Harnessing networks and machine learning in neuropsychiatric care

Abstract

The development of next-generation therapies for neuropsychiatric illness will likely rely on a precise and accurate understanding of human brain dynamics. Toward this end, researchers have focused on collecting large quantities of neuroimaging data. For simplicity, we will refer to large cross-sectional neuroimaging studies as broad studies and to intensive longitudinal studies as deep studies. Recent progress in identifying illness subtypes and predicting treatment response in neuropsychiatry has been supported by these study designs, along with methods bridging machine learning and network science. Such methods combine analytic power, interpretability, and direct connection to underlying theory in cognitive neuroscience. Ultimately, we propose a general framework for the treatment of neuropsychiatric illness relying on the findings from broad and deep studies combined with basic cognitive and physiologic measurements.

Introduction

Neuropsychiatric illness has widespread and devastating effects on populations around the world, affecting approximately 20% of individuals in the U.S. alone [1]. Converging evidence from genetic, behavioral, and neuroimaging [2,3] studies has demonstrated overlapping pathological features in these disorders, suggesting that both common and unique pathophysiological mechanisms underlie clinical symptoms such as anxiety, depression, and psychosis. Accordingly, the classic notion of discrete psychiatric syndromes defined by clinical symptoms [4] is being challenged by more biologically and empirically driven models that link brain and behavior [5]. High rates of comorbidity between disorders hamper the identification of generalizable pathophysiological principles, similar to those that allow us to understand dysfunction of less complex internal organs. The dearth of such principles may partially explain the fact that a large cohort of patients do not respond to psychotherapy, psychopharmacologics [6], and brain stimulation protocols [7•]. Indeed, a marked consequence of the brain’s vast complexity is the existence of many distinct and overlapping pathways for cognitive function and dysfunction, constituting a major challenge in developing accurate diagnoses and predicting individual responses to treatment.

How, if ever, can we elucidate and intervene on these overlapping pathophysiological mechanisms that underlie neuropsychiatric illness? Recent efforts toward this aim have focused on the acquisition of human neuroimaging data sets with samples of unprecedented size [8•,9–11]. These so-called broad studies provide an excellent picture of between-individual or population-level variance, allowing the prediction of treatment response from high-dimensional neuroimaging and affective phenotypes based on methods from network neuroscience and machine learning [12••]. The widespread application of such methods has been facilitated by advances in computer processing power and repurposing of graphics processing units (GPUs) for machine learning. In complementary efforts, researchers have also collected data with repeated measures on a small cohort or single individual. These so-called deep studies have generated insights into the substantial within-individual variation in neuroimaging phenotypes that occurs on the scale of days, weeks, and months [13•,14,15]. Daily changes in neuroimaging phenotypes have also been linked to variability in behavioral and affective profiles [16], suggesting that temporal derivatives of neuroimaging phenotypes may contain unique, neuropsychiatrically relevant information. As such, both broad and deep studies have uniquely contributed to our understanding of healthy neurophysiology and neuropsychopathology.

While initial progress has been made through these unique forms of big data, neuropsychiatry still remains far from the goal of using generalizable principles to develop and deliver treatment. In this review, we begin by describing the results of broad and deep studies in more detail, along with methods well-suited for each study type. Next, we describe a framework to maximize the clinical translatability of broad and deep neuroimaging studies. Specifically, we posit that broad studies can inform models that identify who would benefit from intervention and how to intervene, while deep studies can inform models that suggest when to intervene. Network science and machine learning serve as the foundations for these models and will undoubtedly play a critical role in the coming generation of neuropsychiatric care.

Informing diagnosis and treatment through large population-level studies

Within the past decade, the neuroimaging community has seen the emergence of broad neuroimaging studies with historically large sample sizes (Figure 1). The Human Connectome Project [9], the UK Biobank [8•], the Philadelphia Neurodevelopmental Cohort [11], and Enhancing Neuroimaging Genetics through Meta-Analysis (ENIGMA) [10] have each generated valuable insights into the relations among brain structure, brain function, and behavior. Importantly, analyses of these data have increasingly relied on methods from network neuroscience [17•], an emerging field that provides elegant approaches for the quantitative description of complex multivariable phenotypes in brain anatomy and physiology. In large-scale brain networks, one can succinctly capture the collective role of several regions simultaneously through node-level metrics, such as the participation coefficient [18,19], controllability [20], hubness, nodal efficiency [21], and weighted degree [18] (Box 1). These metrics can be readily computed using freely available code [22,23]. Notably, these statistics are influenced by the topology of the entire network, change dynamically over time in functional networks [24], and are altered in neuropsychiatric disease [25–27]. To combine brain network models with clinical, behavioral, genetic, and cognitive data requires the use of multivariate statistical approaches that acknowledge the complexity of each of these data types by jointly accounting for their covariance structure [8•,28]. Sparse canonical correlation analysis (sCCA) [29] and partial least squares (PLS) [30] are two examples of such multivariate statistical methods that are well-suited to identify covariance patterns between brain networks and high dimensional behavioral, clinical, and genetic data (Figure 1).

Figure 1.

Broad and deep neuroimaging studies. The neuroimaging community has seen the rise of studies with increasingly large sample sizes (broad studies) and increasingly intensive sampling (deep studies). Broad studies (left) typically involve cross-sectional sampling of a specific population. Multivariate statistics and deep neural-network classifiers are two examples of methods that are well-suited to identify high-dimensional patterns between brain network, behavioral, clinical, and genetic phenotypes. Ultimately, such models might lead to the automated classification of neuropsychiatric illness based on neurobehavioral phenotypes, along with predictions of responses to various treatment options. Deep studies (right) typically involve intensive, repeated sampling of a small number of individuals longitudinally over days, months, or years. Multilayer network models can capture how the interaction between different components of brain activity, transcriptomes, metabolomes, or behavior changes over time. Here (middle right), each node represents a component and the time-varying edges represent their time-dependent interactions. Such models could prove particularly powerful for delineating how variability in individual networks over time affects treatment response.

Box 1. Modularity

Complex networks often contain non-trivial clustering in the form of modularity, in which groups of nodes exist that are more densely connected with each other than with nodes in other modules [18,31].

Network density

A fully dense network is one in which a connection exists between every possible pair of nodes. The density of a network is the number of existing connections divided by the number of possible connections.

Participation coefficient

Participation coefficient quantifies the extent to which a node sits on the boundary of multiple modules [18,19,31,32], poised to coordinate activity between functional systems of the brain [33].

Weighted degree

The weighted degree, or strength, of a node is the sum of its connection weights. Weighted degree can be further broken down into within-module and between-module degree, referring to the strength of a node’s connections to nodes in the same module or in other modules.

Hubs

Hubs are brain regions with unique roles in structural and functional networks due to their many, and often diverse, connections with other brain regions. Hubs are often disrupted in neuropsychiatric illness [34,26]. Hubs can be defined in several ways [18,19,32,35,36], often relying on a balance between participation coefficient and within-module degree.

Nodal efficiency

Nodal efficiency is a measure related to the average number of nodes that must be traversed to go from a given node to all other nodes [21,37]. This quantifies how an individual node contributes to the small world properties of brain networks [38].

Controllability

Unlike the above metrics, which quantify the static topological role of nodes in a network, network control theory [17•,23] uses a dynamical systems perspective to quantify the ability of each node to support transitions between states of activity. Two common metrics are average and modal controllability, which capture the ability of regional input to drive nearby or distant state transitions [23,39]. These principles have been explored in neuropsychiatric illness [25,40], over development [41], and across species [42].

Machine learning classifiers have demonstrated clear promise for neuropsychiatric diagnosis [43••], even with unimodal neuroimaging data. In a multisite study (n = 941), the ENIGMA schizophrenia working group utilized consensus-based classifiers to distinguish individuals with schizophrenia from healthy controls with 76% accuracy using structural MRI alone [44]. However, the clinical utility of such classifiers may not be realized until they are able to distinguish a particular disorder from a heterogeneous clinical population rather than healthy controls. Results from studies based on multivariate statistics suggest that the inclusion of clinical and behavioral data may help classifiers resolve this heterogeneity. In a large multisite study (n = 1188), CCA was used to define biotypes of major depressive disorder (MDD) based on resting state functional connectivity and clinical symptoms, allowing for diagnosis of depression with 85–90% accuracy in a replication set and prediction of positive response to transcranial magnetic stimulation (TMS) [12••]. The model was also able to distinguish MDD from schizophrenia more easily than from generalized anxiety disorder, reflecting the varying degrees of overlap in neurobehavioral phenotypes between different forms of mental illness. Importantly, individual patient data from independent samples can be fed into these models to generate priors for clinicians. In the near future, these models are likely to become increasingly powerful as open data sharing practices facilitate the growth of training data sets [45]. Such efforts will be critical for generating low dimensional representations of clinical symptoms and network measures of brain structure and function that are useful in the diagnosis and sub-diagnosis of disease, and in the selection of interventions and treatments (Figure 1).

Harnessing individual differences and within-subject dynamics

In contrast to broad studies, which leverage large sample sizes to make inferences about individuals in a defined population, deep studies are particularly suited for investigating the interdependencies between a diverse range of phenotypes that might vary meaningfully over time in single individuals (Figure 1). Perhaps the most impressive deep study is the MyConnectome project [13•], which is the first to describe the existence and nature of a complex interactome between resting state functional connectivity, transcriptomics, metabolomics, food intake, and behavior over the course of 532 days. It is interesting to consider the potential for such an interactome to inform the development of targeted neuromodulatory interventions that depend on the state of the brain at the time of stimulation [46]. Indeed, daily variation in brain network connectivity could confound the effects of stimulation, leading to mixed responses to such treatments for depression [7•]. To better understand these temporal variations, one can consider using multilayer network models, which can identify changes in network structure over time by taking into account the interactions between network components and the interactions within each network component with time [47] (Figure 1). One can also use linear autoregressive models, Hidden Markov Models, or Long Short-Term Memory recurrent neural networks [48] to predict how a complex, interacting system evolves over time. While the level of depth reported in the MyConnectome project is currently impractical for patient care, it illustrates the complex origins of day-to-day individual variation and — alongside other intensive sampling studies — offers useful benchmarks to inform future data collection [14,49,50••,51].

Typical approaches for ‘parcellation,’ – obtaining representative signals within anatomically [52] or functionally [33] similar regions or ‘parcels’ – tend to rely on registering brain images to a common template space. However, performing targeted manipulations of distributed cognitive systems that exhibit dysfunction in neuropsychiatric illness demands exceptional precision in mapping brain network architecture and function. Thus, the growing focus on subject-specific parcellation to define these parcels independently for each participant or patient is a critical complement to the intensive sampling of deep studies (Figure 2) [14,15,53,54]. Constructing subject-specific parcellations builds on historical work in tumor resection, where neurosurgeons and anesthesiologists perform patient-specific functional mapping of language and motor circuits with fMRI, pharmacology, and electrical stimulation [55,56]. When seeking to map all circuits across the entire brain, one would focus on mapping individual differences in functional topography that might hold diagnostic and prognostic value, with methods that do not depend on warping subject-level volumes to an average brain [50••,57]. Recently, such individualized parcellation techniques have been combined with resting state fMRI to identify novel subnetworks within the default mode network (DMN) [58], a system that has been broadly implicated in virtually every neuropsychiatric illness [59–62]. These observations motivate further studies of individual differences in the distribution of cortical real estate between particular functional networks [63] and their finer subdivisions in the context of neuro-psychiatric illness (Figure 2). In these efforts, deep neuroimaging studies will be particularly important, by providing sufficient data to use subject-specific parcellations. This approach will account for — rather than average over — individual network topographies (Figure 2). Resolving individual differences in spatial topography will facilitate an accurate study of the neural basis of temporal fluctuations in individual symptoms. The richly sampled temporal dimension of deep studies adds a layer of complexity untouched by most broad studies and the individual-oriented methodology improves the accuracy of patient-specific predictions. Ultimately, meta-analysis of deep studies might inform a generalizable approach, if not pathophysiological principles, for making individual predictions of the optimal treatment as a function of time and a more easily measurable subset of variables.

Figure 2.

Subject-specific parcellation uncovers individualized topography of functional networks. (a) Sample depiction of the default mode network in a group-level parcellation, which warps subject volumes to standard space, potentially averaging over important differences in functional network topography. Yellow overlay indicates activity, while colored lines indicate parcel boundaries. (b) Sample depiction of subject-specific activation map (yellow overlay) with group-level parcellation borders (colored lines) overlayed, illustrating the possible variation in subject-specific functional topography.

Using network models to link intensive behavioral assessment with neuroimaging findings

A key counterpart to accurately interpreting changes in functional brain dynamics over time in a patient cohort is the ability to concurrently measure changes in behavior, emotions, and mood — core symptoms of neuropsychiatric illness, which are typically assessed retrospectively in the clinical setting. Experience-sampling (ES) encompasses the measurement of these factors, in addition to physiology, in real time through the use of personal data recording tools [64]. Subject-specific symptom networks can be constructed by computing cross-correlations between measures of different emotions over time, quantifying the cofluctuation of psychiatric symptoms [65•,66,67] (Figure 3). Additionally, directed networks can be constructed using pairwise regression between time-lagged measures, capturing the temporal precedence of symptom fluctuations [68]. Higher order features of these symptom graphs [65•], such as network density (Box 1), are greater in individuals with MDD than in healthy controls [69], suggesting that the temporally dynamic interplay between symptoms may be altered by disease processes. ES also lends itself well to the study of substance use disorders, in which daily emotional variability can trigger relapse [70].

Figure 3.

Measuring mood dynamics with experience sampling. (a) Illustration of symptom measurement by retrospective report, demonstrating how scales of mood symptoms that ask for retrospective reporting average over rich mood dynamics and are subject to recall bias [72–74]. (b) Example time series of mood measurements from the Profile of Mood States [76]. Spikes in the time series indicate rapid changes in mood induced by brief events. The numbering on the x-axis indicates windowing for network construction. (c) Illustration of multilayer emotional network construction from subsequent windows of time series shown in panel (b). Nodes represent mood features, with the letter label corresponding to features in (b), solid gray edges represent the correlation between mood features within a particular time window, and dashed black edges link mood features together across time. Constructing such a network facilitates the application of numerous methods for analyzing the temporal dynamics of multivariate relationships [47].

Major limitations of ES are the burden of repeated assessment on participants and the potential for the process of ES itself to influence symptoms (i.e. reactivity [71]). Nevertheless, commonly used forms [72–74] for evaluating mental health utilize retrospective reporting, assuming stationarity in these dynamic phenotypes [75] (Figure 3). Schizophrenia, for example, is characterized by a lack of insight and poor working memory, and therefore real-time assessment may be more likely to accurately capture cognitive and emotional state than single-shot clinical evaluations or self-report measures. Thus, ES is a highly promising approach for identifying neural correlates of symptom dynamics in complex, overlapping neuropsychiatric pathologies.

The use of ES has begun to enter into neuroimaging studies, though not with the same force as the machine learning techniques described above. In schizophrenia patients, corticostriatal task activation and reduced motor activity were found to predict negative symptoms [77]. Similarly, physiological signs of autonomic dysfunction acquired through wearable technology were associated with positive symptom severity [78]. In a study of patients with anorexia nervosa, reward circuit activity was related to longitudinal body-mass index measurements and body-related rumination [79]. Notably, this particular study used group-level parcellation techniques, indicative of a common disconnect between the use of cutting edge methods in social science and those in neuroscience.

Despite these intriguing findings, no parallels have yet been drawn between basic ES measures, network models of psychiatric symptoms, and structural or functional brain networks. Functional brain network dynamics have been extensively characterized [24,80], and there are likely rich relationships with behavioral and symptom dynamics, as suggested by the MyConnectome project. One could gain traction on these relationships using multilayer network construction with subsequent community detection [47] to draw parallels between dynamic functional networks and symptom networks. Brain regions with high inter-scan variability in functional connectivity and high within-scan community change, that is, flexibility [16], may confer similar variability onto behavior phenotypes.

The use of advanced machine learning techniques for time series analysis, such as recurrent neural networks and Hidden Markov Models, as well as unsupervised multivariate statistical methods, are promising underexplored avenues for finding covariance between complex neural and behavioral phenotypes in neuropsychatric illness. Furthermore, ES could explain temporal variance in cortical excitability [81,82], an important factor in TMS response [83••], and allow for its targeted control. While targeted neuromodulatory treatment paradigms are currently being refined, with the aid of findings from broad studies, ES provides us with useful methods that will help to identify the optimal time in a disease course to deliver these treatments.

Conclusion

Across many academic disciplines, the use of machine learning techniques, often informed by network theory, has skyrocketed in the last decade, concordant with the collection of data with larger (broad) and more intensive (deep) samples. Both broad and deep studies provide the neuroscience community with unique opportunities to advance the diagnosis and treatment of neuropsychiatric illness, with the aid of network science and machine learning. Broad studies allow for network analysis followed by dimensionality reduction and classification for identifying meaningful symptom-neuropathology correspondence and predicting treatment responses. Deep studies demonstrate the importance of individual variability and provide a framework for understanding and manipulating complex, individual phenomes. Experience sampling provides the tools for acquiring intensive repeated physiologic and behavioral measures, the network models of which may have critical unexplored neural correlates. Ultimately, a model using priors derived from broad studies, patient-specific neuroimaging data, and symptom networks might predict the optimal timing and type of treatment for individual patients in real time based on a subset of measurements captured through a personal device. The merger of these techniques has the potential to usher in a next-generation approach to psychiatric care and contribute to our fundamental understanding of the complex relationship between mind, body, and behavior.