The main findings in the present study were that after controlling for the whole-brain volume and age, the SCZ individuals with a history of violence showed reductions in several brain regions involved in emotion and cognition processing, including left frontal pole volume as well as right bankssts and inferior parietal cortical thickness, compared to the patients with no history of violence. Subsequently, these abnormal brain regions were used to develop the prediction model for violence among SCZ patients using machine learning method. Ultimately, seven predictive models were established. Through comparing with each other, the SVM had the best performance with a balanced accuracy of 0.8231 and an AUC of 0.841.
In the present study, the finding of reduced whole brain volume in patients with SCZ who had a history of violence, compared to those without a history of violence, is in consistence with results of previous studies [21.22], suggesting the possibility that certain general cognitive impairments associated with whole brain volume reduction are involved into violence. Besides, we also found abnormalities in several regions implicated in the neuropathology of violence, including left frontal pole, right bankssts and inferior parietal cortical regions. Patients with SCZ who had a history of violent behavior displayed decreased gray matter volume in frontal pole in relation to those without violence, which is consistent with another study [23]. This indicates that changes in the prefrontal cortex including the frontal lobe might be involved in the pathophysiology of violent behavior. The prefrontal cortex is thought to play an important role in executive functioning capacity, including regulation of inhibition, emotions and movement. Damage or dysfunction of this area may interrupt the sending of inhibitory inputs to the limbic system which is composed of hippocampus and parahippocampal gyrus, and increases the risk of unregulated behavior [24], speculating that the region can be regarded as structural markers for violence [25]. Temporal lobe is implicated in emotional processing and its abnormalities are linked to the onset of psychosis, hallucinations and delusions in SCZ [26.27]. Delusions as one of the important features of SCZ have been most consistently related to violent behavior [28]. In addition, the relation of abnormal temporal lobe with violence is also supported by other evidence that alternations in this region can lead to impaired aggression control and increased impulsivity which belong to aspects of characteristic antisocial personality disorder [29.30]. We also found subjects with SCZ and a history of violence displayed decreased cortex thickness in right parietal lobe in relation to those without violence, in consistent with findings from the previous studies [31] which suggested parietal lobe as a part of the default-mode network (DMN) which is responsible for self-referential and reflective activity as well as attending to internal and external stimuli [32].
In our study, we found SVM was appropriate to structural MRI data and had better predictive performance in differentiating violent from non-violent patients with SCZ than other six machine learning algorithms, with its balanced accuracy and AUC reaching 0.8231 and 0.841, respectively. SVM belongs to one of the machine learning algorithms which can process high-dimensional data and capture nonlinear variable relations. SVM belongs to a supervised discriminative classification method that treats each sample to be classified as a point in high-dimensional space, and constructs hyperplanes, or a decision boundary between the prediction classes to optimally separate the data into different groups. To date, studies predicting risk of violence in SCZ patients using neuro-imaging data are sparse. The only research employed multimodal machine learning method to identify SCZ patients at high risk of violence. The model based on the single modality of gray matter volume showed an accuracy of 77.33% and an AUC of 0.80 [20]. The possible reason for different predictive power is that our model included more characteristics of cortical morphology indexes, namely cortical surface area, gray matter volume and cortical thickness, to improve the power of recognizing those patients with greater risks of violence. Besides, there have been a few studies which combined machine learning algorithms and demographic and clinical data to differentiate patients with violence from those without violence, but the performance of prediction models was unsatisfactory [9–11]. In this study, the prediction model integrating structural MRI characteristics demonstrated good performance. Above evidence suggests that due to high anatomical resolution of cortical volume, area and thickness, structural MRI features can be suggested to be biologically-based predictive markers.
Several limitations need to be considered. First, our sample size was relatively small, Future studies should recruit more participants to improve the power of predicting violence in SCZ. Second, the inpatients with SCZ enrolled by us were receiving treatment with a medication. Despite effect of medication on cortical morphology is still unknown, the brain structure of SCZ patients might be influenced by antipsychotic treatment. In order to validate our results, future studies should be conducted in first-episode, medication-naive patients with SCZ. Third, the present study lacked external validation, which could limit the generalization of our findings. Future research should perform the external validation in another sample. Fourth, the model developed by male SCZ patients was not applicable to female individuals. Considering difference in brain structure and risk factors between male and female patients, the models based on gender need to be built in future studies.