The main goal of this study was to determine whether machine learning and fMRI could help to differentiate, at an individual level, healthy adolescent offspring at genetic risk for bipolar disorder and other Axis I psychiatric disorders from healthy adolescents at low risk of developing bipolar disorder or other Axis I psychiatric disorders. We also investigated whether the predictive probability of those healthy at-risk adolescents who subsequently developed a future psychiatric illness were statistically different from the predictive probability of those at-risk adolescents who did not develop such a disorder in longitudinal clinical follow up. Our findings indicate that machine learning combined with fMRI helped to discriminate healthy low-risk control adolescents from healthy adolescents at genetic risk of future psychiatric disorders. Our findings also indicate that the magnitude of the predictive probabilities for group classification that were derived from these techniques could potentially be used as a score to predict which at-risk adolescents subsequently went on to develop an Axis I psychiatric disorder, namely mood and anxiety disorders.
The advantage of pattern recognition techniques such as the one we employed in the present study is that they provide information at the individual – rather than the group - level (GPC based on whole brain neuroimaging data). Specifically, our findings from GPC indicate that we can make predictions at the individual level considering the discrimination between who are genetically at risk from who are not. Furthermore, ROC analysis with predictive probabilities derived from GPC suggests that pattern recognition techniques such as GPC have the potential in the future to help identify which at-risk adolescents are most likely to develop future Axis I disorders.
Recent studies demonstrated the utility of pattern recognition approaches in helping with classification of different psychiatric disorders, including Alzheimer's diseases and autism 
. One study 
evaluated early recognition and disease prediction using multivariate pattern classification, and demonstrated that this approach could be used to predict transition to psychosis. Until now, however, it was unknown whether the method could distinguish completely asymptomatic, genetically at-risk individuals from healthy, low-risk individuals.
In the present study, none of the at-risk adolescents developed bipolar disorder following the fMRI scan. Over half of the at-risk adolescents developed an affective disorder (anxiety or depression), however. It is possible that the prediction findings in the present study may be attributed to similarities across affective disorders with regard to abnormal patterns of activation to emotional facial expressions relative to age-matched healthy controls. However, the fact that healthy at-risk adolescent who went on to develop these disorders have higher predictive probabilities is noteworthy in light of evidence indicating that major depressive disorder and anxiety disorders often emerge prior to the onset of mania and episodes of depression characterizing bipolar disorder 
. For example, recent evidence from the Bipolar Offspring Study suggests that a larger number of bipolar disorder episodes of offspring in that study started with depressive episodes 
. Such evidence supports previous findings indicating that depressive symptoms emerge prior to the onset of bipolar disorder symptoms 
, particularly if there is family history of bipolar disorder 
. With regard to anxiety disorders, some studies suggest that anxiety symptoms often precede and may hasten the onset of bipolar disorder in adults 
. For example, prospective data from a large community sample suggested that individuals who reported experiencing anxiety as adolescents were at increased risk of developing bipolar disorder as adults 
, which is consistent with previous findings 
. Ongoing prospective follow-up of the at-risk adolescents in the current study will help to further elucidate the role of these psychiatric disorders in the developmental course of bipolar disorder. Identifying differences in patterns of neural activity to emotionally salient information in these at-risk adolescents can contribute valuable information to this research.
Gaussian process classifiers are discriminative approaches and therefore are able to find a discriminating boundary between two classes (e.g. healthy vs. a patient group), and then use this information to classify new individuals. Discriminative models, however, should not be confused with statistical approaches based on mean group differences, such as the General Linear Model 
. GLM analyses treat every voxel independently and extract measures of interest from them, such as the average response during a particular experimental condition. More specifically, the GLM approach searches for voxels whose activation time series is well reconstructed by the combination of regressor time series related with each experimental condition and some noise terms. The analysis of fMRI data with discriminative models differs from the traditional GLM analysis by investigating a different question. Instead of finding voxels whose time series respond to a specific experimental condition, such models ask whether it is possible to make a prediction about a variable of interest (e.g. patients vs. controls or task 1 vs. task 2) based on the pattern of activation over a set of voxels 
. Furthermore, discriminative models provide a map that shows the discriminating boundary
between the different groups. Specifically, a high value in a particular voxel indicates a strong contribution to the discrimination boundary, but does not necessarily imply greater activity in one group versus another. In summary, pattern recognition approaches such as GPC are multivariate techniques, where discrimination is based on the whole pattern rather than on regional activity, which is typically reported in the traditional GLM-based analyses comparing psychiatric patient and healthy control groups. Based on the GPC used in the current study, the spatial pattern that best discriminated at-risk adolescents vs. healthy controls included ventromedial prefrontal cortex and superior temporal sulcus, which are key regions supporting emotion regulation and face processing, and are regions that have been shown to be functionally abnormal in individuals with bipolar, and other mood disorders 
Interestingly, the best discrimination between at risk and low-risk adolescents was found to be neutral faces presented in the happy face experiment. Furthermore, the fact that there were no significant findings for mild or intense happy faces, or for any of the faces in the fear face experiment, suggested that accurate classification was specific to neutral faces presented in the context of happy faces, and not generalizable to the other emotional faces. Neutral faces especially are often perceived as ambiguous and potentially threatening by individuals diagnosed with anxiety or mood disorder 
. One study, for example, reported abnormally elevated subcortical activity to neutral faces in youth with bipolar type I disorder, particularly in those who perceived these faces as threatening 
. Another study reported that depressed patients did not differ from healthy controls in their ability to accurately recognize sad and happy facial expressions, but they were less accurate at recognizing neutral expressions 
. Specifically, depressed patients misclassified a higher number of neutral expressions as sad, suggesting a negative interpretative bias. Other studies used neutral facial expressions as a control condition for other emotional facial expressions, and also found evidence of a negative emotional interpretative bias in depressed patients 
, or found that depressed patients were slower to respond to neutral expressions compared with emotional expressions 
Findings from the behavioural data analyses indicate that there were no differences between groups, or group by face condition interactions, neither in accuracy nor for reaction times. This is interesting because it shows that behavioural differences cannot explain the results found in the present study, indicating that the classification results are related to the underlying neural circuitry which is already different in at-risk individuals as compared to healthy individuals. Furthermore, the absence of behavioural differences between these groups suggests that there was no impairment in the performance of the at-risk adolescents indicating that those adolescents were healthy during the neuroimaging experiment. There was, however, a main effect of face condition, such that all adolescents were less accurate in gender labeling neutral faces. Performance on the out-of-the scanner emotion labeling task also indicated that all adolescents were less accurate in labeling neutral faces relative to fearful and happy faces. These behavioral data are consistent with previous findings that children and adolescents find neutral faces more ambiguous and more difficult to identify than emotional faces 
. Taken together, these findings suggest that neutral faces may have been more difficult to label in all adolescents in the present study, and that this greater level of difficulty in perceptual discrimination may have warranted greater recruitment of neural regions, including ventromedial prefrontal cortex and superior temporal sulcus, which contributed to the classifier that differentiated the two groups. Moreover, at-risk adolescents may have perceived neutral faces presented in the context of happy faces not only as ambiguous and “non-neutral” but potentially as more threatening than did healthy controls, which would have influenced their pattern of activation in these key neural regions. This interpretation would be consistent with the idea that other emotions can influence the interpretation of neutral faces 
, and may help to explain why groups were classified based on recruitment of key neural regions implicated in face processing and emotion regulation. Nevertheless, including subjective emotional ratings of neutral and other emotional facial expressions in future studies may help elucidate these findings further.
The following limitations to the current study merit some discussion. Although the sample size in the current study was sufficient to provide adequate power to train a GPC to discriminate between adolescents at risk for mood disorders vs. adolescents at low risk, further analyses to distinguish between those who developed vs. those who not develop mood disorder could be conducted only on a subset of the at-risk adolescents who had complete diagnostic interview data. Our findings do, however, provide a rationale for future studies, with larger samples of converters and nonconverters from the at-risk group, to examine the extent to which pattern recognition techniques can identify at the individual level those at-risk adolescents who are most likely to develop in the future different Axis I disorders.
In addition, while the leave-one-out cross-validation is the recommended technique for evaluating classifier performance on small samples due to its almost unbiased estimation of the true error rate it has high variance for small sample sizes. Therefore our results should be validated using independent and bigger samples. Nevertheless, our findings are an important first step toward the ultimate goal of using neuroimaging to help predict future clinical course in at-risk adolescents, and highlight the utility of combining neuroimaging and machine learning techniques to identify neuroimaging measures that may ultimately be able to act as predictors of future onset of psychiatric disorders in at-risk adolescents. We recognize the importance of other risk factors (e.g., psychosocial functioning) in the development of psychiatric disorders such as bipolar disorder. Additionally, it is possible that the environmental effects of being raised by a bipolar parent were a potential confound in the present study. Future studies with larger samples will allow us to integrate these factors and examine these more complex prediction models. In future work we also aim to investigate whether other fMRI tasks could lead to better discrimination between the groups and also strategies to combine different information into the model (e.g. different imaging modalities, clinical and behavioural information).
In summary, our findings indicate that the combination of machine learning and neuroimaging have great potential, especially in situations where there is limited clinical and genetic information, to help to identify which individual at-risk adolescents are at true risk of developing future Axis I disorders. This in turn can help guide early and appropriate interventions for these adolescents and their families, to relieve the significant psychological problems associated with lack of knowledge about the future likelihood of psychiatric disorders in individual at-risk adolescents.