Approaches to Defining Common and Dissociable Neurobiological Deficits Associated with Psychopathology in Youth

Abstract

Psychiatric disorders show high rates of co-morbidity and non-specificity of presenting clinical symptoms, while at the same time demonstrating substantial heterogeneity within diagnostic categories. Notably, many of these psychiatric disorders first manifest in youth. Here we review progress and next steps in efforts to parse heterogeneity in psychiatric symptoms in youth by identifying abnormalities within neural circuits. To address this fundamental challenge in psychiatry, a number of methods have been proposed. We provide an overview of these methods, broadly organized into dimensional vs. categorical approaches and single-view vs. multi-view approaches. Dimensional approaches including factor analysis and canonical correlation analysis aim to capture dimensional associations between psychopathology and brain measures across a continuous spectrum from health to disease. In contrast, categorical approaches such as clustering and community detection aim to identify subtypes of individuals within a class of symptoms or brain features. We highlight several studies that apply these methods to samples of youth, and discuss issues to consider when using these approaches. Finally, we end by highlighting avenues for future research.

1. Introduction

Youth, which we define broadly as childhood, adolescence, and young adulthood, is a period during which many psychiatric disorders first manifest (1, 2). It is also a time of marked development in brain structure and function (3, 4). Despite dramatic advances in neuroimaging methodology that have made it possible to measure the structure and function of neural circuits, attempts to isolate neurobiological substrates of psychopathology have been hampered by the simultaneous comorbidity among and heterogeneity within psychiatric diagnoses. For example, psychiatric disorders share many presenting clinical symptoms (5) which likely contributes to the apparent non-specificity of neural mechanisms associated with psychopathology (6–13). At the same time, there is considerable heterogeneity in the presentation of clinical symptoms (14–17). There is increasing interest in understanding clinical heterogeneity in psychopathology in terms of underlying biological mechanisms (18), which may provide the basis for a biologically-based nosology for mental disorders. However, thus far, the majority of this work has been restricted to adult samples.

The goal of this review is to introduce different approaches to understanding neurobiological heterogeneity within psychiatric disorders, with a focus on studies using these approaches in samples of youth. Specifically, we concentrate on methods for defining common and dissociable neurobiological deficits associated with psychopathology. Notably, the focus of this review is not on trajectories of developmental change or theoretical nosological debates. For an overview of patterns of brain development using longitudinal studies, we refer to the review by Becht and Mills in this special issue. For current nosological debates, Bornovalova et al. in this special issue provide a comprehensive review of issues concerning the bifactor model. Instead, our goal is to introduce the reader to a range of approaches for parsing common and dissociable neurobiological heterogeneity, while highlighting the similarities and differences between the methods. Of note, we focus only on methods that have been applied in samples of youth. While we do not provide a comprehensive tutorial on any particular method, we refer the reader to more detailed treatments of these methods when available. We organize this review according to two axes that aim to understand the complex mapping between neural deficits and clinical symptoms: dimensional vs. categorical approaches and single-view vs. multi-view approaches. We begin by defining the four broad approaches, and then highlight studies that apply these methods to samples of youth. We follow this with a discussion of the issues to consider when using these approaches, and end by discussing considerations for future research.

2. Dimensional vs. categorical and single-view vs. multi-view approaches

When reviewing methods for parsing heterogeneity, one important distinction is between dimensional vs. categorical approaches. Here. we are referring to whether the approach produces dimensions (continuous variables) or categories (clusters or subtypes) of the measure of interest. Specifically, dimensions represent the loadings onto symptoms and each individual receives a dimensional score, while categories represent subtypes of people who share features in common and each individual is classified into a category. Research has shown that many psychiatric symptoms exist on a continuum, with diagnosable psychopathology being an extreme phenotype of variation that is present in the general population (19). Dimensional approaches are able to account for this continuous spectrum from health to disease including subthreshold levels of psychopathology. However, multiple mechanisms may drive the extreme phenotypes of psychopathology and one can also use categorical approaches to identify subtypes of psychopathology.

A second important distinction is between single-view and multi-view methodological approaches. This refers to the nature of the input data, which can be symptoms, brain features, or both. Any data type can be considered, but here we limit our review to clinical symptoms and neuroimaging measures given our focus on neurobiological heterogeneity in psychopathology. Single-view approaches consider data from a single feature set (e.g., symptoms). The output can be dimensional (a spectrum of symptoms) or categorical (subtypes of people), but the input reflects a single data type. In contrast, multi-view approaches use input data from multiple feature sets, such as integrating both symptoms and brain features (20). Again, the output can be dimensional (continuous dimensions representing combinations of symptoms and brain features) or categorical (subtypes characterized by different combinations of symptoms and brain features). Collectively, we can organize the approaches discussed into this dimensional vs. categorical and single-view vs. multi-view framework.

2.1. Single-view dimensional approaches

Single-view dimensional approaches take as input a single feature set (clinical symptoms or brain measures) and produce dimensional output. One use of this approach is to reduce a large feature set into a smaller number of latent summary variables for discovery of hidden relationships and/or for data reduction. For example, approaches such as independent component analysis (ICA), principal component analysis (PCA), and non-negative matrix factorization (NMF) can be used to reduce high-dimensional symptoms or brain features into a smaller number of components (21–31). Another common approach involves factor analysis, which summarizes a large number of psychiatric symptoms into latent dimensions that can then be related to various neurobiological measures (32–39). Two models for this purpose include correlated traits and bifactor models (40). Correlated traits models (e.g., factors from an exploratory factor analysis) produce correlated symptom factors. In contrast, bifactor models reveal a hierarchical structure of symptoms including a general psychopathology (p) factor that represents the overall burden of psychopathology across disorders (Figure 1) (41). Akin to the “g” factor in general intelligence, the p factor represents the symptoms that psychiatric disorders share in common (42). In addition to p, a bifactor model identifies uncorrelated subfactors of psychopathology, such as factors for internalizing/fear, anxious-misery/distress, externalizing/behavioral, and psychosis/thought disorder (32–39).

Figure 1. Examples of representative approaches: single-view vs. multi-view and dimensional vs. categorical.

Schematic illustrating the four broad approaches surveyed in this review. A) The input data for each approach may include clinical symptoms, imaging features, or a combination of the two. B) Each approach is illustrated with an exemplar technique (note that many other methods are available as well). Single-view dimensional approaches (e.g., bifactor model) take as input a single data type, such as clinical symptoms, and output latent dimensions that summarize the data. Multi-view dimensional approaches (e.g., CCA) take as input two data types and identify linear combinations of the two. Single-view categorical approaches (e.g., clustering) find subtypes based on a single feature set. Multi-view categorical approaches (e.g., multi-view biclustering) find subtypes based on multiple views of the data, where the input is two or more feature sets.

The p factor has been associated with a number of neurobiological measures in youth, including reduced gray matter volume (22, 43), reduced activity in executive regions (44), elevated resting-state CBF (45), reduced fractional anisotropy (46), and delay in connectome distinctiveness (47) (Figure 2). In addition, there also exist dissociable deficits specific to the symptom domains of fear, anxious-misery/distress, behavioral/externalizing, and psychosis/thought disorder. For example, factors related to internalizing symptoms are associated with reduced gray matter volume in specific regions (43), widespread hyperactivation of the executive network (44), reduced cortical thickness (22), and specific abnormalities in CBF (45). The bifactor model is useful for summarizing symptoms into dimensions that capture common variance across disorders (p factor) and unique variance within specific classes of symptoms. The studies applying this method to samples of youth and then relating these factors to neuroimaging measures suggest that there may exist both common and dissociable neurobiological substrates of psychopathology in youth.

Figure 2. General psychopathology is associated with common neurobiological deficits in youth.

Single-view dimensional approaches that identify a general factor common across disorders in youth reveal: A) globally reduced gray matter volume (22); B) bilateral gray matter volume reductions in regions chosen a priori including dorsal prefrontal cortex (DPFC), ventrolateral prefrontal cortex (VLPFC), and orbitofrontal cortex (OFC) (43); C) reduced activation in regions within the cingulo-opercular control network during an N-back working memory task (44); D) reduced fractional anisotropy in the genu and body of the corpus callosum (46); E) elevated cerebral blood flow in the dorsal and rostral anterior cingulate (45); and F) delay in connectome distinctiveness compared to healthy controls across the whole brain (47). All figures reprinted with permission (Copyright ©2017, Nature Neuroscience, Springer Nature; Copyright ©2019, Behavioural Brain Research, Elsevier; Copyright ©2017, Clinical Psychological Science, SAGE Publications; Copyright ©2016 and ©2019, American Journal of Psychiatry, American Psychiatric Association. All Rights Reserved).

2.2. Multi-view dimensional approaches

While single-view dimensional approaches only consider a single data type as input, multi-view dimensional approaches take as input two or more feature sets (symptoms and brain) and produce dimensional summaries of the interrelationships between the feature sets. Two commonly used multi-view dimensional methods are partial least squares (PLS) regression and canonical correlation analysis (CCA) (48–51). Both methods seek to find linear combinations of brain features that predict linear combinations of clinical symptoms (Figure 1). A growing number of studies have used PLS regression and CCA to link neurobiological measures to psychopathology in adults (52–66).

In youth, CCA has been used to link functional connectivity patterns to behavioral measures including demographics, IQ, and a variety of self-report measures (67). However, the participants of this study included healthy youth (n=281) and a much smaller number with major depression (n=25); as a result, there was relatively limited psychopathology present. It should also be noted that the assumptions of CCA often do not hold in high-dimensional imaging data (68). Alternative methods such as sparse CCA (sCCA) are able to overcome these limitations in high-dimensional data by simplifying the model to avoid overfitting and increase interpretability (69–71). Our group has used sCCA to link dimensional psychopathology symptoms to functional connectivity measures in a large sample of 663 youth (72). The results revealed that mood, psychosis, fear, and externalizing behavior were associated with distinct patterns of connectivity, while loss of network segregation between the default mode and executive networks was common across these dimensions (72) (Figure 3). Methods such as PLS regression and CCA are useful for measuring brain-behavior relationships. However, compared to adults, there have been relatively few studies applying these methods in younger populations, suggesting the need for additional studies in samples of youth in order to replicate these results.

Figure 3. Sparse CCA links dimensions of psychopathology to functional connectivity patterns in youth.

A) sCCA reveals three symptom dimensions (mood, psychosis, and fear) that were statistically significant, with the fourth dimension (externalizing behavior) showing an effect at uncorrected thresholds. B) Scatter plots showing linear combinations of functional connectivity and psychiatric symptoms demonstrate the correlated multivariate patterns of connectomic and clinical features. C) Connectivity-informed dimensions of psychopathology cross clinical diagnostic categories. Specifically, the mood dimension was composed of a mixture of depressive symptoms, suicidality, irritability, and recurrent thoughts of self-harm. The psychotic dimension was composed of psychosis-spectrum symptoms, as well as two manic symptoms. The fear dimension was comprised of social phobia and agoraphobia symptoms. The externalizing behavior dimension showed a mixture of symptoms from attention-deficit/hyperactivity and oppositional defiant disorders, as well as irritability from the depression section (72). Reprinted with permission (http://creativecommons.org/licenses/by/4.0/).

2.3. Single-view categorical approaches

Single-view categorical approaches take a single modality as input (symptoms or brain features) and produce clusters or subtypes of individuals. Notably, these techniques may also be used to discover clusters of variables; however, here we focus on the use of these methods to reveal clusters of individuals with common features. While many studies have clustered on psychiatric symptoms, behavioral measures, or neuropsychological performance in both adults and youth (73–93), here we focus on studies that cluster based on imaging measures. A number of studies have identified subtypes based on imaging features in adults (60, 94–108).

Neurobiological subtypes of internalizing/externalizing symptoms have been identified in samples as young as infants (109). Additionally, community detection has been used to identify subtypes of youth with ADHD using measures of intrinsic functional connectivity (110, 111). Our group has used a recently developed machine learning method, HYDRA (94), in 1,141 youth to identify two subtypes of internalizing youth differentiated by abnormalities in brain structure, function, and white matter integrity, with one subtype showing poorer functioning across multiple domains (112) (Figure 4). Unlike many clustering approaches, HYDRA accounts for variability in the controls while uncovering subtypes in the symptomatic group. These approaches are potentially useful because they seek to identify subtypes that “carve nature at its joints” based on underlying neurobiology, rather than relying on symptom measures. Studies using this approach in samples of youth are just beginning to emerge.

Figure 4. HYDRA identifies two neurostructural subtypes of internalizing youth.

A) A linear support vector machine (SVM) is a discriminative classifier defined by a separating hyperplane, shown here as a gray line. Linear SVMs fail to capture the heterogeneity that exists in the patients, indicated by blue, green, and purple subtypes within the cases. B) Conversely, methods such as HYDRA (94) can estimate multiple linear hyperplanes (green lines) whose segments separate the clusters of cases from the controls. This approach makes HYDRA more flexible than SVMs, facilitating the identification of heterogeneous subtypes of patients. C) HYDRA identified two subtypes of internalizing youth with a high degree of reliability. Subtype 1 and 2 both showed significant psychopathology compared to typically developing youth. D) Subtype 1 showed significantly worse performance than the other two groups on cognitive measures, especially executive functioning tasks. E) In terms of structural measures, subtype 1 showed smaller volumes than the other two groups consistently across the brain. F) Subtype 1 also showed reduced cortical thickness in most regions. G) Subtype 1 demonstrated reduced resting-state ALFF (amplitude of low-frequency fluctuations) in frontal regions, the right amygdala, and the right hippocampus. H) Finally, subtype 1 showed reduced fractional anisotropy in a number of white matter tracts. Taken together, this study showed that clustering approaches can be used to identify reliable subtypes of internalizing youth, with subtype 1 showing greater deficits across symptoms, cognition, and brain structure (112).

2.4. Multi-view categorical approaches

Multi-view categorical approaches take as input two or more feature sets (e.g., symptoms and brain measures) and produce clusters or subtypes characterized by different combinations of those features (20). Biclustering clusters on both rows (subjects) and columns (features) simultaneously to generate clusters representing subsets of subjects related to subsets of features (113). While still a single-view approach, this method can be adapted to be multi-view by including more than one feature set. Multi-view biclustering has been used to simultaneously cluster symptoms and brain features in adults with a range of psychopathology (114–116). There are currently no studies applying multi-view biclustering to samples of youth, suggesting a promising area for future work.

A method conceptually to related multi-view biclustering is Similarity Network Fusion (SNF). SNF is a multi-view approach that creates networks of individuals based on each feature set separately and then integrates these into a single network (117). SNF has identified clusters of youth with schizophrenia-spectrum, autism-spectrum, or bipolar disorder using demographics, brain imaging, and behavioral data (118). Methods that consider multiple features sets will likely better represent the complex interactions that exist between clinical symptoms and biological data; as such, this approach may have great relevance to studies of brain development in youth.

3. Methodological considerations in studies of heterogeneity

3.1. Data and subject inclusion

Single-view and multi-view approaches share several issues in common across both dimensional and categorical methods. First are the issues of which subjects and what data (imaging features or symptoms) to include. This is especially important since the results will inevitably depend heavily on the input data. For example, bifactor studies that do not measure psychosis-spectrum symptoms will not find a thought disorder factor. Any variance associated with psychosis-spectrum symptoms may be aliased into the remaining factors, possibly impacting the dimensions or clusters discovered. Likewise, many clustering methods require defining a patient group for clustering; thus, how the sample is defined (e.g., transdiagnostic or only a particular disorder) will have a large impact the clusters found. Specific to clustering, there is also debate as to whether healthy participants should be clustered separately or in combination with patients (19, 79). Despite evidence of neurobiological heterogeneity within healthy controls (79, 110, 111, 119), they are often treated as a homogeneous group (120). Taken together, it is important to carefully consider which individuals and measures will be included for all methods discussed.

3.2. Choice of approach

A second common issue across approaches is the existence of multiple ways to parse heterogeneity in clinical groups depending on the approach chosen (19). While this was noted by Marquand et al (19) in regard to clustering methods, this issue also applies to dimensional approaches as well. There are many supervised, semi-supervised, and unsupervised algorithms for clustering, and while beyond the scope of this review, each has its own strengths and limitations (121). Different algorithms may suggest different clustering solutions, and there is currently no strong consensus on the best method for choosing the optimal number of clusters (19). Similarly, there is considerable debate regarding the appropriate method to delineate psychopathology factors using dimensional approaches. These methods differ to the degree to which they are theory-driven (confirmatory) or data-driven (exploratory), which will impact the interpretation of the resulting factors. Bornovalova et al. provide a comprehensive review of these issues as they relate to the bifactor model in this special issue. As Feczko and colleagues (119) point out, the goal or question at hand should drive methodological choices, as different approaches may be valid for different purposes (e.g., differentiating subtypes of patients, investigating symptom latent structure, predicting treatment response, etc.).

3.3. Consideration of covariates

Also related to the approach used is the issue of whether or not to control for covariates. This is not commonly discussed in single-view dimensional approaches such as factor analyses of symptom data, and it remains an open question as to whether it is necessary or even desirable to consider covariates in approaches that only examine symptoms. However, controlling for covariates in multi-view approaches that include both symptoms and neurobiological measures is important, as there are known confounds associated neuroimaging measures. In cross-sectional data spanning a large age range in youth, both age and sex are commonly controlled for (see Future Directions for a discussion on the need for longitudinal designs), as well as motion or data quality. Controlling for covariates may be especially important for methods that cluster based on neurobiological measures, as clustering algorithms may inadvertently produce clusters based on irrelevant variables with known relationships with brain features (e.g., clusters separated by younger and older individuals, or by females and males). For many approaches, covariates can simply be regressed out of the features of interest prior to clustering or subject-level weighting can be used to address confounds (122). Additionally, multi-view factor analytic approaches that consider covariates have been developed (123).

3.4. Issues regarding sample size

An additional consideration that applies to all approaches is the issue of adequate sample size. The methods reviewed typically require very large sample sizes to increase the stability of the results. Some methods are constrained by sample size. For example, factor analytic approaches and CCA are limited by the ratio of observations (subjects) to model features (variables), with more observations than variables required (51). While clustering methods and multi-view techniques like PLS may produce results in smaller samples that have many features, the results may have poor stability, generalizability, or both. Cross-validation and replication in independent samples can increase our confidence in the results, but may require even greater sample sizes and access to independently collected datasets with comparable measures. Leveraging large, publicly available neuroimaging studies will be especially helpful moving forward. Available resources include the NKI-Rockland Sample (124), the Philadelphia Neurodevelopmental Cohort (125), the Pediatric, Imaging, Neurocognition, and Genetics study (126), the Human Connectome Project-Lifespan studies (127), the Healthy Brain Network (128), and the Adolescent Brain Cognitive Development Study (129).

3.5. Interpretation issues

There are a number of interpretation issues to take into consideration for the methods discussed. For single-view dimensional approaches such as factor analysis, methodological choices for addressing correlations between factors will impact the interpretation of the results. For example, while correlated traits models allow resultant factors to be highly correlated with one another resulting in a high degree of overlap between these factors (40), bifactor models produce orthogonal (uncorrelated) factors (40). The symptom factors in a bifactor model represent the specific variance not accounted for by general psychopathology (e.g., the fear factor represents the unique fear symptoms not shared across disorders). Thus, the presence or lack of orthogonality needs to be taken into consideration when interpreting these symptom factors. Likewise, interpretability is also an important issue in categorical approaches. The progress in developing and implementing different clustering algorithms has out-paced the research on evaluating the validity of the results (130). Notably, some clusters may be so smallas to no longer be meaningful and some individuals may not fit into any cluster (19). Finally, for both dimensional and categorical approaches, it is also unclear whether data reduction for high-dimensional data will help or hinder interpretability. Data reduction is widely used, but few studies address its impact on the results. Validation on an independent dataset or feature type will be useful for evaluating whether the results are biologically meaningful.

3.6. Reproducibility and generalizability

Finally, it is important to consider the reproducibility and generalizability of results in studies that attempt to parse heterogeneity. While there are varying opinions in the field as to the definition of reproducibility as opposed to replicability (131), here we take Plesser’s recommendation to adapt Goodman’s definitions of: 1) methods reproducibility - providing sufficient detail about procedures and data so that the same procedures could be exactly repeated; 2) results reproducibility - obtaining the same results from an independent study with procedures as closely matched to the original study as possible; and 3) inferential reproducibility - drawing the same interpretive conclusions (132). Results reproducibility and inferential reproducibility are both related to the generalizability of the results. Importantly, the stability of subgroups derived from clustering methods over time has been brought into question (19), and generalizability to new samples remains challenging for all methods discussed (133). Reproducibility and generalizability may be improved when out-of-sample validation methods are employed using best practices (119, 134). While the reproducibility crisis in the field is not exclusive to the methods covered in this review (135), the high-dimensional, multivariate nature of the data in conjunction with relatively small sample sizes inevitably reduces reproducibility. Thus, it will be important for researchers to provide transparent documentation of the decisions made at each step and to validate their results using appropriate methods.

4. Considerations for future research

4.1. The importance of considering brain maturation in youth

As should be apparent from this review, there are only a limited number of studies that have applied these approaches to samples of youth, suggesting a potential area growth for the field. Additionally, the studies reviewed have relied primarily on cross-sectional data spanning wide age ranges, which has clear limitations for studying development. Currently, most cross-sectional studies simply control for age effects, either by including age as a covariate in the model, or by removing age effects before clustering. Treating age as a confound – rather than the primary effect of interest – is not ideal for approaching primary developmental questions of interest. Longitudinal designs have long been considered the gold standard in developmental research (136) and the dynamic nature of longitudinal trajectories of brain development are well illustrated in the review by Becht and Mills in this special issue. Thus, it will be important for future work to apply the approaches reviewed to longitudinal samples to determine whether neurobiological patterns change throughout development. Future work that capitalizes on large, longitudinal studies such as the ABCD study (129) will be extremely useful in this regard.

4.2. Considering circularity in data-driven approaches

One of the primary goals of data-driven approaches that aim to reconceptualize psychopathology is to provide an alternative to clinically-defined DSM categorical diagnoses. However, many of the methods reviewed still rely to varying degrees on DSM defined symptoms, introducing potential circularity into the data-driven discovery of brain-behavior relationships. For example, the clinical symptoms used in a bifactor analysis or CCA often come from clinical interviews or self-report measures based on symptoms established by the DSM. Likewise, many clustering methods require defining a patient group to cluster on beforehand, which is typically defined using DSM criteria. If our input symptom measures or patient groups are heavily influenced by DSM defined diagnoses, then it is possible that our output may broadly align with these DSM categories. Such circularity may conflict with the goal of identifying underlying heterogeneity in order to redefine traditional diagnostic categories. Importantly, it is possible the that symptoms defined by the DSM through years of observation and scientific study could be the most relevant for data-driven exploration. Future work may be able to mitigate this circularity to some degree by applying these methods to representative samples that 1) do not exclude comorbidity, 2) include symptoms spanning the continuum from health to disorder, 3) assess the full range of psychiatric symptoms, and 4) include atypical or less common symptoms. Additionally, alternatives to using DSM-defined constructs should be explored, including studies framed around longitudinal functional outcomes (137) and treatment response across disorders (60).

4.3. The utility of modeling symptoms and neurobiology simultaneously

Studies that employ data-driven approaches to drive discovery of neural circuits associated with psychopathology represent a potential advance over the traditional case-control approach that dominated the field for years. The approaches discussed have propelled research beyond simple group differences to consider common circuit-level deficits which drive co-morbidity, and heterogenous biological mechanisms within clinical syndromes. Methods that take into account more than one feature set at the same time may be particularly valuable. While single-view approaches that consider a single set of features remain useful for specific goals (e.g., redefining the classification of psychopathology symptoms), modeling more complex relationships between symptoms and biological measures will necessitate the development and further refinement of methods that can consider multiple features sets simultaneously. Moving forward, it will be important to assess for convergence using multimodal data including clinical symptoms, neuroimaging features, neuropsychological measures, genetics, as well as cellular and molecular measures. Thus, future work would benefit from the application of advanced methods such as multi-kernel learning, generalized CCA, and multi-view biclustering to integrate these increasing numbers of data types.

5. Conclusion

Taken together, the approaches reviewed here can be useful for redefining our understanding of heterogeneity in psychopathology. A reconceptualization of psychiatric disorders using these data-driven methods may move the field forward beyond traditional symptom-defined categories. Critically, additional research is needed using these methods in younger samples with longitudinal designs. Parsing heterogeneity in youth is a critical first step to advancing interventions that target the pathophysiological mechanisms underlying psychopathology.