genetic similarities [
15,
16]. This complex relationship between MDD and SCZ sometimes can impede diagnoses by psychiatrists.
Adolescence is the critical period during psychological development in which most psychiatric disorders are initially detected [
17,
18]. Similarly, there is an overlap in the clinical characteristics of adolescents with MDD and SCZ [
19]. At this stage, although there are not wide effects on behaviors influenced by psychiatric disorders, they can have significant negative effects later in life and are potential health threat for future generations [
20,
21]. Prior to severe symptoms during adulthood, SCZ often begins developing during early adolescence [
18,
22,
23]. Compared to patients with adult-onset SCZ, those with early-onset often exhibit more severe psychotic symptoms, poorer therapeutic outcomes, and greater disability [
24,
25]. In structural neuroimaging studies, although a meta-analysis by Van et al. found a thinner cortex in adult patients with SCZ (especially in the frontal and temporal lobe regions), Thormodsen et al. reported no significant difference in the cortical thickness between adolescent patients with SCZ and healthy adolescents [
26,
27]. These findings indicate changes in cerebral cortex of the adolescent patients with SCZ may take time to develop.
The first episodes of affective disorder, including MDD, appear at adolescence and cause serious distress to the patients and their guardians [
28,
29]. In the symptoms of MDD, appetite and weight changes, energy loss, and insomnia can be seen among adolescents while concentration problems and anhedonia/loss of interest are more frequent among adults [
30]. Structural neuroimaging studies have reported reduced cortical thickness in the dorsal lateral prefrontal cortex, lingual gyrus, and pre- and postcentral gyrus in patients with early-onset depression [
31]. Reynolds et al. reported a thicker bilateral dorsal-lateral prefrontal cortex and left caudal anterior cingulate cortex in MDD adolescents [
32]. Contrastingly, previous studies have reported no significant differences in the cortical thickness at the whole-brain level between adult patients with MDD and healthy controls [
31,
33]. Consistent with findings on SCZ, these findings indicate there are a lot of differences in symptoms and brain morphology between adolescents and adults with depression. Therefore, pathophysiological mechanisms might differ between adolescents and adults with MDD.
Machine learning is an emerging technology in recent years that can help us better understand the pathophysiological mechanisms of the brain. It involves assessing the similarity of a brain MRI scan with images obtained from a group of individuals to determine whether the tissue is more likely from a patient or a healthy individual [
34]. Over the past decade, different machine learning methods using various brain features have been developed to distinguish between psychotic and affective disorders with a good accuracy ranging from 60 to 90% [
35‐
38]. However, these studies on disease identification have mostly focused on adults without accounting for pathophysiology differences at different stages, especially in adolescence, which is a critical development period. It remains unclear whether adolescent patients with MDD and SCZ can be distinguished via structural brain MRI scans.
Consequently, we used a support vector machine (SVM) to determine whether it could be used to accurately identify adolescent patients with SCZ and MDD at the individual level based on anatomic brain parameters, as well as to determine their key brain characteristics [
39]. We hypothesized that SVM could accurately distinguish among MDD, SCZ, and healthy controls. To our knowledge, this is the first study to examine psychiatric disorders (MDD and SCZ) in adolescents using a machine learning technique. This study provides the insight of MDD and SCZ. Moreover, this study may contribute toward the identification of adolescent psychiatric disorders based on MRI scans and a scientific basis for early clinical diagnosis of psychiatric disorder.
Discussion
To our knowledge, this is the first MRI study to distinguish between adolescent psychiatric disorders (MDD and SCZ) using machine learning techniques. We employed a linear kernel nested SVM to create data-driven models for classifying patients with MDD, patients with SCZ, and HC based on whole-brain neuroanatomical features in MRI scans. The models using cortical thickness could distinguish adolescents with MDD and SCZ from healthy adolescents with an accuracy of 79.21, and 69.88%, respectively. The classification between adolescents with MDD and those with SCZ had a lower accuracy of 62.93%. Our findings indicate that machine learning using cortical thickness as the features can allow effective classification of psychiatric disorders among adolescents at an individual level.
Our findings indicate that structural brain MRI imaging can be used to effectively identify MDD and SCZ in adolescents. Previous studies widely used different structural indexes, including subcortical volume, gray matter density, gray matter volume, cortical thickness, and cortical area, and all of them could distinguish between adults with psychiatric disorders and healthy adults [
38,
58‐
60]. In this study, we focused on adolescent psychiatric patients and collected structural MRI data from patients with MDD and SCZ. Different structural indexes were used as SVM model features respectively for case- classification and MDD-SCZ classification. Unlike previous findings on adults, only cortical thickness could provide the best accuracy in three adolescent classification models in our study. This is consistent with the findings by Qiu et al., who used an SVM based on various brain morphometric features to distinguish between 32 adult patients with first-episode MDD and 32 HC [
61]. They reported that multiple cortical features could discriminate them with cortical thickness providing the highest accuracy. Our findings indicate that the cortical thickness is already altered in adolescent patients with MDD and SCZ. This is consistent with previous findings that patients with childhood-onset schizophrenia presented with bilateral deficits in the temporal, prefrontal, and parietal cortices [
62]. Moreover, using the machine learning technique, cortical thickness has been reported to predict future-onset of depression in adolescents with an accuracy of 70% [
63]. Besides, with volumes of both subcortical and cerebellar regions as the feature set, the classification model of MDD-SCZ resulted in a significant accuracy of 62.93% (
p = .044). This is because some subcortical nuclei are also linked to MDD and SCZ, such as amygdala, which is associated with emotion [
64,
65]. In addition, there is an interesting finding that we succeed to distinguish adolescents with SCZ and HC but Thormodsen et al. not [
26]. In their study., there are no significant evidence of cortical thickness difference between adolescent with SCZ and HC based on univariate analysis. In our study, to further explore the brain morphology of adolescents with SCZ, we succeed to distinguish them using multivariate analysis. Although no significant evidence is found in cortical thickness of each brain region between the two groups, there may be a particular spatial pattern of abnormal changes in cortical thickness across brain regions in adolescents with SCZ. That may be why we are successful. In a word, cortical thickness is a crucial structural brain index for identifying adolescent patients with psychiatric disorders.
For distinguishing adolescent patients with MDD from HC, the most important brain region was the temporal pole. The temporal pole, which is a node of the paralimbic system, plays an important role in socioemotional and cognitive processing [
66]. Defects in these processes are associated with depression [
67,
68]. Gray matter abnormities in the temporal pole have been reported in medication-naive patients with first-episode MDD [
69,
70]. Compared to healthy controls, individuals with depression present with greater activation of the right anterior temporal pole [
71]. Previous studies also reported abnormal functional connections between the right temporal pole and other brain regions in patients with MDD [
72‐
74]. Given the emotional instability in adolescents and the abnormal emotional response to external stimuli, abnormal changes are more likely to occur in the temporal pole [
75]. Therefore, the structure of this region could be used as a crucial biomarker for adolescent depression.
The left banks of the superior temporal sulcus, which is a crucial association area for biological motion perception, was the most significant brain region for distinguishing between adolescents with SCZ and HC [
76]. The superior temporal sulcus is part of a neural circuit involved in perceiving intention from action and reactions to social and emotional events [
77,
78]. Many studies have reported a reduced ability to extract social information from bodily cues in patients with SCZ [
79‐
82]. Neuroimaging studies have reported that patients with SCZ present with an aberrant pattern of superior temporal sulcus activity during basic biological motion tasks [
83,
84]. Matsumoto et al. reported a negative correlation of the behavioral performance on basic biological motion perception tasks and the gray matter volume of the superior temporal sulcus in patients with SCZ [
85]. Similarly, we observed adolescents with SCZ had thinning cortical thickness of the left banks of the superior temporal sulcus than HC (
p = 0.005, FDR corrected). Our findings indicate that the superior temporal sulcus could be associated with impaired extraction of social information in adolescents with SCZ.
In our study, the most important brain region that distinguishing between MDD and SCZ was the right pars triangularis. The pars triangular is located in the inferior frontal gyrus, which is a crucial brain region for emotional and cognitive control circuits [
86]. Deng et al. reported that the right inferior frontal gyrus is highly activated in a stop-signal task involving motor inhibitory responses [
87]. Damage to this area impairs the performance of the stop-signal task [
88]. Moreover, individuals with higher depression levels were found to have poorer response inhibition and to perform worse on the stop-signal task [
89]. Neuroimaging studies have reported that patients with MDD have increased functional connectivity in the right pars triangularis of the inferior frontal gyrus [
90,
91]. This indicates a strong correlation of the right pars triangularis with depression in adolescents. Patients with SCZ also present with reduced gray matter volume in the right inferior frontal gyrus [
92]. However, this is attributed to the generalized neuropsychological impairment associated with SCZ rather than impaired inhibitory behavioral control, which is a specific cognitive impairment [
93]. Taken together, these findings indicated that the right pars triangularis is associated with response inhibition in adolescents with MDD and could be used to distinguish between adolescent patients with MDD and SCZ.
This study has several limitations. First, the accuracies of our models were all < 80%. To improve accuracy, we combined other indexes (cortical volume, cortical area, and cerebellar-subcortical volume) with cortical thickness as the feature set. An additional table presents these results [see Additional file
2]. After adding additional indexes into the feature set, no improved prediction accuracy was found. To prevent model overfitting and to improve accuracy, we applied the least absolute shrinkage and selection operator for feature selection and dimensionality reduction [
94]. However, this did not reduce dimensions, which could be attributed to the complexity of brain structures and small sample size. In addition, it is a limitation that we exclude those patients with psychiatric comorbidities to maximize the group difference to train the classifier. The patients with comorbidities are valuable cases to investigate for diagnostic purposes. In the future, we will apply the models here to these groups. Moreover, we obtained our sample from a single center. It is not clear whether our results are reproducible and generalizable. In future studies, we will obtain multi-center samples to validate these findings and continue to focus on early psychiatric disorders.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.