The promise of precision functional mapping for neuroimaging in psychiatry

Abstract

Precision functional mapping (PFM) is a neuroimaging approach to reliably estimate metrics of brain function from individual people via the collection of large amounts of fMRI data (hours per person). This method has revealed much about the inter-individual variation of functional brain networks. While standard group-level studies, in which we average brain measures across groups of people, are important in understanding the generalizable neural underpinnings of neuropsychiatric disorders, many disorders are heterogeneous in nature. This heterogeneity often complicates clinical care, leading to patient uncertainty when considering prognosis or treatment options. We posit that PFM methods may help streamline clinical care in the future, fast-tracking the choice of personalized treatment that is most compatible with the individual. In this review, we provide a history of PFM studies, foundational results highlighting the benefits of PFM methods in the pursuit of an advanced understanding of individual differences in functional network organization, and possible avenues where PFM can contribute to clinical translation of neuroimaging research results in the way of personalized treatment in psychiatry.

Introduction

Over the past decade, advancements in neuroimaging methods have contributed to a richer understanding of inter-individual differences in brain function. One method in particular, Precision Functional Mapping (PFM), has propelled our understanding of functional brain network organization and has the potential to clarify how inter-individual variation of these networks may be associated with behavior or clinical diagnoses. Given the heterogeneous nature of common neuropsychiatric disorders such as depression [1–3], attention-deficit hyperactivity disorder (ADHD) [4, 5], and autism spectrum disorder (ASD) [6, 7], PFM holds clinical promise as individual variations in the organization and integration of functional brain networks may help guide personalized medicine and provide a higher standard of clinical care.

Functional MRI (fMRI) research intending to establish clinical translation of results has traditionally focused on large group studies that aim for approximately 10–30 min of functional imaging data per participant, per task. These studies strive for a balance of sufficient data collection for accurate group-level analyses, while staying within the limitations of what can be reasonably collected during a single, ~1-h scan visit. When applied to clinical populations, these studies aim to identify group differences between large cohorts of individuals or clinical sub-populations [8, 9]. While group-level studies (both cross-sectional and longitudinal) have improved our understanding of the generalized makeup and stability [10, 11] of canonical functional brain networks and provides valuable insights for a broader understanding of the neural correlates of behavior, inter-individual variability is necessarily muted when averaging across a group. This reduction in individual specificity in the data can make translation and adoption of those results into individualized clinical practice difficult. Overcoming this obstacle is important in the pursuit of successful individualized clinical care, as generally, neuropsychiatric disorders have not been found to be associated with vast alterations of functional network organization. Rather, group differences in brain function relevant to behavior or clinical sub-populations are often associated with smaller disruptions in how functional networks organize and integrate.

There are many ways in which PFM may benefit psychiatry, both in furthering our understanding of mechanism and in guiding individualized clinical care. Below in “Bringing together PFM and psychiatry”, we discuss three potential utilities of PFM, including linking brain and clinical features in individuals, individualized brain targeting for interventions, and individualized tracking of change. However, we first provide the reader with foundational background, including operationalization of PFM as a method, the origins of PFM, and current insights regarding brain function afforded by PFM methods. Then, we discuss the use and potential of PFM in psychiatry, as well as considerations towards a path forward to incorporating PFM into clinical practice.

An operational definition of PFM

PFM is the precise characterization of individual brain function, currently made possible by the collection of hours of non-invasive fMRI data from an individual, typically collected over multiple visits. PFM relies on large amounts of resting state or task fMRI data in order to extract highly reliable, individualized estimates of brain function and functional connectivity (FC), which allow for the creation of high-resolution, high fidelity individual-specific maps of functional brain networks or activation. Although ever-evolving computational methods can improve the individual-specific accuracy of more common, shorter fMRI collection times [12–15], there is often a trade-off due to signal manipulation that can result in difficulties accurately interpreting and translating those results to a clinical setting. PFM is an emerging method for neuroimaging research that is currently gaining more attention and adoption. As such, a set of best practices or criteria to deem a study as “PFM” is not established. In fact, it is difficult to prescribe a one-size-fits-all requirement for PFM, as the best practices for a given study depend on the questions and metrics of interest. Our contention is that high within-participant reliability of the metric of interest (e.g., whole-brain FC, delineation of functional network topology) is the target for PFM analyses. This measure can be quantified with split-half or incremental reliability estimates, with the goal of reaching r > 0.8–0.9, thus ensuring a high confidence of within-participant measurement accuracy. It has been shown that longer scan times improve reliability of FC [16], yet numerous factors can influence measures of reliability, such as the spatial scale of investigation (e.g., networks vs. parcels) and MRI sequence parameters (e.g., voxel size, single vs. multi-echo). Therefore, the amount of data required will vary study to study.

The current review defines PFM for neuroimaging studies as the analysis of fMRI data that achieves individual specificity by ensuring that estimates of brain function (e.g., correlation of signal between two regions or between all regions of the brain) are reproducible with a separate dataset from the same person. Although PFM itself is a newer method, it is heavily rooted in the repeated collection of large amounts of targeted data, pioneered by functional localizer studies [17–20], which have provided fundamental insight to the specific function of brain regions that cannot be delineated solely by anatomical markers. While similar in their methods, PFM is typically based on a more macro view of brain function and aims to accurately measure whole-brain functional network organization, connectivity, and activation, to provide insights to brain function as a larger, interconnected network of brain activity.

Foundational PFM collections

Notable collections of PFM data started to emerge around 2015. These studies were aimed at quantifying the inter-individual variability of FC – pairwise correlations of timeseries derived from the blood oxygenation level-dependent (BOLD) signal – that is obscured in group-average analyses, as well as establishing suggested, baseline methods for the empirical testing of PFM data. The “MyConnectome Project” [21] was a single-participant collection (45 yr, male) of both neuroimaging and genetic data, whose collection spanned over 18 months. Neuroimaging data consisted of resting-state fMRI, cognitive task fMRI, and fMRI localizer tasks. This collection not only focused on the establishment of a precision, longitudinal dataset to test the dynamics of brain function across multiple tasks and over a vast timescale, but also included biological samples to measure the relationship of bodily functions, such as metabolism, to measures of brain function.

Following the initial publication and release of the MyConnectome dataset, a follow-up paper [22] reported the detailed description of that participant’s brain organization, highlighting both functional and anatomical variations from established group-average estimates and including a second adult PFM participant for between-participant comparisons. One notable result showed the MyConnectome participant exhibited general functional network topography that was broadly similar to group-level estimates of functional network organization. However, distinct individual variation of FC was identified by comparing connectivity maps created from two seed regions of interest (ROIs) in the lateral frontal cortex. In the group average, both ROIs exhibited similar connectivity and were assigned to the fronto-parietal network (FPN). However, connectivity maps in the highly-sampled PFM individual revealed one seed with varied, yet similar connectivity to the FPN, and the second seed with connectivity more aligned with the cingulo-opercular network (recently termed the action-mode network [23] (preprint) and hence will be referred to as CON/AMN hereafter). This difference in FC was a direct example of instances where group-defined cortical network parcels can obscure individual topography, thereby ignoring or muting individual variation. The MyConnectome Project and the analyses that followed were a “first of its kind” study, establishing a rich, phenome-wide, longitudinal, publicly available sample of single-participant data and setting the stage for modern PFM data collection and analysis.

Inspired by the MyConnectome study, the Midnight Scan Club [24] (MSC) study sought to create an extended PFM dataset including 10 adult participants (24–34 yr, 5 M, 5 F), allowing for greater inter-individual comparisons. The MSC dataset included five hours of resting-state fMRI, six hours of motor and perceptual fMRI task data, and four neuropsychological behavioral assessments, per participant. With the inclusion of more PFM participants and an expanded range of clinical measures, the MSC dataset was able to better address inter-individual variability of functional brain networks (Fig. 1). Further, this study quantified within-participant reliability of brain metrics commonly used to measure network organization and integration. One significant impact of this publication was the identification of “networks variants” within every individual and evidence of shared network variant “subgroups” that may exist in the population and could be related to demographics, behavior, or cognitive abilities (described more below). The MSC data continue to serve as a key, publicly available adult dataset that is often used for PFM analyses or as a PFM comparison to group data.

Fig. 1. Individual variation in cortical functional network organization revealed by PFM.

a Common deviations from the group average (blue and purple arrows) in 6 individuals. b Seed ROIs (white spheres A&B) delineate differences in functional connectivity and network assignments in a group average (MSCavg) and in PFM data from one individual (MSC06). (Figure adapted from Gordon et al. [24]).

During this time, researchers also sought to establish a better understanding of variables that contribute to daily intra-individual variability of MRI data. In 2017, the Day2day study [25] released a publicly available MRI dataset (including various MRI modalities) to allow for the study of intra-individual differences due to day-to-day effects on scanner measurements. While the MyConnectome dataset provided longitudinal scanning of one individual that could estimate these differences, the Day2day dataset expanded its collection to eight adult participants (24–32 yrs, 2 male) who underwent near-daily scanning. Further, the Day2day collection increased the specificity of its acquisitions, including high-resolution anatomical imaging of the hippocampus, along with the commonly collected anatomical, functional, and diffusion imaging data. These three collections, the MyConnectome, MSC, and Day2day datasets, have made significant publicly available contributions for future research questions, while also setting the stage for PFM analyses to come.

Current advancements using PFM

Individual differences

The aforementioned foundational PFM studies provided the groundwork for understanding the stability and reliability of fMRI metrics for an individual person. An important step towards clinical applicability of this approach is to quantify individual differences in brain function that ultimately relate to individual differences in clinical features.

Given the identification of individual variability in functional network organization, a vital question for understanding the relevance of this variability to psychiatry is to parse out the sources of such variation. For example, is the observed individual variability driven by differences in what people are thinking during a resting state fMRI scan? Is it driven by stable characteristics of the individual that do not change over time? If we are measuring stable person-specific brain network features, and not different states or thoughts, then these features would be more useful as biomarkers relating to individual differences in psychopathology. Gratton and colleagues [26] leveraged the extended amount of fMRI data from single individuals collected over many sessions along with the acquisition of multiple tasks and resting state in the MSC dataset to examine the contribution of different sources of variation. Their findings revealed that functional network organization was dominated by commonalities across the group as well as stable individual-specific features. Contributions of day-to-day (session) and task state (e.g., rest vs. motor task vs. memory task) variation were measurable, but the effects were much smaller than those of stable individual characteristics. Thus, individual differences in functional networks reflect stable characteristics of a person, such as genetics and environmental factors, more than short timescale changes over sessions or different task states. These results are promising for the use of PFM in psychiatry, as individual-level measures of functional networks that represent stable characteristics of a person may capture individual differences in each person’s behavior or even collection of symptoms.

Further work has investigated the nature of this individual variability. Up to this point, PFM studies showed that while individuals’ functional networks follow a consistent broad organization, there are measurable, specific features that are individually unique. For example, some individuals exhibit part of the salience network in ventromedial prefrontal cortex, which is typically represented as the default-mode network in other individuals and in the group average network map [24]. Thus, one goal was to study the stability, cortical location, and possible dynamic nature of these features. Seitzman and colleagues [27] sought to characterize the stability and trait-like properties of regions of the brain in which FC in an individual differed from that of the group average, termed “network variants.” Across three independent datasets, they found that network variants did not coincide with anatomical differences. Rather, variants were linked to inter-individual differences in task-evoked signals and were a product of their functional network association, not cortical location. In addition, like whole-brain functional network organization, the spatial locations of variants were highly stable across time (session) within an individual, supporting the idea that variants represent characteristic features of an individual. Further, this work provided supporting data for the existence of variant types that cluster into subgroups, are shared across individuals, and may be associated with behavioral differences. Two main network variant clusters were found: (1) variants more strongly associated with processing and control systems, exhibiting FC organization most aligned with sensorimotor networks, the CON/AMN, and the dorsal-attention network (DAN), and (2) variants primarily associated with the default mode network (DMN). These two subgroups were anticorrelated, suggesting two types of network variants that may emerge for distinct roles in brain function.

Although inter-individual differences in functional network organization do not appear to be due to or linked to specific anatomical differences, network variants do tend to appear in common brain regions across people. Work aimed at producing probabilistic maps of functional networks using PFM data explored the topographical stability of cortical networks and, in turn, highlighted brain regions most prone to exhibiting individual-specific variations [28]. While core regions of all networks exhibit high consensus across individuals, resulting in functional networks that are vastly stable across development, populations, and differences in scanner model or acquisition settings, regional boundaries of networks showed the lowest consensus and thus greater inter-individual variability. Further, regions of low network consensus, such as the temporoparietal junction (TPJ), align with the previously reported variant locations [27], helping to establish more generalizable locations of inter-individual network variability. The spatial locations of network variants have also been shown to be stable between rest and task states within a given individual [29]. Interestingly, though the locations of variants are broadly similar across people, the intra-individual spatial overlap of variants between rest and task states was higher than the overlap between-individuals. Such findings provide additional evidence that individual variability in functional network features reflects stable characteristics of the individual.

Machine learning classification results further bolster the findings of reliable inter-individual variability in brain function. Ridge regression classifiers trained on densely sampled single-person fMRI data from resting state, motor, semantic, perceptual coherence, and memory tasks showed highest classification accuracy of task state within an individual. Interestingly, these individualized classifiers were also generalizable to new individuals, yet cross-individual classification required approximately four times as much data to achieve comparable accuracy [30]. Similar results were achieved using Support Vector Machines (SVM), Multi-Layer Perceptron (MLP), and Graph Convolutional Neural Networks (GCN) decoding methods across seven tasks (emotion processing, gambling, language processing, motor, relational processing, social cognition, and working memory) [31]. This study found intra-individual accuracy higher than inter-individual accuracy and that orders of magnitude less PFM training data was required to achieve accuracy comparable to group decoding models.

Recent work has also provided evidence that PFM is necessary to accurately investigate sub-properties of canonical resting state networks, which may be obscured in group averages due to their interconnected nature or may not exist at all in certain individuals. Group average studies typically characterize the DMN as either a single network or consisting of three sub-components [32, 33]: (1) a core system involved in self-referential processing, (2) a dorsal medial system attributed to Theory of Mind and semantic meaning, and (3) a medial temporal system associated with contextual association or episodic remembering. However, PFM work has shown that certain sub-properties of the DMN identified in group data may be an artifact of group average mapping of functional networks [34]. Specifically, the “core system” of the DMN may be in fact regions of inter-individual variation of the two interdigitated DMN sub-system boundaries. Similarly, the organization of the parietal memory network (PMN) often includes regions in the posterior inferior parietal lobule and dorsal angular gyrus, yet when examined in PFM data, not all participants showed connectivity to the PMN in these regions [35]. It is hypothesized that individuals who lack this PMN connectivity in certain regions may exhibit lowered sensitivity to certain stimuli, which would be obscured using group-derived parcellations of the PMN. The behavioral associations of sub-components of the DMN were further investigated using two different domains of cognition: episodic projection and Theory of Mind. To determine if these two domains recruit different brain regions, despite both being attributed to the DMN in group level studies [36], six participants were scanned four times each, completing episodic projection tasks (Past Self, Future Self) and Theory of Mind tasks estimating others’ mental states, beliefs, or emotional pain or suffering. PFM revealed two distinct networks, one for each of the tested cognitive domains. Both networks were primarily linked to regions associated with the DMN (parahippocampal cortex, TPJ), yet the preferential distinction for each type of task is typically obscured in group average data due to these sub-properties of the DMN being interwoven together.

Functional brain networks can also be examined dynamically, that is changing over time, which may provide another measure of clinically relevant inter-individual differences [37]. Peng and colleagues used a method called Individualized Network-based Single-frame Coactivation Pattern Estimation (INSCAPE) to generate templates of brain states, aimed to delineate how functional networks reorganize based on moment-to-moment changes in fMRI data. This work found that dynamic brain states were highly reliable at the individual subject level and also exhibited sensitivity that could reliably capture inter-individual differences in transient network dynamics within both resting state fMRI data and language task fMRI data. Similarly, an emerging method of functional network analysis that focuses on network edges to reveal network dynamics at fine timescales - edge time series (ETS) [38, 39] - can also estimate network fluctuations at the individual level. With this approach, functional network states refer to the presence of repeated patterns or events in co-fluctuations of timeseries data that form clusters or communities of similar connectivity. Using a densely sampled PFM dataset [40] (N = 1) with fMRI scans and blood draws occurring daily for 30 days over each of two endocrine states (one during a natural menstrual cycle, and one while taking oral contraceptives for selective progesterone suppression), ETS revealed two distinct and repeating functional network states (communities) that coincided with differences in hormone concentration. Cluster analysis revealed repeating connectivity states within the individual’s functional network fluctuations. One state in particular showed opposed activation of the default mode and control regions with attention and sensorimotor regions, and was significantly correlated with concentrations of luteinizing and follicle-stimulating hormone concentrations. These results suggest that tracking clusters of fluctuating connectivity may provide focused and clinically relevant associations with hormonal changes in individualized clinical care.

Cortical organization and neuroplasticity

PFM has also afforded insight about functional network organization across the cortex that standard group-level approaches are not sensitive enough to detect. The importance of elucidating network organization at a more fine-grained level is important for the promise of these methods to be useful in psychiatry and account for the heterogeneity across individual patients. Our previous reliance on a group-level, population-average understanding of functional network organization washes out details that may be key for understanding this clinical heterogeneity.

Network analyses are sensitive to methodological decisions (e.g., thresholding) [41]. Thus, methodological decisions in functional network assignment often struggle with how to appropriately handle regions with significant connectivity to multiple functional networks [42]. Cortical regions that are associated with multiple networks have often been characterized as “hubs” or “network integration zones.” PFM provides a powerful avenue to interrogate these cortical regions more deeply. Previously, network hubs (regions thought to integrate discrete cortical networks) were considered as a single category of brain regions with similar properties [43–46]. However, PFM has revealed three distinct hub categories: (1) “control-default” hubs with strong connectivity to the FPN, DMN, and contextual association network, (2) “cross-control” hubs with strong connectivity to the CON/AMN, dorsal attention network (DAN), and FPN, and (3) “control-processing” hubs with strong connectivity to sensory and motor processing systems, along with CON/AMN and DAN [47]. Analyses of task activations associated with these hub categories revealed preferential association to specific tasks, with hub category being a significant explanatory factor for all tasks, and hub category explaining more variance of activation than the network identity of those cortical regions alone. These results suggest that cortical hub regions are associated with task performance, not only by the functional network those regions are assigned to, but in how they specifically integrate networks across the cortex. Additionally, the observed impact and isolation of functional networks following the removal of cortical hub nodes (simulating damage to that cortical region) may assist in formulating hypotheses about specific behaviors that can be affected by perturbation of these hub regions in clinical scenarios.

The observed variation of network boundaries in individual-specific functional network maps may also be influenced by local minima brought on by standard group-average data. One candidate case example is the substructure of the DMN. A series of PFM studies, including six participants that underwent 31 or 24 scan sessions on a 3 Tesla scanner, and three participants who completed shorter 7 Tesla scans [48, 49], investigated the fine-grained organization of the DMN. This work, as with work conducted with task data [36], demonstrated a division of the DMN, with each of the two distinct networks showing unique but interdigitated connectivity in regions attributed to the single canonical DMN. One network showed connectivity primarily within ventromedial prefrontal cortex (PFC), retrosplenial/ventral posteromedial cortex, and parahippocampal cortex, while the other network showed connectivity primarily in the TPJ, lateral temporal cortex, and posterior cingulate cortex. This substructure of what is typically described as a single network (DMN) may allow for more informed investigation of the role of the DMN and its components in psychiatric disorders and symptoms.

As the body of PFM research grows, it is becoming more clear that the previously group-defined architecture of functional brain networks has obscured nuance in the behavioral roles of these networks. One recent and pronounced demonstration of this obscured nuance is the discovery of the somato-cognitive action network (SCAN) [50]. At the group average level, the somatomotor area of the cortex is typically represented as a single larger network or subdivided into two or three effector specific functional networks [42, 51, 52]. However, Gordon and colleagues leveraged PFM to identify three inter-effector regions within the somatomotor networks that lack activation specificity to movement. Rather, the inter-effector regions were highly connected to one another and to the CON/AMN - a functional network involved in cognitive and action control [23, 53, 54] - and show co-activation during the planning of actions (Fig. 2a). The researchers posit that the interdigitated nature of the SCAN and somatomotor networks reveals parallel systems that integrate effector-specific roles of motor control with the higher-level cognitive task of action planning. As the field’s methods for collecting and analyzing precision data improve, a refined understanding of functional brain networks shows great promise in advancing our understanding of how these networks are involved in behavior.

Fig. 2. Novel discoveries of cortical organization and plasticity made possible by PFM.

a Seed ROIs (white spheres) show functional connectivity of inter-effector (somato-cognitive action network), foot, hand, and mouth regions. b Seed maps show functional connectivity of left primary somatomotor cortex before (Pre), during (Cast), and after casting (Post) in one participant, showing cast-induced changes. c Daily time course of functional connectivity between left and right upper extremity for each participant. A time-varying exponential decay model was used to calculate Δr values. (Nico: p = 0.002, Ashley: p < 0.001, Omar: p < 0.001) (a adapted from Gordon et al. [50]; b, c adapted from Newbold et al. [58]).

Given the ability to delineate functional brain networks with finer-grained specificity using PFM, we can also leverage this approach to investigate changes to network organization in the face of perturbation. The brain’s ability to reorganize neural networks during typical development and in response to injury (neuroplasticity) is vital for successful adaptation to changing life events. Recent work has demonstrated the power of PFM to non-invasively study how functional brain networks can adapt and change in response to both injury and experimental disuse of selected motor function. In a case study of one adolescent male who experienced significant brain trauma at three weeks of age due to bilateral perinatal strokes resulting in bilateral cystic cortical lesions [55], researchers were able to gain a better understanding of the adaptive nature of functional brain networks. Surprisingly given the extent of cortical damage - approximately 20% loss of supratentorial brain volume - PFM of this patient revealed a highly adapted and intact network structure, which behaviorally only presented with temporary gait asymmetry during development and a persistent left-hand preference. The patient’s cognitive function was similar to age-matched standardized scores, and motor function tests detected only slight deficits to upper limb strength and dexterity. Overall functional networks appeared intact, with networks remapped from the frontal and parietal regions that experienced significant trauma. Thus, to make up for the tissue loss, there was a posterior shift in somatomotor network mapping with task-based activation falling within those individualized network boundaries. This case study demonstrates the power of PFM for studying brain network organization in an individual patient with unique neurophysiology, and revealed the incredible capacity of the brain to reorganize.

Aside from brain trauma, reorganization of neural networks can also occur due to extended disuse [56, 57], which can be brought about by bodily injury and the following recovery. To test and quantify the plasticity of functional brain networks, three individuals voluntarily casted their dominant arm from bicep to fingertips for two weeks, while undergoing daily, 30 min resting-state fMRI scans and motor assessment (e.g., accelerometry of the casted arm, grip strength [58, 59]). Most notably, the regions of somatomotor cortex associated with the disused limb became functionally uncoupled from the contralateral somatomotor cortex (Fig. 2b), with significant decreases occurring within just 48 h of casting, displaying a rapid adaptation of functional network organization. Further, homotopic functional recoupling of the somatomotor network regions recovered quickly upon cast removal. Changes in FC were also tested using the graph metric modularity, which quantifies the degree to which a network is divided into multiple, smaller networks or clusters. During casting, the somatomotor cortex associated with the casted arm showed significantly higher modularity in all three participants, signifying a functional separation of regions associated with the unused extremity, that rebounded to pre-cast levels or higher upon removal of the cast (Fig. 2c). Additionally, “disuse pulses” (pulses of spontaneous activity) were observed in the sub-circuit of the disused somatomotor region, which the researchers posit may have protected the circuit during the period of disuse and contributed to the rapid recovery of network architecture once the cast was removed.

With PFM, we have been able to see the juxtaposed nature of individual-level functional network organization, such that we can measure high stability within an individual, but also neuroplasticity in networks in the face of perturbation. The ability to capture both stability and change is highly relevant to psychiatry, as we strive to identify stable individual differences that contribute to the heterogeneity in many neuropsychiatric disorders, but also the plasticity that can occur with major events or changes in a patient’s life, such as trauma or treatment.

Other brain structures

While a substantial body of neuroimaging work has characterized the cortical organization of functional networks over the past decade, the functional organization of other structures, such as the cerebellum, basal ganglia, and thalamus, has received less attention. With the advancement of fMRI protocols and analytical methods, researchers have begun to fill this gap in the literature [60–62]. However, much like with our understanding of cortical functional network organization, PFM datasets are improving our ability to measure intricate organization of non-cortical brain structures. Subcortical structures are fundamentally delineated by their anatomical boundaries based on cytoarchitecture. Yet, the putative functions of these structures may be better represented by divisions that take into account functional activity and connectivity to cortical brain networks [63]. Importantly, subcortical structures, the cerebellum, and the circuitry connecting these structures to cortical regions have been implicated in a number of neuropsychiatric disorders, including attention-deficit/hyperactivity disorder, obsessive compulsive disorder, Tourette syndrome, and depression [64–68].

PFM has improved our understanding of the representation of functional brain networks in subcortical structures that do not clearly align with anatomical boundaries, such as functional subdivisions within a structure and integration zones where multiple cortical networks converge [69]. Investigation of the basal ganglia and thalamus using PFM identified subcortical regions that showed strong connectivity to either a single network or multiple networks, the latter termed “integration zones” [70]. These integration zones clustered into three distinct categories: (1) visual attention integration zones, located in the pulvinar integrating the DAN and visual network, (2) cognitive integration zones, located in the caudate nucleus integrating the DMN, salience network (SAL), FPN, and ventral attention network (VAN), and (3) motor integration zones, located in the ventral intermediate thalamus integrating the CON/AMN and somatomotor networks (Fig. 3a). Identification of these integration zones was possible with PFM data, providing confidence that the convergence of cortical networks was not an artifact of group averaging. Further, PFM revealed that functional subdivisions within the subcortex were either common across all individuals or individually-specific. It was posited that individual-specific functional organization within subcortical structures may improve clinical outcomes of deep brain stimulation (DBS) treatment. This idea is theoretically supported by common DBS targets for essential tremor aligning with group-common (low-variability) motor integration zones and exhibiting higher efficacy than DBS targets for Parkinson disease and dystonia, which align with individually-specific (high-variability) regions of the globus pallidus and exhibit widely variable efficacy.

Fig. 3. Functional network organization of the subcortex and cerebellum in PFM data.

a Functional network assignment of subcortical structures in one representative male (MSC02), one representative female (MSC04), and the group average. Voxels with preferential functional connectivity to a single network are indicated with solid colors and voxels functionally connected to multiple networks are shown with cross-hatching. Three clusters of integration zones are detailed on the right. b Functional network representation of two individuals (MSC01 and MSC09) in the cerebellum. c Flatmaps of cerebellar network parcellations in two individuals (Subject1 and Subject2). (a adapted from Greene et al. [70]; b adapted from Marek et al. [83]; c adapted from Xue et al. [84]).

Functional subdivisions with individual specificity have also been described in the hippocampus and amygdala using PFM data. The hippocampus showed preferential FC with two functional networks, the DMN connecting with the head and body of the hippocampus and the PMN connecting with the hippocampal tail [71]. Along with the FC results, deactivation during spatial coherence and noun-verb discrimination tasks [24, 54] was observed in the hippocampus head and body (DMN), but not the hippocampal tail (PMN), supporting the observed network split based on expected deactivation of the DMN during attention-demanding tasks [72]. In the amygdala, PFM defined functional subdivisions with a data-driven approach [73]. Each individual-specific subdivision of the amygdala was generated based on connectivity to the cortex, and compared to empirically defined subdivisions and subdivisions from a predefined template. Results demonstrated that the centromedial subdivision (central and medial nuclei) showed strong and preferential connectivity to the DMN, the superficial subdivision (anterior amygdaloid area) showed preferential connectivity to the DAN and FPN, and the laterobasal subdivision (lateral, basolateral, basomedial, and paralaminar nuclei) did not show preferential FC to any networks. Interestingly, the spatial locations of amygdala subdivisions was variable across individuals. Taking into account the body of literature demonstrating relationships between clinical features (e.g., symptoms, response to treatment) and FC of the amygdala and hippocampus [74–82], delineating individual-specific FC in these structures appears critical for the individualized approach to medicine needed in psychiatry.

Within the cerebellum, stable maps of functional network representation have been characterized in two PFM studies [83, 84] (Fig. 3b, c). Although these studies differed in several ways (e.g., 10 vs 2 participants, 13 vs 10 a priori cortical networks), functional network representation and organization within the cerebellum was highly stable within individuals and similar across the studies. Network representation in the cerebellum exhibits more segregated delineation than observed in subcortical structures and topographical organization. While this cerebellar functional organization was more generalizable than that found in subcortical structures, both studies observed noticeable inter-individual variation of network topography. Given the implications for the cerebellum’s role in a number of neuropsychiatric disorders [85–87], again capturing inter-individual variability holds more promise for linking brain organization to collections of symptoms that are heterogeneous across patients.

Bringing together PFM and psychiatry

PFM has the potential to bring clinical utility to functional neuroimaging. Here we discuss three major directions in which PFM can benefit psychiatry. First, with precise and reliable estimates of an individual person’s brain function, we can improve our ability to relate individualized brain features to individual differences in clinical and behavioral features. Second, individual-specific measurement of brain function (e.g., mapping an individual’s brain network topography, localizing peak FC) has the potential to guide precise target sites for intervention using brain stimulation, such as DBS or transcranial magnetic stimulation (TMS), or surgical methods. Third, PFM can allow for tracking individualized changes in brain function over development and disease trajectory or in response to treatment. These three paths for clinical relevance of PFM (illustrated in Fig. 4) have the potential to improve patient outcomes by moving toward a patient-specific understanding of the disorder, and guiding individualized treatment.

Fig. 4. Three promising paths towards clinical application of PFM.

The center brain image displays functional network topology from a single individual, representing one metric that can be measured with PFM (any metric of interest could be used). a Identifying individual-specific brain features that relate to clinically relevant individual differences. b Localization of individual-specific targets for intervention, here showing transcranial magnetic stimulation (TMS). c Tracking individual-specific changes in the brain over time due, here in response to treatment.

Linking brain function to clinical features

Given the significant heterogeneity in the majority of neuropsychiatric disorders, relating measures of brain function to clinical features is important for understanding this individual variability across patients. PFM affords the opportunity to reliably measure individual differences that may relate to clinical features, but that are muted in group average neuroimaging data. Additionally, recent evaluations of effect size and reliability of brain-behavior associations have found that reproducible results require thousands of participants [88] (with standard amounts of data, e.g., 10 min resting-state fMRI data), marking a need for larger, public and collaborative datasets. Therefore, the road to effective neuroimaging research that translates to clinical utility lies in results informed by both PFM and extremely large, group average studies [89, 90].

Many neuropsychiatric disorders, such as ADHD, ASD, and learning disabilities, often share clinical and diagnostic behavioral traits as well as comorbid symptoms, [91–93] which adds considerable complexity to linking unique brain network features to specific disorders. One domain that has been considered transdiagnostic is cognitive control, which consists of many subdomains (e.g., inhibition, working memory, flexibility) that are affected sometimes similarly across different diagnostic disorders. It has been proposed that the shared variance in group-level assessment of cognitive control, which may be a broad, overlapping feature of multiple disorders, could reflect a bias in the particular tasks studied. Since tasks are grouped into domains conceptualized to capture a general underlying executive function, this approach may result in the assumption of domain generality across disorders [94]. PFM methods can help elucidate clinical relevance of group differences by parsing out domain generality with a more precise understanding of non-relevant individual variability. Additionally, the development of more individual-focused assessments that incorporate PFM may also help to disentangle those overlapping [95] or similar behavioral traits that can complicate diagnosis or treatment. One such potential use case is distinguishing ASD from temporal lobe epilepsy [96], which can share similar physiology and behavioral commonalities [97–99], making accurate diagnostic discrimination difficult at times.

There is evidence that individual-focused fMRI methods may aid in a nuanced understanding of clinical features relevant to an individual or subtype of a disorder [100–104], which can streamline appropriate treatment options. PFM work examining personalized brain activation associated with response inhibition via a stop-signal task showed individualized activation maps that deviated from the group-average activations, as well as individualized topography of six network parcels of interest [105]. The individualized parcels showed unique brain activity that was related to response inhibition performance, suggesting improved accuracy compared to atlas-derived parcels. There is also evidence, though not PFM, that computational techniques aimed at modeling individualized network organization with standard amounts of data may show utility in accurately predicting dimensions of psychopathology while taking into account individual variability [106]. Using spatially regularized nonnegative matrix factorization, individualized functional network maps significantly predicted dimensional scores of psychopathology [107]. There is also evidence that dimensions of anxiety (state vs trait anxiety) are differentially associated with cortical representation of functional networks: trait anxiety associated with the DMN, though inconsistently depending on the node location, and state anxiety associated with both salience and DMN nodes [108]. The reported inconsistency of node location within the DMN that is associated with trait anxiety may be due to the interdigitation of sub-components within the DMN that has been shown with PFM [34]. The use of PFM can illuminate the individual variation of network topography that may be driving this inconsistency, resulting in a more precise understanding of cortical regions (or sub-components) within the DMN associated with trait anxiety. Although clinical use of neuroimaging data in psychiatry is not currently common practice, this improved understanding of individual differences in neural circuits that are applicable to actionable clinical tools [107] or data-driven predictive models aiding in personalized diagnoses [100] may help drive the transition to standardized clinical use and personalized treatment options that better cater to individuals’ needs or even predictions of clinical effectiveness of treatment options [109](preprint).

While there are currently no peer-reviewed publications reporting PFM-derived results that link differences in brain networks to clinically relevant features in psychiatric patients, this work is currently underway. A recent preprint [110] leveraged 11 datasets that included both PFM data and standard-collection, large-sample, longitudinal data from the Adolescent Brain Cognitive Development study [111]. This work found a two-fold expansion of the frontostriatal salience network in most individuals with depression. This expansion was both stable over time and detected in childhood before the onset of depression symptoms that occurred later in adolescence. This study is the first of its kind, outlining a proof of concept for methodology that can recruit both PFM and large, longitudinal datasets to more accurately predict the emergence of psychiatric symptoms over time.

Individualized mapping for personalized treatment targeting

The high-resolution, individual-specific measurement and representation of functional networks and connectivity provided by PFM can help with individualized targeting for brain stimulation treatment methods [112]. TMS is an evolving option for first line treatment-resistant neuropsychiatric disorders, such as major depressive disorder, that is becoming more widely accessible. By its nature, TMS is also a prime use case for the clinical application of PFM, as target sites for TMS show greater efficacy of treatment with individualized network maps [113]. While the dorsolateral prefrontal cortex (DLPFC) has been established as a target TMS site for the treatment of depression [114], clinical outcomes vary among patients. Such variable efficacy is thought to be a product of suboptimal targeting, most likely due to individual variation in functional network connectivity [115]. A pair of studies using PFM data suggest that using individualized functional network connectivity maps (rather than previously defined cortical locations) to establish target sites for depression may improve the antidepressant efficacy of TMS treatment [116, 117]. DLPFC target sites whose FC was anticorrelated with the subgenual cingulate cortex demonstrated better antidepressant outcomes [118]. This region of anticorrelation in the DLPFC aligns with one node of the DAN, yet its precise location varies between individuals [119, 120]. Across the two studies, individuals’ cortical regions in the DLPFC that showed strong correlation with the DMN and strong anticorrelation with the DAN, most reliably matched subgenual cingulate connectivity profiles, establishing a reliable and individualized connectivity-driven method for identifying target TMS sites for depression. Further, there is evidence that individualized TMS sites developed from PFM data are both stable over time and heritable [116]. Although methodological best practices for establishing individualized target sites and reducing off-target effects during TMS application are still developing [117, 121–123], the evolution of TMS treatments may act as a guide for the development of best practices of using PFM data to improve other treatment options.

The highly individualized nature of PFM also shows great promise in potentially improving surgical outcomes for patients in a hospital setting. The inclusion of functional neuroimaging data, alongside traditional structural imaging, may aid in the advancement of neurosurgical applications, or even reduce the need for some invasive pre-surgical procedures. In the case of brain tumors, healthy and viable brain tissue can be surrounded by tumor tissue, calling into question if the healthy, neighboring tissue can be preserved [124, 125]. Recent work addressing the utility of fMRI as an aid to presurgical procedures posits that PFM can delineate neighboring tissue that is functionally intact, thus reducing the negative impact of tumor removal [126]. Further, cortical stimulation was used to verify fMRI results, indicating that around 33.2% of tumor-invaded cells, across 20 patients, were functionally preserved. This work highlights one possible avenue for preserving healthy brain tissue during surgical removal of brain tumors, though further investigation is needed. Specifically, a number of considerations that impact fMRI data must be reviewed, such as medication interaction with the BOLD response, patients’ ability to complete a scan (especially when task fMRI is needed for personalized mapping; e.g., motor function), motion artifacts, and data processing considerations, to name a few [127]. Although the use of PFM for presurgical mapping is barely in its infancy, future work can establish best practices and the potential for improving patient outcomes is promising.

Tracking clinical and treatment progression

The majority of psychiatric disorders involve symptoms that change in their nature and severity over time, the trajectory of which can vary widely across individuals with the same diagnosis. As such, establishing accurate measures to track such longitudinal change within an individual can aid in understanding disorder progression and inform decisions in treatment planning. PFM provides an approach to track the neural correlates of clinical features over time, allowing for individual-specific measures of how severity and symptoms are progressing. While current PFM studies most often concatenate data collected over time to reach sufficient amounts of data for high reliability, individually-precise longitudinal tracking will require reaching such reliability at each timepoint of interest. One intriguing, though speculative, potential use case would be tracking functional network variants, as these brain regions of high inter-individual variability have been shown to be stable within an individual over time, suggesting trait-like properties [29]. If network variants that relate to clinical features are identified, tracking their function and connectivity over time may provide key insights into the trajectory of trait-like characteristics of a disorder.

PFM can also offer precise individual measures for treatment response tracking [128] by longitudinally measuring changes to brain function over the course of an intervention. In cases of interventions that are expected to alter network organization or the strength of within- or between-network connectivity towards that of the normative control population, a more precise measure of FC for that individual is necessary. For example, when translating the research findings of disrupted frontostriatal salience network organization associated with depression [110] to clinical intervention, individual-specific metrics of FC pre-, during, and post-intervention can be used to quantify reductions in salience network overrepresentation and track treatment progression. Since FC has been shown to rapidly change within 48 h of significant physiological disruption, specifically casting of a limb [58], individually precise measures of changes in FC can be viable estimates of treatment progression over a relatively short timescale. Given the level of individual specificity desired for this type of clinical application, PFM data will be vital. Inaccurate or noisy representations of FC may imply change that is falsely attributed to treatment, rather than methodological error.

Future research directions

Advancements in methodology to aid clinical applications of PFM

Group average functional neuroimaging research has faced barriers with clinical utility [129]. While PFM may overcome the barrier of individual-specificity of the data, other factors must be considered for the adoption of PFM methods in clinical settings. One such factor is the additional skill and knowledge required to accurately apply PFM methods. Preprocessing methods for PFM data are currently less automated than standard fMRI preprocessing pipelines, for example by processing data in a given subject’s native space rather than using standardized atlases (see methods in Gordon et al. 2017, Laumann et al. [22], and Poldrack et al. [21] for details on suggested PFM-specific preprocessing methods). As modern fMRI preprocessing pipelines have become more accessible to researchers of all technical backgrounds through software advancements, such as the release of containerized pipelines like the Human Connectome Pipeline [130] and fMRIPrep [131], and as tools focusing on PFM methods advance and become more widely available, it is reasonable to expect that the need for expert intervention at the technical stage can be minimized, allowing experts to focus on the analysis and implementation of intervention.

Another significant barrier is the financial cost and patient burden associated with the collection of hours of fMRI data from individuals. Fortunately, rapidly advancing methodological options show great promise in reducing this burden. Recent advancements in multi-echo acquisitions for functional neuroimaging have shown a dramatic reduction in the amount of fMRI data needed to produce accurate and reliable functional network maps at the individual level [132, 133]. While a set of standardized best practices are still evolving, preliminary results using multi-echo fMRI for PFM show a considerable reduction in the amount of fMRI data needed per individual to measure accurate, precision level FC. One recently published result shows that 10 min of multi-echo fMRI data can produce better test-retest reliability than 30 min of standard, single-echo fMRI data [132].

Considerations for the clinical application of PFM

In order for PFM to have clinical utility, we need to consider the criteria required to move it from the research to the clinical domain. One key criterion should be high reliability of FC metrics for an individual, in line with our operational definition of PFM. Several of the foundational PFM studies described above established increasing reliability of FC and network measures with increasing amounts of data in healthy adults [24]. Depending on the specific metrics of interest with varying spatial scale (e.g., network-level connectivity, voxelwise connectivity) and which brain structures were of interest (e.g., cortex, subcortex), these studies showed that 30–90 min of low-motion fMRI data can achieve reliability r > 0.9 [22, 24, 70, 83]. This level of reliability is promising, yet entails substantial data collection often beyond 30–90 min, as most individuals will have data points exceeding strict motion criteria that will need to be removed. Even so, establishing this high level of reliability is an important first step. However, reliability measures have not yet been reported for other populations, such as children, older adults, and people with psychiatric disorders. While these populations tend to exhibit higher levels of in-scanner motion compared to healthy, young adults [134–136], strategies such as pre-scanning practice, mock scanner exposure, and real-time motion monitoring [134] can help mitigate motion for these individuals. Importantly, those most in need of personalized care can be the most highly motivated to endure repeated scanning with the hopes of better outcomes (as evidenced by individuals willing to complete multiple scans for presurgical mapping). There is currently work underway by our group and others to establish feasibility of collecting PFM data in populations that are presumed difficult to scan, e.g., children with Tourette syndrome [137] and older adults with Parkinson’s disease [138]. Assessing reliability in these populations is the next important step.

A second criterion is demonstration that the estimates we are measuring capture characteristics of the person and are not simply epiphenomenal to fleeting thoughts and behaviors. For example, if network architecture changes when someone is thinking about their grocery list vs. thinking about a favorite song, such measures would be less likely to capture brain features related to a chronic psychiatric disorder. At the same time, these measures should be able to detect large changes expected to affect clinical phenotypes, such as trauma or treatments. As discussed above, estimates of FC derived from large quantities of data using PFM are stable within an individual across varied time scales (i.e., days or months) in healthy adults [22, 24, 25], while also sensitive to neuroplasticity caused by perturbation in the cases of perinatal stroke and arm casting [55, 58]. Again, while this second criterion has been demonstrated in healthy adults and one specific case study, stability and sensitivity to neuroplasticity needs to be investigated and established in neuropsychiatric populations.

Once it is established that brain metrics measured with PFM are highly reliable and do, in fact, represent the intrinsic properties of brain function for an individual across different populations, what are the next steps necessary to apply PFM methods directly to the clinic? A next major criterion will be to establish that these metrics underlie symptoms and clinical phenotypic presentations. It is possible that some facets of individual variation are epiphenomenal (like eye color) and will not relate to individual differences in clinical phenotypes. Therefore, future work must test and establish which metrics of brain function are clinically relevant and which are not, as discussed above. Considering the recent findings that brain-behavior correlations can only be estimated reliably in extremely large samples (i.e., thousands of participants) [88] - albeit with standard amounts of data - the collection of PFM data on thousands of individuals would be untenable. A thorough evaluation of the impact that collection of more data (and the resulting increase of intra-individual reliability) has on effect sizes and reliability estimates of brain-behavior correlations is needed. If we have 60 min of high quality, low-motion data per person, would we still need thousands of participants for reproducible brain-wide behavior correlations [88]? A recent preprint has directly tested the influences of scan time and sample size when predicting brain-wide associations of behavior, finding that sample size and scan time are interchangeable at a certain point [139]. For specific populations or disorders where a sample size of thousands may not be attainable, this work suggests that smaller samples with longer scan times can achieve similar brain-wide association accuracy that is thought to require thousands of participants with shorter scan times. The authors also point out that this interchangeable relationship between scan time and sample size is not considered with standard power calculations and hence developed an online calculator to help future research studies balance these factors.

There is also promise that less data may be required with ongoing advancements in MRI technology, such as the use of multi-echo sequences that may reduce the amount of data to ~20 min to achieve the same intra-subject reliability seen with one hour from standard sequences [133]. If such results are confirmed and generalize, acquisition of 20 min of low-motion multi-echo fMRI data is a much more achievable goal within the scope of currently available resources. Hence, the ability to test for reliable brain-clinical feature relationships may become within our grasp. In addition, we will need to test if these individually-defined clinically relevant features guide clinical care substantially better than group-defined features. For example, to test the efficacy of individually-defined TMS targets, a randomized controlled trial would need to demonstrate that individually-defined sites of stimulation lead to more clinically meaningful improvement than anatomically-defined sites. Thus, a group-level study would test the efficacy of the individual-level approach.

An ultimate criterion for clinical adoption of PFM is a deep evaluation of the cost-benefit ratio of collecting and interpreting such fMRI data from patients seen in the clinic. Indeed, MRI scans are expensive and not typically administered clinically unless necessary. Thus, it will be pivotal to be able to collect enough data in as short of time as possible. If the above criteria can be attained with one single MRI scan session at the hospital, it is theoretically possible that such a scan could become a more standard medical test that would not be overly burdensome for many patients. Of course, one issue among many is that MRI data and FC measures in particular are highly susceptible to small, submillimeter head movement during scanning [10, 140]. Therefore, certain individuals would be unlikely to complete the scan, unless the medical benefit of obtaining MRI data was large enough to opt for sedation. Thus, the benefits to the patient will need to be clearly established. These benefits will then need to be evaluated against the cost for different strategies of administration. For example, what is the cost-benefit ratio for collecting scan data on potentially every patient seen in a psychiatry clinic? Alternatively, what is the cost-benefit ratio of a more titrated approach, in which first line treatments are administered before the inclusion of MRI data, and only those who do not respond well to that treatment would undergo an MRI scan? Ultimately, determination of the benefits and costs association with PFM in psychiatry will require immense effort and resources. At present, we can continue to push the field forward testing the first two criteria discussed above with hopeful promise for the future.

With personalized medicine moving to the forefront of discussion on improving patient outcomes, PFM is poised as a promising tool toward accomplishing this goal. As more work using PFM methods is conducted with a focus on psychiatry, and with improving methods for measuring individual brain features, we hope to move towards meeting the criteria needed for PFM use in the clinic, ultimately improving clinical outcomes for neuropsychiatric disorders.

Author contributions

DVD and DJG drafted and reviewed the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by NIMH grant R01 MH118217.

Competing interests

The authors declare no competing interests.