The objective of the current study was to identify task-related fMRI markers associated with past suicide attempts, assessing two putatively relevant aspects of adolescent suicidality: perception of social inclusion/exclusion and cognitive control. In summary, our study revealed two major findings. First, we observed a set of neural disturbances during both social inclusion/exclusion and cognitive inhibition tests in adolescent suicide with past suicide attempt in comparison to both control groups. These alterations were mainly located in the left insula and prefrontal cortex for social perception and interaction; and motor and prefrontal cortices for inhibition of action (at a more stringent level of correction). Interestingly, behavioral performances (self-report of emotional feelings during the Cyberball Game and reactions times and omission/commission errors at the Go-NoGo task) were similar between both patient groups suggesting that functional MRI may yield finer properties than neuropsychological tests and questionnaires to discriminate patients with and without a personal history of suicide attempt. Second, activity in these significant brain clusters enhanced the identification of a past history of suicidal attempt, outperforming clinical and sociodemographic variables.
During the Cyberball Game, adolescent SA displayed distinct neural functional alterations in relation to both inclusion and exclusion conditions, as compared to PC and HC. In contrast to previous studies (Harm et al. 2019), we found that the set of anatomical regions distinguishing suicidal adolescents depended on the valence of social circumstances (i.e., inclusion or exclusion). In the inclusion situation, SA showed reduced neural activity in the left insula, a region involved in the salience network [80]. Decreased activity in left insula (although more posterior) during inclusion vs. control condition at the Cyberball Game was also found in adult females with past attempts (Olié et al, 2017). In our study, lower insular activity was associated with lower feeling of belongingness and having a less meaningful existence during perceived social exclusion. These two factors are in line with psychological risk factors associated with suicide [81], somewhat paralleling the concepts of “thwarted belongingness” and “perceived burdensomeness” found in the Interpersonal Theory of Suicide [27]. Thus, it is possible that the lower activation of the left insula during inclusion in SA may interfere with the normally reinforcing feelings associated with inclusive social interactions [82]. These findings may also reflect a lower ability of SA to feel connected to others. It may be hypothesized that in SA this may both limit their pleasantness of being with others and also their will to seek help when in difficulty. Reduced disclosure of suicidal ideas has indeed been associated with increased risk of social isolation and suicide attempt in adolescents [83, 84]. Of note, lower insular activity during inclusion was also correlated with higher depressive state, suggesting a significant effect of the negative mood state that may be stronger in individuals at risk of suicide.
In contrast to inclusion, social exclusion in SA was associated with increased activity in right middle/superior frontal gyrus (surviving at a more stringent correction level) and lower activity in the right inferior frontal gyrus compared to PC (and in the left inferior frontal gyrus compared to HC although activity also seems to be lower in SA than PC). These differences in prefrontal cortex activation point toward possible difficulties in regulating emotions and behaviors [85, 86]. Of note, activity in the left (but not right) inferior frontal gyrus during exclusion correlated negatively with depressive symptoms, suggesting that depression might again interfere with regulatory prefrontal control during social interaction. Previous studies using the Cyberball Game also observed alterations in prefrontal cortex activation during exclusion in adolescents with a history of suicide attempt or non-suicidal self-injury [39, 87]. Social processing abnormalities in the prefrontal cortex have also previously been described in SA, for instance while viewing angry faces in adults [88, 89] and adolescents [22]. Interestingly, while neuroimaging during exclusion was able to discriminate patients with and without a history of suicidal acts, this was not the case for the social threat questionnaire, questioning the limits of self-reports.
During the Go-NoGo Task, adolescent SA mainly exhibited greater bilateral activity in the precentral gyri –primary motor brain regions - compared to PC (surviving more stringent correction). We did not observe behavioral differences between SA and PC, only between patients and HC. Activity in these motor regions during the Go-NoGo task did not correlate with impulsivity measured with self-questionnaires. However, they were negatively correlated with response time in our study, suggesting that increased activity may lead to slower response times. The Go-NoGo task specifically demands inhibition of a prepotent motor response [68]. Increased activation in SA may therefore reflect excessive activity to achieve the same outcome, and, therefore indirectly inefficient functioning. Furthermore, these findings corroborate an earlier adolescent study (Pan et al. 2011) which also found that neural activity patterns during response inhibition discriminated between SA and PC, but that behavioral measures did not. Hence, functional markers might be more sensitive than neuropsychological measures.
We found that the activities of left insula during inclusion and right precentral gyrus during inhibition were correlated, suggesting a functional connection between the networks underlying social perception/interactions and cognitive inhibition. Overall, it could be hypothesized that in situation of stress and emotional disturbances (e.g., a depressive episode), this inefficient functioning of brain networks encompassing both the left insula and motor and precentral regions may translate into a lower feeling of social connectedness, lower ability for self-disclosure and help-seeking and inefficient emotional and behavioral regulation facilitating both the emergence of suicidal ideas and acting out. Recent studies have shown that suicidal behaviors are associated with the dysfunctional connectivity of various brain regions [90].
Finally, exploratory results from our study indicate that considering functional alterations related to cognitive and socio-emotional processing significantly enhanced the identification accuracy of adolescents with past suicidal behaviors from PC and SA. According to two classification performance metrics (ROC AUC, classification accuracy), neuroimaging modalities significantly improved baseline models that used sociodemographic data with or without clinical variables. It should be noted, however, that while the most accurate model was provided by the combination of both fMRI modalities without clinical variables (> 90% accuracy, > 95% AUC), the difference between combined models and single modalities was marginal. An important caveat is that the diagnostic validity of these functional markers were tested on the same cohort from which they were derived, which can lead to an overestimation of their effect [91]. Their combination within this sample might not be as effective as their combined use in an independent sample, where each neuroimaging modality might separately be less diagnostically efficacious [92]. We also found that an SVM-based learning method did not outperform conventional logistic regressions. While non-linear kernels with SVM, or other supervised learning methods such as decision trees might have provided better fit [78], logistic regressions already provided high diagnostic utility. Deep learning methods are promising methods to integrate neuroimaging and clinical data to assess suicide risk, but they require large-size data for training [93].
We must highlight a few limitations of the current study. Firstly, even with a substantial number of participants with past suicidal behaviors, we may have lacked the statistical power to detect effects at more conservative levels of correction [94]. Our main results are based on a voxel-wise threshold of p < 0.005, which carries a higher likelihood of false-positives than a typically recommended voxel-wise threshold of p < 0.001 [95]. However, supplementary analyses at a threshold of p < 0.001 yielded significant and concordant results across all contrasts. Besides, these analyses did not account for the variability of cluster-size threshold across all brain regions. While our strategy to simulate noise volume with a mixed-model ACF is deemed acceptable, newer methods using randomization and permutation are considered more accurate at controlling false positive rates [96]. Second, participants had a wide age range (11–18), which probably introduced developmental variability in our analysis despite controlling for age. We previously reported structural brain differences across ages in this sample [58]. Third, despite recruiting in a relatively homogeneous second-line psychiatric clinic, several clinical heterogeneities still characterized our sample, including features of suicidal behaviors, psychiatric comorbidities, and environmental factors. These heterogeneities may prevent generalization to different populations. Finally, while controlling for medication status did not change our findings, we did not account for more complex pharmacological effects, such as dose effects, duration of treatment, and interactions.
In conclusion, our study revealed discrete sets of functional alterations associated with suicidal behaviors during social interactions and cognitive inhibition, two important features that underlie the development of a suicidal crisis. These alterations may furthermore enhance diagnostic identification of adolescents with previous suicide attempts. These findings highlight the complex mechanisms underlying suicidal behaviors. The capacity of neural measures to predict suicidal behavior beyond clinical and sociodemographic variables will have to be tested in large-scale longitudinal studies.