Identification of male schizophrenia patients using brain morphology based on machine learning algorithms

Abstract

BACKGROUND

Schizophrenia is a severe psychiatric disease, and its prevalence is higher. However, diagnosis of early-stage schizophrenia is still considered a challenging task.

AIM

To employ brain morphological features and machine learning method to differentiate male individuals with schizophrenia from healthy controls.

METHODS

The least absolute shrinkage and selection operator and t tests were applied to select important features from structural magnetic resonance images as input features for classification. Four commonly used machine learning algorithms, the general linear model, random forest (RF), k-nearest neighbors, and support vector machine algorithms, were used to develop the classification models. The performance of the classification models was evaluated according to the area under the receiver operating characteristic curve (AUC).

RESULTS

A total of 8 important features with significant differences between groups were considered as input features for the establishment of classification models based on the four machine learning algorithms. Compared to other machine learning algorithms, RF yielded better performance in the discrimination of male schizophrenic individuals from healthy controls, with an AUC of 0.886.

CONCLUSION

Our research suggests that brain morphological features can be used to improve the early diagnosis of schizophrenia in male patients.

Key Words: Schizophrenia, Machine learning, Classification, Structure, Magnetic Resonance imaging

Core Tip: Schizophrenia is a severe psychiatric disease characterized by impairments in cognition, positive and negative symptoms, affecting about 1% of the general population worldwide. A fast diagnosis of schizophrenia is crucial to prescription of an appropriate anti-psychotic in the early stage, which is able to make treatment more efficient. Many studies have demonstrated widespread functional and structural brain alternations from magnetic resonance imaging in individuals with schizophrenia in relation to healthy controls. our aims were to employ four commonly used machine learning algorithms including general linear model, random forest, k-nearest neighbors, and support vector machine and a wider range of brain morphological features to avoid bias towards a particular machine learning algorithm and improve the performance of classification between male individuals with schizophrenia and healthy controls in the present study.

INTRODUCTION

Schizophrenia is a severe psychiatric disease characterized by impairments in cognition and positive and negative symptoms, and it affects approximately 1% of the general population worldwide[1,2]. A fast diagnosis of schizophrenia is crucial for the early-stage prescription of an appropriate antipsychotic agent, which can improve treatment efficacy[3]. Currently, the diagnosis of schizophrenia is carried out based on doctors’ judgment through observing the behaviors and psychiatric symptoms of patients[4-7]. However, in many cases, the diagnosis is not highly reliable due to clinician subjectivity, patient heterogeneity and nomenclature inadequacy, and the diagnosis of schizophrenia in the early stages is still considered a challenging task.

Many studies have found widespread functional and structural brain alterations from magnetic resonance imaging (MRI) data in individuals with schizophrenia in comparison to healthy controls[8-12]. Among various MRI techniques, structural MRI, which has the advantages of high spatial resolution, low cost and low sensitivity to noise, has been widely applied to detect differences in brain structures between schizophrenia patients and normal controls. For example, a multimodal meta-analysis with a sample of 801 schizophrenia patients and 957 healthy controls showed that patients with schizophrenia exhibited gray matter volume (GMV) abnormalities in many brain regions relative to healthy controls[13]. One multicenter meta-analytic study including 572 first-episode psychosis (FEP) patients and 502 healthy controls collected from 5 sites showed decreased GMV in the fronto-temporal, insular and occipital regions bilaterally in FEP patients compared to healthy controls[14]. Takayanagi et al[15] reported decreased cortical thickness in the frontal and temporal regions in schizophrenia patients compared to healthy controls. Although these abnormal brain structures were detected based on group-level statistical analyses, it is difficult to identify persons at high risk for schizophrenia at the individual level.

Machine learning, as a computational technique, can learn from data inputs and deliver a solution automatically. It is able to overcome the drawbacks of conventional statistical analysis and calculate the probability of an individual being diagnosed with a disease[16-18]. To date, a few studies using machine learning-assisted MRI characterization have attempted to provide diagnostic information for individual patients[19-24]. However, the majority of these studies utilized only a single machine learning algorithm to perform the classification and did not confirm whether the selected machine learning algorithm was optimal. This is because machine learning uses a series of algorithms, and each algorithm has unique methodologies for data processing and model development[25]. In addition, the number of MRI features extracted for discriminating schizophrenic individuals from healthy controls varies between studies. Considering that the type of data is closely associated with the performance of each algorithm[26], our aims in the present study were to employ four commonly used machine learning algorithms, namely, the general linear model (GLM), the random forest (RF), k-nearest neighbors (KNN), and the support vector machine (SVM), with a wide range of brain morphological features to avoid bias toward a particular machine learning algorithm and improve the classification performance in discriminating male individuals with schizophrenia from healthy controls.

MATERIALS AND METHODS

Participants

A total of 78 male subjects, 60 patients with schizophrenia and 18 healthy controls, participated in this study. The patients diagnosed with schizophrenia were recruited from inpatient units of Hefei Fourth People’s Hospital from July 2021 to December 2021. The inclusion criteria were as follows: Meeting the diagnostic criteria for schizophrenia according to the Diagnostic and Statistical Manual of Mental Disorders-5 and right-handedness. In addition, healthy controls were recruited from the community via advertisements. The inclusion criteria were as follows: Willingness to participate in the study, no history of any psychiatric disorder, and no family history of psychosis in first-degree relatives. The exclusion criteria for patients with schizophrenia and healthy controls were as follows: Intellectual disability, severe contraindications for MRI, left-handedness, substance abuse, and a history of head injuries. This study was approved by the ethics committee of Hefei Fourth People’s Hospital, and all participants provided written informed consent.

MRI acquisition

The structural MRI data of all subjects were collected on a 3.0-Tesla Siemens MRI scanner at Hefei Fourth People’s Hospital. T1-weighted images were collected with the following sequence: repetition time = 8.5 ms, echo time = 3.2 ms, inversion time (TI) = 450 ms, flip angle (FA) = 12°, field of view = 256 mm × 256 mm, matrix size= 256 × 256, slice thickness= 1 mm, no graph, voxel size = 1 mm³ × 1 mm³ × 1 mm³, 188 sagittal slices, and acquisition time = 296 s. Precautions were provided for noise attenuation, and a birdcage head coil with foam padding was placed around the subject’s head to minimize movement. The subjects were instructed to remain motionless and close their eyes during scanning.

MRI processing

Before processing the structural MR data, we visually inspected the MR images of all the subjects and excluded those of three participants with schizophrenia because of motion artifacts. Finally, structural T1 images of 57 schizophrenia patients and 18 healthy controls were converted to the NIfTI format using dcm²nii software[27] and then processed using FreeSurfer version 5.0 (https://surfer.nmr.mgh.harvard.edu/). The reconstruction of the cortical surface was performed using a standard automatic reconstruction algorithm, which generated several types of brain morphological features, including cortical volume (CV), cortical thickness (CT), and cortical surface area (CSA). Each feature was calculated for 34 brain regions for each hemisphere according to the Desikan-Killiany-Atlas[28].

Feature selection

When the CV, CT, and CSA features of 68 cortical regions were analyzed simultaneously, some uninformative, irrelevant, or redundant features could not be removed, leading to multicollinearity between features. Eliminating redundant features not only highlights important features but also improves classification accuracy. This study adopted the least absolute shrinkage and selection operator (LASSO) to address multicollinearity by compressing the coefficients of these unimportant variables to 0 and achieve feature selection.

Development and performance of the models

The establishment and evaluation of the models were conducted in R software. All the data were randomly divided into training (70%) and testing sets (30%). The training set was used for developing the models, and the testing set was used for assessing the performance of the models. Furthermore, to adjust the model parameters and avoid model overfitting, 10-fold cross-validation was performed on the training set. In 10-fold cross-validation, the entire dataset in the training set was randomly divided into 10 equal subsets. One subset was selected to test the model, and the remaining 9 subsets were used for training. This process was repeated 10 times. All algorithms were evaluated on the testing set using the area under the receiver operating characteristic curve (AUC) as a performance metric, and the algorithms were then compared to select the optimal one. An AUC of 0.5 indicates no discriminative ability, and 0.7-0.8 indicates good discrimination.

Statistical analysis

Statistical analysis was conducted in SPSS version 16.0. Differences in age, whole-brain volume and brain morphological features between schizophrenia patients and healthy controls were determined using t tests. P < 0.05 was considered to indicate statistical significance.

RESULTS

Comparison of demographic characteristics

There were no significant differences in age or whole-brain volume between male patients with schizophrenia and healthy controls, as shown in Table 1.

Table 1 Comparison of age and whole volume between two groups.

Variables	Healthy controls	Schizophrenia patients	t value	P value
Age (yr)	34.06 ± 8.88	35.04 ± 11.21	-0.338	0.736
Whole volume (mm3)	1203600.00 ± 88093.08	1164100.00 ± 112071.00	1.365	0.176

Selection of the important features

The important features selected by LASSO are shown in Table 2, including the banks of the superior temporal sulcus (bankssts), cuneus, inferior temporal, isthmus cingulate, lateral occipital, lingual, paracentral, parsopercularis, superior frontal, temporal pole, and insula cortex regions.

Table 2 The important features extracted by least absolute shrinkage and selection operator.

Variables		Coefficients
Surface area	Bankssts	-0.00005308106
	Inferior temporal	-0.00004908730
	Lateral occipital	-0.00003610912
	Lingual	-0.00022333510
	Insula	0.000007207081
	Isthmus cingulate	0.000156094500
	Paracentral	0.000248385300
Gray matter volume	Superior frontal	-0.00001763374
	Temporal pole	0.000149290400
Cortical thickness	Lingual	0.062200590000
	Cuneus	0.037246120000
	Lateral occipital	0.239341600000
	Par sopercularis	0.000000005200

Comparison of features between groups

After the important features were selected using LASSO, we compared differences in these features between groups. Patients with schizophrenia showed reductions in the left bankssts area, left inferior temporal area, left lateral occipital area and left superior frontal volume as well as increased cortical thickness in the left lingual, right cuneus and lateral occipital regions compared with healthy controls, as shown in Table 3.

Table 3 Comparison of morphological features between two groups.

Variables	Schizophrenia patients	Healthy controls	t value	P value
Left bankssts area	1016.20 ± 170.13	1133.20 ± 186.00	-2.488	0.015
Left inferior temporal area	3492.30 ± 557.45	3856.30 ± 423.39	-2.544	0.013
Left lateral occipital area	4925.30 ± 719.19	5489.80 ± 480.09	-3.111	0.003
Left lingual area	2812.00 ± 444.30	3236.70 ± 479.05	-3.471	0.001
Left lingual thickness	2.08 ± 0.17	1.99 ± 0.10	2.121	0.037
Left superior frontal volume	23123.00 ± 2824.16	24744.00 ± 2448.14	-2.187	0.032
Right cuneus thickness	2.01 ± 0.14	1.89 ± 0.14	2.938	0.004
Right lateral occipital thickness	2.18 ± 0.15	2.07 ± 0.15	2.738	0.008

Performance of models

RF outperformed the other machine learning algorithms and achieved better classification performance, with an AUC of 0.886, as shown in Table 4. Subsequently, CV, CT, and CSA were used for the development of classification models using the RF algorithm. Among the three brain morphological features, CT performed best, and its AUC and balanced accuracy reached 0.605, as shown in Table 5.

Table 4 Performance of each machine learning algorithm.

Algorithms	AUC	Balanced accuracy, %	Sensitivity	Specificity
GLM	0.728 (0.470-0.986)	61.84	0.737	0.500
RF	0.886 (0.754-1.000)	64.04	0.947	0.333
KNN	0.601 (0.257-0.945)	55.70	0.947	0.167
SVM	0.842 (0.670-1.000)	50.00	1.000	0.000

AUC: Area under receiver operating characteristic curves; GLM: General linear model; RF: Random forest; KNN: K-nearest neighbor; SVM: Support vector machine.

Table 5 The performance for each structural feature.

Features	AUC	Balanced accuracy, %	Sensitivity	Specificity
Surface area	0.474 (0.241-0.706)	39.50	0.789	0.000
Gray matter volume	0.553 (0.235-0.871)	56.60	0.632	0.500
Cortical thickness	0.605 (0.327-0.884)	55.70	0.947	0.167

AUC: Area under receiver operating characteristic curves.

DISCUSSION

In the present study, we employed the important features extracted from several brain morphological features using LASSO and t tests, including CV, CT, and CSA, to establish models for the classification of male schizophrenia patients and healthy controls based on four commonly used machine learning algorithms, namely, GLM, RF, KNN, and SVM, and then compared the classification performance of these models in terms of AUC to determine the optimal machine learning algorithm.

In this study, there were a total of 8 morphological features with significant differences between groups, including the left bankssts CSA, right cuneus CT, left inferior temporal CSA, left lateral occipital CSA, left lingual CT, left superior frontal CV, right cuneus CT, and right lateral occipital CT, which are largely consistent with the findings of many previous studies. For example, Shi et al[29] reported abnormalities in the inferior temporal gyrus and superior frontal gyrus. Zhao et al[30] reported that schizophrenia patients had extensive structural abnormalities, such as in the occipital lobe and superior frontal gyrus. These results suggest extensive changes in the brains of schizophrenia patients, which may be important regions in the classification of schizophrenia patients and healthy controls.

We found that RF performed better (with an AUC of 0.886) in discriminating male schizophrenia patients from healthy controls than did the other three machine learning algorithms, which was consistent with the results of other studies[14,24]. Furthermore, the RF algorithm identified as the best algorithm in this study had higher classification accuracy than those of previous studies[31]. For example, a study utilizing SVM with gray matter and white matter features selected with recursive feature elimination to discriminate individuals with schizophrenia from healthy controls reported that the optimal machine learning algorithm achieved an accuracy of over 85%[19]. In another study, features selected from amygdaloid and hippocampal subregions using sequential backward elimination were employed to distinguish between schizophrenia patients and normal controls, and the SVM classifier achieved an accuracy of 81.75%[24]. Yassin et al[21] used subcortical volumes and cortical thickness features of 64 schizophrenia patients and 106 healthy controls, and the RF classifier applied to subcortical volumes had the highest accuracy of 76.4%. Xiao et al[23] used cortical thickness and CSA data to classify 163 schizophrenia patients and 163 healthy controls and achieved accuracies ranging from 81%-85%. In another study using a diverse set of neuroanatomical measures and ensemble methods, the classification accuracies of the ensemble methods ranged from 83% to 87%[20]. In addition, compared to other brain morphological features, the classification model established using RF and CT achieved greater classification accuracy, which is consistent with results from previous studies reporting that cortical thickness features yielded greater classification performance than other neuroanatomical features[14,21]. The higher performance of the classification models in this study compared with those in other studies may be due to the following reasons. First, LASSO was adopted for the selection of important features in our study. All brain morphology features were ranked according to their importance, and the unimportant features were filtered out. This not only increases the computation speed but also improves the accuracy of classification[32]. Second, we employed more brain morphological indices as input features for classification than did other studies. Finally, due to the close association of the type of data with the performance of each machine learning method, we compared the performance of several machine learning algorithms in discriminating schizophrenia patients from normal controls and found that RF performed best, suggesting that RF is more appropriate for brain morphological features extracted using LASSO. Additionally, RF is regarded as an ensemble method that can combine individual weak classifiers to generate a more robust classification system.

Several limitations in this study need to be considered. First, the size of the sample was relatively small. Future studies should recruit more subjects to improve the ability to classify patients with schizophrenia and healthy controls. Second, validation data from another center were not available, which impacted the generalizability of the classification models. Future research should perform external validation at other sites. Third, the schizophrenia patients included in this study had chronic schizophrenia, and their brain structures could have been influenced by antipsychotic use. Future studies should be conducted in first-episode, medication-naive patients with schizophrenia to validate our results.

CONCLUSION

In this study, we found structural abnormalities in some brain regions primarily involved in the temporal and frontal lobes in schizophrenia patients. Compared with other algorithms, RF showed better performance in discriminating male schizophrenia patients from healthy controls.

ACKNOWLEDGEMENTS

The authors thank all participants in the study, as well as investigators involved in conducting the study.