How could brain fingerprinting lead to the early detection of mental illness in adolescents and what are the next steps?

KEYWORDS:

• Artificial Intelligence
• Brain fingerprinting
• Early intervention
• Future application
• Mental illness
• Neuroimaging
• Youth mental health

1. Introduction – Brain fingerprinting as a predictive technology

‘Brain fingerprinting’ is a neuroimaging technique that determines the individual uniqueness of brain activity or structure. Miranda-Dominguez et al. [1] and Finn et al. [2] were the first to show that functional connectivity profiles are unique and reliable, at the individual level. Unique functional synchronization patterns (‘functional connectomes’) index brain activity across various networks or circuits in the brain and are thought to underpin behaviors. Finn et al. [2] demonstrated that functional connectome fingerprints reliably predict individual cognitive attributes. In their study of adolescents, Kaufman et al. [3] found that connectome uniqueness (labeled ‘distinctiveness’) increases with age, however those with lower uniqueness had increased mental health problems (MHPs) and a subsequent delay in connectome uniqueness (more on this below). Such evidence suggests great promise of ‘connectotyping’ as a predictor of mental health outcomes, and potentially, a key component of future personalized mental health care [1–3].

Recently, Shan et al. [4] were the first to show that the uniqueness of an adolescent’s brain fingerprint can predict a mental health outcome. More specifically, low uniqueness of the cingulo-opercular network (implicated in goal-directed behavior) was significantly associated with subsequent increased psychological distress (four months later). Shan et al. [4] hypothesized that maturational delays in the ‘fine tuning’ of such executive function networks may lead to increased MHPs. This was evidence of a brain marker for early detection of psychological distress before behavioral abnormalities manifest, which is critical for early intervention of mental disorders.

The ability to reliably predict who may develop a mental disorder is vital when we consider the increasing MHPs, and burden faced by society. Most mental disorders first emerge in youth and those with childhood/adolescent onsets tend to be more severe, are frequently undetected early, accrue additional comorbidity, and have an economic cost ~ 10 times higher than adult-onset illnesses [5]. In youth, mental health and substance use disorders are the leading causes of disability, accounting for 25% of all years lived with disability [6]. In spite of the best efforts of clinicians and researchers, it is still not fully understood why some individuals develop mental disorders while others do not. However, changes in the brain are more than likely to provide strong clues about mental health outcomes. Indeed, the adolescent brain in particular is important in this pursuit as there are rapid and dynamic changes during this period, sculpting individual uniqueness. MHPs likely stem from aberrations/exaggerations in normal maturational brain changes [7]. Such anomalies act in concert with other biological, psychosocial and environmental factors which can then manifest as emerging mental disorders, if left unchecked [7].

By tracking brain changes as they occur, we can proactively tackle emerging MHPs in adolescence and make targeted early interventions. The challenge is accurately predicting the likelihood of a mental disorder, well before it happens. The major rationale for the ‘Longitudinal Adolescent Brain Study’ (LABS) is to track brain changes during adolescence, to gain a deeper understanding of the factors that impact mental health and wellbeing [4,8]. LABS researchers conduct assessments and brain scans in youth from the general population every four months, for five years, starting at age 12; data collected is detailed and its ‘temporal richness’ (many scans over shorter periods of time, for each individual) allowed Shan et al. to conduct their predictive ‘brain fingerprinting’ study [4].

The ‘Adolescent Brain Cognitive Development’ (ABCD) study has the potential to expand on this, as it also tracks brain and mental health changes in youth as they progress into adulthood [9]. The ABCD starting age is 9–10 years and sites involved conduct brain scans every 2 years. According to the latest data release [https://abcdstudy.org/] there is baseline and early longitudinal data from ~ 11,800 participants. Hopefully, replication of brain fingerprinting findings [4], as well as discovery of other brain biomarkers that can reliably predict specific mental health outcomes in youth, will emerge in the not-too-distant future. ABCD and LABS are complementary and synergistic studies; while the former includes a large sample size to capture the general developmental trajectory over two-year periods, the latter has less age variance and monitors brain changes at shorter, fixed intervals.

Increasingly, brain imaging research is utilizing artificial intelligence (AI) to discover novel patterns linked to MHPs. Rahman et al. [10] implemented sophisticated machine learning (ML) models in large control datasets to learn to understand brain patterns and then applied these models to smaller datasets of individuals with autism, schizophrenia, or dementia to identify new brain patterns linked to each disorder [10]. Their ‘deep learning’ (DL) models accurately distinguished patient groups from controls and discriminated when such distinctions occurred. The general premise is that neuroimaging may identify when relevant, predictive patterns first emerge before disorder/disease is apparent. Similarly, there is evidence showing that different aspects of connectome fingerprints relate to different time scales [11].

As suggested above, brain fingerprinting is anticipated to be a key component of the personalized mental health landscape. In the context of youth mental health, such AI-based predictive technology could greatly help young people and their families make informed decisions about potential risk for MHPs – working with clinicians to make early interventions and track brain changes that ensue. The evidence would suggest that an ideal time to undertake such predictive measurement is at the beginning of adolescence, around the time of puberty. Such screening could take place at the population level, that is, all individuals at ~12 years of age are provided with the opportunity to undertake a government-funded brain scan utilized to predict mental health risks. This concept is not dissimilar to something like bowel cancer screening for all individuals aged 50–74 years [12].

2. The next steps

For brain fingerprinting to succeed as a predictive technology for mental health outcomes the following steps are needed. First, is to obtain a high level of confidence of its predictive capacity – this will only come from more research and replication of findings. Clearly, more longitudinal research is required to address this, as the only way to predict mental health outcomes from brain imaging data is via studies that assess these measures throughout the childhood into adolescent years. Regarding the generalizability of brain fingerprinting findings, there are some very important opportunities available now – by applying a fundamental principle of AI/ML approaches – that is, to utilize one study for a training dataset and another comparable study for the testing dataset [13]. Unfortunately, to date, very few studies have subjected their brain models to this type of approach [14]. For example, researchers could determine brain fingerprinting-behavior relationships in a ‘sample-rich’ study (such as ABCD), and then track these in a separate, ‘temporally-rich’ study (such as LABS). More specifically, functional connectomes found to predict mental health outcomes two-years later via ABCD could then be implemented and tracked via LABS to determine more fine-grained brain changes and corresponding fluctuations in mental health over the same period. This would help to determine the reliability of brain fingerprints as well as their stability over time. As more longitudinal neuroimaging studies are undertaken, the breadth and depth of brain fingerprinting research is anticipated to develop to the point whereby the prediction of mental health outcomes will approach substantive, and then meet, clinical utility.

To date, few studies have utilized functional connectome fingerprinting to distinguish those with psychiatric conditions from healthy controls. In the above-mentioned study by Kaufman et al. [3], among young participants with symptoms of attention deficit, depression or prodromal schizophrenia, N = 137 were identified as having markedly increased clinical symptoms (‘general psychopathology’) and compared to healthy controls (N = 153) they showed weaker connectome uniqueness, across all ages (8–22 years). Although this study was based on cross-sectional data, the authors concluded that individuals with increased clinical symptoms had a marked delay in their connectome uniqueness. In a subsequent study [15], the same group evaluated the stability of functional connectome fingerprints in individuals with schizophrenia, aged 16–52 years. Compared to healthy controls (N = 202), those with schizophrenia spectrum disorders (N = 167) showed reduced and delayed trajectories in their full brain connectome stability across the adulthood age range. In a transdiagnostic brain fingerprinting study of adults (aged 18–71 years), Baker et al. [16] compiled cross-sectional functional connectome data from N = 210 participants with a primary psychotic disorder, N = 192 with a primary affective disorder [without psychosis] and N = 608 healthy controls. Essentially, this study revealed impairments in the frontoparietal network, whereby a graded pattern of dysconnectivity was amplified in those with more severe psychopathology (particularly those who were currently treatment seeking) and not necessarily according to the presence of an affective versus psychotic illness, per se. Baker et al. [16] note that case-control, group-level studies examining single psychiatric illnesses in isolation give the illusion of group specificity and likely mask important variations in functional connectomes. Rather, they argue that the field would advance via analyses that specifically link individual functional connectome patterns to their behavioral, symptom and severity profiles [16].

It is also important that other imaging modalities are incorporated into this research. There are already studies demonstrating brain fingerprinting based on structural connectomes. For example, Yeh et al. [17] demonstrated that in N = 213 adults (44% of whom had repeat scans, within 3 months), local structural (white matter) connectome fingerprinting was highly specific and was accordingly described as a reliable measure of the connective architecture of individual brains. Yeh et al. [17] suggest that their local connectome fingerprinting approach is complementary to region-to-region connectivity approaches, as it provides high dimensional data characterizing structural connectivity at the voxel level. Such evidence suggests that multimodal approaches utilizing functional and structural, as well as neurochemical and electroencephalography (to name a few), measures should be undertaken to identify the best brain fingerprinting profiles predictive of mental health and other behavioral outcomes. According to Sui et al. [14], the integration of multimodal brain data is an effective way to capitalize on the strengths of different neuroimaging modalities, with evidence that multimodal integration increases the accuracy of predicting behavioral outcomes, compared to unimodal data. Furthermore, in addition to general population studies, there is a need for more brain fingerprinting research conducted in clinical samples, particularly with individuals currently accessing youth mental health services. Such studies would enhance the clinical utility of brain fingerprinting and they could also expand on its application as a theranostic biomarker [18], since these samples would be experiencing (to a greater extent) the use of psychotropic medications.

There are numerous factors to consider when anticipating the future application of brain fingerprinting to predict mental health outcomes. A factor key to its widespread uptake and use is equitable access to brain scans across societies [19]. Linking brain fingerprinting techniques to other technologies, such as electroencephalography, will help address access and affordability, as well as allow the application of ‘wearables’ [20] that could be used to track changes in brain patterns and corresponding changes in mental health and wellbeing across different time scales. Another issue to address in the context of AI-based predictive technologies in mental health is the appropriate attribution of doubt in contemporary and emerging ML algorithms, which are increasingly playing major roles in biomedical science. A predictive algorithm is contingent on its source dataset and the way it learns from this ‘ground truth’ data may be influenced by human rules (supervised ML) or not (unsupervised ML). While there are strengths and weaknesses of each, there is an argument for the ‘human-in-the-loop’ approach whereby there may be a need for an ethical guarantor [21]. In the context of predicting mental illnesses in youth, researchers utilizing techniques such as brain fingerprinting should iteratively and recursively experiment, refine and include doubtful accounts of predictive models [21]. Increasingly, DL methods are being applied to neuroimaging data and employed to predict individual differences in behavior [14]. The advantage of DL is that it automatically discovers (learns) the representations needed for feature detection or classification, typically from raw (or less processed) data. Thus, DL operates at a more abstract level and as such it has the capacity to identify neural features that may be missed by standard ML methods; seemingly better at detecting subtle or complex brain signatures [14]. A potential limitation of DL approaches is the criticism that it requires very large datasets in order to outperform standard ML approaches. However, recent research by Abrol et al. [22] demonstrates that, when utilized properly (i.e. trained following DL practices, with evaluation of findings on secondary data sets), DL consistently outperforms standard ML in larger datasets, and with small sample sizes (n = 50, n = 100) it performs at the same level as standard ML. Abrol et al. [22] strongly recommend that neuroimaging researchers utilize DL methods and suggest that since it has the combination of superior representational learning and finer interpretability it has great potential in learning specific changes in the brain that relate to the characterization of mental illnesses. There are some promising signs of such application, with evidence that a DL pipeline specifically developed for functional connectome fingerprinting can achieve better classification than the Pearson correlation approach commonly used [2,4].

3. Conclusions

Neuroimaging and other technologies, available now, have great potential in predicting who may develop mental illness. Initial findings of brain fingerprinting predicting cognitive [2] and mental health outcomes [4] are major and exciting advances. Replication studies, via extensive longitudinal research that focusses on mental health outcomes, are urgently needed. For the personalized predictive approach to flourish, brain fingerprinting should be combined with multiple measures of mental health, as well as cognition and functional measures. Data-driven analyses, via AI/ML/DL approaches, are likely to reveal the most important patterns of brain and behavior that best predict an individual’s future mental health. The utility of brain fingerprinting as a predictive technology for mental health outcomes in youth can only be achieved if there is also careful consideration of the ethical and scientific issues surrounding data-driven approaches and hence the anticipatory governance of this going forward is vital.

Declaration of interest

The author has no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.

Reviewer Disclosures

Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Additional information

Funding

This article is supported by the Prioritizing Mental Health Initiative from the Australian Commonwealth Government.