Neuroimaging for precision medicine in psychiatry

Abstract

Although the lifetime burden due to mental disorders is increasing, we lack tools for more precise diagnosing and treating prevalent and disabling disorders such as major depressive disorder. We lack strategies for selecting among available treatments or expediting access to new treatment options. This critical review concentrates on functional neuroimaging as a modality of measurement for precision psychiatry, focusing on major depressive and anxiety disorders. We begin by outlining evidence for the use of functional neuroimaging to stratify the heterogeneity of these disorders, based on underlying circuit dysfunction. We then review the current landscape of how functional neuroimaging-derived circuit predictors can predict treatment outcomes and clinical trajectories in depression and anxiety. Future directions for advancing clinically appliable neuroimaging measures are considered. We conclude by considering the opportunities and challenges of translating neuroimaging measures into practice. As an illustration, we highlight one approach for quantifying brain circuit function at an individual level, which could serve as a model for clinical translation.

Introduction

Mental health disorders pose a significant global public health challenge [1] and a state of emergency has been declared [2]. Anxiety disorders are the world’s most common mental disorders, affecting over 300 million people, as reported in the 2019 Global Burden of Disease study [1, 2]. With the COVID-19 pandemic, the global prevalence of major depression and anxiety disorders increased by a dramatic 27.6% and 25.6% respectively [3]. Depressive and anxiety disorders are the top ranked contributors to years and days lost to disability globally among all age groups [4]. All too often these disorders are fatal, and suicidal ideation in youth is surging, outstripping clinical capacity [5, 6].

Treatment of these disorders has been hindered by etiological and phenotypic heterogeneity. Depressive and anxiety disorders involve the dysfunction of multiple and overlapping neurobiological processes that, in turn, would likely each require a different treatment. This is evident by the fact that as many as one third to one half of patients with major depressive disorder do not respond to treatment, even after multiple attempts [7]. Precision medicine applied in psychiatry seeks to address this need. The precision strategy uses objective testing to personalize treatments for an individual, thereby moving away from the current one-size-fit-all approach. ‘Stratified psychiatry’ is an intermediate step that aims to parse the heterogeneity of depression and anxiety into biologically coherent subtypes that implicate different treatment approaches [8–10]. This strategy seeks to determine which treatment will ameliorate which root cause of the illness. The goal is to quickly rule out ineffective treatments and expedite effective treatments, thereby substantially increasing the number of people who achieve remission on their first treatment.

In this review we focus on the contribution of functional magnetic resonance imaging (fMRI) for advancing precision medicine in psychiatry, visualized in Fig. 1. fMRI provides a direct measure of the organ of interest for precision psychiatry, the brain [11]. Functional neuroimaging markers haves greatly enhanced the cognitive neuroscience understanding of individual variation in the human brain in both health and disease offering the opportunity to personalize clinical practices that lead to better outcomes for people [12].

Fig. 1. A conceptual overview of precision psychiatry informed by functional neuroimaging.

Panel 1 illustrates the heterogenous nature of major depression and anxiety both within the disorders and their overlap. Panel 2 shows a path forward for precision medicine in psychiatry incorporating direct measures of neural circuit function underlying clinical heterogeneity using techniques such as functional magnetic resonance imaging. Panel 3 illustrates how measures of neural circuit function can be used to aid in more personalized selection of treatments.

We first provide an overview of brain circuits identified using fMRI where there is consensus regarding their implication in disorders like major depressive disorder and anxiety disorder, in both adults and adolescents. We then illustrate the use of circuit measures for stratifying the heterogeneity of these disorders into biological subtypes, identifying predictors of treatment outcomes and clinical trajectories. We explore future avenues to enhance the clinical implementation of fMRI tools and imaging metrics.

Grounding precision psychiatry in neuroimaging

Key to precision approaches is the focus on measurements, paralleling recent advances in other medical fields. Just 74 years ago, cardiovascular medicine was limited by the lack of imaging tools such as ultrasound or computed tomography. This prevented quantitative, personalized assessments of the heart’s structure and function in relation to observable symptoms. Today, in cardiology heart imaging during both rest and stress conditions is the gold standard in patient management (Fig. 2). Commonly, the presenting symptom is broad such as the experience of chest pain. Having a direct means to observe the heart provides information about whether the source of the chest pain is structural, such as whether the heart is enlarged or heart muscles are thickened, or functional, such as a problem with blood flow. What if these principles of precision medicine already at work in cardiology - and in other fields of medicine—were applied to psychiatry? We can envision an analogous approach in which imaging during rest and task conditions identifies specific circuit dysfunctions in depression and anxiety, aiding in selecting treatments that are personalized for targeting these dysfunctions (Fig. 2). Supporting this view, fMRI in both rest and task conditions in MDD has been found to predict response to antidepressants for each patient with at least 75% accuracy over and above baseline symptom severity [13–15]. fMRI in rest and task fMRI also show differential moderation of treatment response for antidepressants [13, 15, 16]. Similarly, baseline rest and task fMRI in social anxiety also predicts response to CBT with up to a fivefold improvement over baseline symptom severity [17, 18]. This perspective assumes that understanding the brain is a necessary foundation for advancing precision psychiatry. It does not presume that imaging is the only measure important for diagnosis and treatment. Rather, once we have a foundation in the brain, we have a basis from which to specify the relationships between brain circuits, behaviors, and symptoms and to interpret clinical profiles with respect to underlying neurobiology.

Fig. 2. An illustration of the analogy between precision medicine in cardiology and precision medicine in psychiatry.

An illustration of the analogy between cardiology, in which imaging of the heart at rest and during stress tasks is a gold standard for aiding diagnosis and treatment decisions (panel A), and the vision for precision medicine in psychiatry in which imaging of the brain at rest and during tasks is used to aid diagnosis and treatment decisions for mental disorders (panel B).

A challenge for precision psychiatry is the current lack of clinically interpretable measures for quantifying neurobiological dysfunctions in individual patients, unlike the situation in cardiology.

A number of national initiatives have been launched to make progress toward addressing this challenge [19–24]. These initiatives all include a focus on neuroimaging measures, along with other biomarker measurements. In the following section we review the larger body of work identifying human brain circuits, or networks, anchored in large scale studies of the healthy brain, then applied in clinical samples.

What do we know about human brain circuits implicated in psychiatric disorders?

Over the past two decades, fMRI studies examining brain regions in the resting state and evoked by tasks have revealed a neural circuit architecture, underpinning the intrinsic, domain-general, and task-related processes of human brain function [25–31]. These processes include self-reflection, salience perception, attention, sensorimotor functions, sensory processing, and reactions to emotional stimuli.

The universality of the intrinsic architecture has been demonstrated with large-scale functional connectivity analysis of hundreds of brain regions encompassing every major brain system at rest and across 64 task-evoked states [27]. During rest, the default mode circuit tends to be upregulated, and other circuits downregulated [27, 30, 32]. Engagement of the default mode may also be driven by the context of responses to task stimuli, whether internally or externally focused [33]. In turn, intrinsic circuit architecture might also have a major role in shaping task processes, potentially with a smaller but important amount of variance contributed by specific task-evoked changes [27]. Regions of activation and connectivity contributing to task-evoked circuits have been identified in systematic reviews and meta-analyses of a large body of studies using cognitive and emotional task stimuli [31, 34, 35]. There is some overlap in the terminology used to describe resting state and task-evoked circuits, such as the frontoparietal central executive network at rest and the cognitive control network evoked by cognitive tasks. An elegant experimental investigation has shown that tasks differentiate the dorsolateral prefrontal cortex (dLPFC) and anterior cingulate cortex (ACC), extending to the pre-supplementary motor area, from the resting state [25].

In the following sections we highlight how dysfunctions in these resting and task-evoked circuits can inform the stratification of heterogeneity within major depressive disorder to identify subgroups—or ‘biotypes’—with common circuit profiles.

How circuit dysfunctions can unravel the heterogeneity of major depressive disorder

Substantial progress has been made in using fMRI-derived circuit measures in parsing the heterogeneity of major depressive disorder (MDD) into brain-based subgroups or biotypes. We acknowledge that progress has also been made in other areas [36], which are beyond the scope of the present review.

‘Stratified psychiatry’ is an intermediate step toward personalized medicine in psychiatry that aims to parse the heterogeneity of mental disorders into biologically coherent subtypes, or biotypes. Stratification of biotypes is an important step in enabling more personalized approaches to treatment selection as biotypes can implicate different treatment approaches [8–10]. Seminal studies characterizing biotypes of MDD with similar brain circuit dysfunctions have implemented data-driven approaches with resting fMRI data [37–40]. For example, one pioneering study has found biotypes characterized by aberrant connectivity in fronto-striatal and limbic networks that respond differently to repetitive transcranial magnetic stimulation (TMS) (Drysdale et al., 2017a). These resting fMRI biotypes, accounting for clinical profiles of depressed mood, anhedonia, anxiety, and insomnia, have been found to generalize in a large single site sample [41]. Other researchers have found biotypes characterized by hyper- and hypo-connectivity of the default mode network (Liang et al., 2020); biotypes that distinguish comorbid anxiety within the context of depression (Price et al., 2017); and biotypes that are associated with poorer response to standard antidepressants (Tokuda et al., 2018). In adolescents at risk of internalizing symptoms such as depression, a multi-model estimation approach using resting state connectivity has identified two biotypes characterized by diffuse connectivity and hyper-connectivity, distinguished by prior internalizing symptoms and by future problems, respectively [42].

A complementary approach has pursued stratification of circuit biotypes using a hybrid approach in which data-driven clustering has been applied to a theoretically informed set of circuit metrics derived from three task-free and three task-evoked circuits implicated in MDD [43]. This approach has identified six biotypes defined by distinct profiles of intrinsic task-free functional connectivity within the default mode, salience and frontoparietal attention circuits and of activation and connectivity within frontal and subcortical regions elicited by emotional and cognitive tasks. These biotypes were also distinguished by symptoms, behavioral performance on general and emotional cognitive tests, and by response to pharmacotherapy as well as behavioral therapy. The findings suggest that engaging circuits with tasks may uncover complementary insights that may not be uncovered in analyses of only task-free data or with large numbers of circuit features.

These findings indicate that both theoretically based and data-driven approaches are necessary to provide complementary insights in different contexts and applications. To prospectively identify biotypes in future studies and clinical applications it will be useful to have a relatively small set of circuit metrics that can be interpreted for each individual patient. However, fully data-driven approaches will be necessary for revealing unexpected results that are not currently incorporated into the theoretical foundations of our field. Both theoretical and data-driven approaches will also be essential to determining the extent to which circuit dysfunction identifying biotypes is the same or different as the circuit dysfunction that predicts treatment efficacy.

Circuit predictors and treatment selection

In this section we provide an overview of fMRI studies that have used randomized controlled or well-designed open label trials to identify circuit predictions of treatment response. An accumulating body of work demonstrates that circuit measures have promise as predictors for commonly used pharmacological and behavioral therapies in MDD. We acknowledge there are additional important therapeutic areas such as neuromodulation, deep brain stimulation, and rapid acting exploratory treatments, but these are beyond the scope of the current review.

In MDD, several collaborative biomarker trials have identified both resting and task-evoked circuit predictors of pharmacotherapy outcomes. These trials include PReDicT (Predicting Response to Depression Treatment) [44], EMBARC (Establishing Moderators and Biosignatures of Antidepressant Response in Clinical Care) [45], iSPOT-D (international Study to Predict Optimized Treatment in Depression) [46], CAN-BIND (Canadian Biomarker Integration Network in Depression) [47] and the Netherlands Study of Depression and Anxiety (NESDA) [48, 49]. Because these trials have focused on first and second-line antidepressants, the findings have implications for broad clinical implementation. As highlighted in these trials, circuits implicated in treatment prediction are broadly aligned with neural systems articulated in the Research Domain Criteria (RDoC) framework [50] (Table 1), specifically, negative valence, positive valence, cognitive systems and the cross-cutting default mode and resting networks.

Table 1.

A summary of fMRI predictors of treatment response and differential moderation of response in major depression, focusing on randomized controlled trials, for which there is consistency across studies and sites.

Circuit	Treatments	Design	Predictors	Treatment outcome	Classification metrics	Convergent evidence	Systematic review
Default Mode, resting fMRI	Escitalopram, Sertraline, Venlafaxine-XR	RCT	Higher baseline PCC-amPFC, default to attention/CEN Lower baseline PCC-amPFC, default to attention/CEN	Greater remission Rate of symptom reduction Lower remission	62–74% sensitivity* [14, 55] Effect size 0.40** [56] 73% specificity* [14, 55]	Well-designed open label trials [59, 134] [60]:[61]; [62]; challenge and single dose studies [63, 64]	Yes [135, 136]
Salience, resting fMRI	Citalopram, Escitalopram, Sertraline, Venflaxine-XR, CBT	RCT	Lower baseline sgACC-insula, amygdala-insula	Higher remission Higher response Lower remission, CBT	72% remission classification [16] Effect size, 0.70 [57] 89% non-remission classification [16]		Yes [136]
Negative affect, emotion task fMRI	Citalopram, Escitalopram, Sertraline, Venflaxine-XR, CBT, PST	RCT	Baseline implicit fear-elicited amygdala hypo-activation Baseline implicit sad-elicited amygdala hyper-activation Early change in implicit fear elicited amygdala activation	Higher response to SSRIs Lower response to venlafaxine Symptom improvement with PST	75% accuracy* [13] 81% accuracy* [13] Effect size,0.47 [75]	Well-designed open label studies, responders to escitalopram show early fear-elicited amygdala attenuation [66, 67] and responders to a course of SSRIs show amygdala normalization [68–72]; Non-response to escitalopram predicted by higher fear-elicited amygdala-ACC connectivity [65]. Explicit sad-elicited amygdala activity predicts response to scopolamine[73]. Implicit sad elicited amygdala activity normalizes with CBT [74]	Yes [137]
Cognitive control, GoNoGo task fMRI	Escitalopram, Sertraline, Venflaxine-XR Duloxetine	RCT	Baseline dLPFC hypo-activation Baseline dLPFC task connectivity	Poor remission to SSRIs Differential response to sertraline v venlafaxine	Effect size,0.91 [84] Sertraline: effect size, 1.37, Sensitivity, 95%, specificity 74%*** [15] Venlafaxine: effect size, 1.19, Sensitivity, 95%, specificity 88%*** [15]	Well-designed open-label trial, baseline dACC activation differentially predicts response to escitalopram v duloxetine, 84% accuracy* [85]

^*Values are for prediction over and above baseline symptom and demographic measures, with leave- one out cross validation (LOOCV); **Values are with LOOCV; *** Values are with LOOCV with nested validation, a procedure in which an inner cross-validation loop is used to tune the model parameters while an outer cross-validation loop is used to estimate classification error; Effect sizes are in cohen’s d units.

Classification metrics for sensitivity, specificity and effect size are highlighted. Convergent evidence from well-designed open-label trials is also presented.

One set of findings highlights dysfunction in the intrinsic connectivity of the default, salience and frontoparietal attention or central executive networks. Profiles of both hyper-connectivity and hypo-connectivity have been observed in depression [51–54] and these distinct profiles may be complementary predictors of treatment outcomes. Relatively higher default mode circuit connectivity has been found to predict better response for antidepressants, including escitalopram, sertraline and venlafaxine, for both seed-based approaches [14, 55, 56] and for whole brain connectomics approaches [14] (Table 1). By contrast, lower default mode connectivity is predictive of lower remission in these studies. Seed-based approaches have demonstrated a leave-one-out cross-validated accuracy of 82% for predicting remission over and above baseline symptom severity, with a specificity of 74% and sensitivity of 73% [55] (Table 1). Connectomics-based prediction using the same data inputs also demonstrates significant improvements over baseline symptoms, achieving a cross-validated sensitivity of 63% and specificity of 72% [14] (Table 1). Higher connectivity between the default mode and frontoparietal attention or central executive network (CEN) is also involved in these predictive associations for both seed-based and connectomics approaches [14, 55, 56] (Table 1). More localized disruptions within default mode subnetworks also inform specific treatment associations. Hypo-connectivity within posterior nodes of the default mode circuit, quantified using a patient-level processing method, specifically predicts response to venflaxine [57]. Normalization of the default mode has been observed after antidepressant treatment in both treatment naïve depression [58] and late-life depression [59, 60], and changes in default mode connectivity show promise as predictor of long-term antidepressant outcomes [61]. By contrast, hyper-connectivity within anterior nodes of the default mode, quantified using amplitude fluctuation, characterizes treatment resistant depression (TRD) [62].

Resting functional connectivity involving regions of the default mode and CEN, has also been found to predict antidepressant response in studies designed to probe early changes in connectivity. For example, using a drug challenge design, one study found that differences in resting connectivity induced by citalopram compared to placebo predicted reductions in clinically rated symptom severity at 7 weeks post-treatment [63]. Another study reported that a signal dose of escitalopram dynamically modified resting-state activity in treatment naive depression, such that changes after 5 h of the dose in the caudate, occipital, and temporal cortices predicted clinical remission at endpoint [64].

In PReDICT, lower connectivity of the subcallosal cingulate with insula and ventral frontal regions was identified as a differential moderator of remission to escitalopram and treatment failure with CBT [16] (Table 1). Lower salience circuit connectivity has also been found to predict response to escitalopram in iSPOT-D, and also to sertraline and venlafaxine [57] (Table 1). In NESDA, lower insula connectivity within the salience network has been identified as a prospective indicator of insufficient response to pharmacotherapy for patients taking multiple medications [49]. While connectivity involving the insula has been found to predict non-response to CBT, reduced intrinsic connectivity within the frontoparietal attention circuit, also referred to as the CEN, has been found to predict better response to problem solving behavioral therapy in depression [57].

Emotion task fMRI has revealed circuit predictors of pharmacotherapy outcomes for both negative and positive affect circuits. Distinct types of amygdala-prefrontal activation predict response to different types of treatment. In adults with MDD, amygdala hyper-activation to implicit sad stimuli is a differential moderator of poor response to venlafaxine with 81% leave-one-out cross- validated accuracy [13]. Relatively lower amygdala activation to implicit fear stimuli, on the other hand, is a predictor of better response across SSRIs and antidepressants, while relatively higher amygdala activation to implicit fear is a predictor of lower response to these antidepressants, with 75% leave-one-out cross- validated accuracy [13] (Table 1). Hyper-connectivity of the amygdala and anterior cingulate evoked by fear stimuli also predicts poor response to SSRIs [65] (Table 1). Responders to SSRIs have also been characterized by an early attenuation of amygdala activation to fear [66, 67] and a relative normalization of amygdala activity after a course of antidepressant treatment [68–72] (Table 1). Amygdala hypoactivity elicited by explicit sad face stimuli has also been found to predict response to scopolamine [73] (Table 1). Amygdala and dACC activation to fear stimuli and dorsal anterior cingulate (dACC) activation to sad stimuli have been further implicated in response to cognitive behavior therapy [74] and problem solving behavior therapy respectively [75] (Table 1). The involvement of both hyper- and hypo-activation of the amygdala in predicting treatment response is consistent with evidence that both of these amygdala profiles have been observed in major depression in adults [69, 76, 77]. Regarding positive affect, reward-related ventral striatal activation has been identified as a predictor of response to sertraline versus placebo in the EMBARC trial [78]. In CAN-BIND, an early increase in ventral striatal and ACC connectivity from baseline to week two was positively correlated with subsequent clinical response to escitalopram [79].

Although these studies have focused on adults, a pilot study found that greater activation of emotion-elicited ACC and MPFC regions, together with greater amygdala resting connectivity, also predicts better outcomes with SSRIs in adolescents with depression [80]. Among the tiny number of studies investigating imaging markers of treatment outcome in adolescents, greater baseline striatal reactivity evoked by a reward task has been associated with the rate of symptom reduction during 8 weeks of CBT or CBT plus SSRI, in adolescents with major depressive disorder, quantified with a growth model [81].

During cognitive tasks like Go-NoGo requiring goal selection and response inhibition, hypo-activation of the dLPFC characterizes a distinct cognitive biotype of depression [82] and poor response to SSRIs [83, 84]. Dorsal ACC activity evoked by a parametric Go-NoGo task has also been identified as a differential moderator of response to the SSRI escitalopram versus the SNRI duloxetine [85]. Go-NoGo-evoked connectivity of the dLPFC within parietal cortical regions is also a differential moderator of response to SSRIs versus the SNRI venlafaxine, beyond symptom and demographic measures [15]. Higher connectivity distinguishes responders to sertraline and lower connectivity non-responders to sertraline with a nested cross- validated accuracy of 83% (sensitivity 95%, specificity 74%), while the opposing profile of lower connectivity distinguishes responders to venlafaxine and higher connectivity distinguishes non-responders to venlafaxine, with an accuracy of 77% (sensitivity 95%, specificity 88%) [15]. In late-life MDD, imaging metrics also predict treatment response beyond clinical and cognitive behavioral measures, while cognitive measures contribute to diagnostic prediction [86].

The above studies have demonstrated that fMRI metrics have predictive utility over and above baseline clinical behavioral characteristics in predicting antidepressant response in MDD. Some studies demonstrate fMRI metrics are differential moderators of response to one treatment versus another, a taller bar to clear than general response prediction based on classic definitions [87, 88]. These findings are exemplars of circuit metrics that could be developed for use in clinical practice, a topic we consider in the next section.

The road to clinical translation

To move circuit metrics from research into clinical applications, it is necessary to have imaging technology that can quantify circuits at the individual patient level and determine the extent of dysfunction in these circuits relative to reference norms. With individualized circuit metrics, it is then possible to prospectively select patients with particular impairments, and evaluate treatments selected to target these impairments. One such imaging technology was developed by LMW to quantify individualized circuit function relative to a healthy reference dataset for six circuits implicated in depression and depression treatments [57]. For example, within the cognitive control circuit, the imaging technology quantifies activation of the bilateral dLPFC and dACC, and the functional connectivity between these regions, and expresses each of these measures in standard deviation units relative to mean and standard deviation of a healthy reference dataset. These measures meet psychometric criteria for construct validity and internal reliability and show associations with clinical behavioral measures that generalize across independent samples [57]. Reliability for the cognitive control metrics is 0.77. The dLPFC-dACC cognitive control circuit metrics have been evaluated as personalized predictors of antidepressant remission in a secondary analysis of 159 patients from iSPOT-D who completed fMRI during a GoNoGo task, 8 weeks of treatment with one of three antidepressants and who were assessed for remission status. Using a prior established signal detection algorithm implementing receiver operating analysis (ROC) [89] the sensitivity and specificity of these predictors were calculated at every cut-point for each circuit measure. ROC models correctly identified 63% of remitters at an initial cut-point of −0.75 standard deviations for dLPFC activity, compared to the base rate of 36% without the inclusion of circuit predictors. To further assess the clinical meaningfulness of the model incorporating circuit predictors, we calculated the number needed to treat (NNT) metric. This NNT metric quantifies the additional number of patients that could achieve remission should the circuit model be used, compared to the current base rate without use of predictors. Applying the NNT formula to the circuit value of 63% compared to the base rate of 36%, the NNT value is 4. This value indicates that if the circuit predictors were used to inform treatment selection, then for every 4 patients assessed with these predictors, we would anticipate one less failure to achieve remission compared to the current situation based on symptom observation alone. Future clinical translational studies are needed to test the replication of these findings and evaluate their generalizability across sites and samples.

Creating an economic example may show the importance of future trials and the potential impact of using circuit predictors to improve remission outcomes. In current practice, we could expect 36 remissions after treating 100 patients, given the base rate of 36%. If fMRI circuit predictors were to work in the field, we could anticipate obtaining this rate of remission after treatment of only 58 patients who meet the circuit selection criteria, given the circuit rate is 63 and 63% of 58 is 36%. Thus, if the treatment cost is $5000 per course, the cost for treating selected patients to achieve remission would be $290,000 (58 times $5000) versus $500,000 (100 times $5000) for the same number of remissions in an unselected population. Such an approach at a health systems level might afford considerable savings, even if one had to spend say $58,000 to get 58 fMRI scans at $1000 each. Finally, it might also afford a considerable reduction in frustration of patients and physicians alike by avoiding a treatment that is unlikely to be effective.

Finally, boosting the number of patients achieving remission based on pre-treatment circuit predictors has the further potential to help drive down the longer-term costs due to burden of illness in depression. Reflecting some aspects of the burden of illness, for every employee experiencing depression, there is an average cost of $15,000 per year in lost productivity, health care costs and turnover [90]. Based on this annual cost, using the current clinical base rate of 36%, one individual achieving remission out of three would reduce lost productivity costs from $45,000 to $30,000. In the future, if the incorporation of circuit predictors enables two out of three individuals to achieve remission, this cost could be reduced further, from $45,000 to $15,000. Depression has a chronic course of disability, resulting in lifetime costs that are currently not routinely factored in when evaluating the introduction of new clinical tools, such as neuroimaging, in psychiatry.

This illustration highlights the potential for using fMRI-derived circuit tools in a precision psychiatry approach to select antidepressants and improve clinical outcomes. In the following section, we explore future pathways to enhance the clinical translational potential of this approach.

Future research and clinical translational directions

We highlight several future directions spanning new opportunities for fMRI designs and approaches and addressing the needs of translating fMRI-based biomarkers into precision trials and clinical settings.

Replication, reliability, norms, and standardization

To ensure the rigor and interpretability of fMRI measures in practice, there is a need to address methodological issues. Increasingly, there is a call for replication of approaches that stratify the heterogeneity of depression based on fMRI data and use fMRI to predict treatment outcomes. When using thousands of inputs in fully unsupervised approaches, there is the potential for obtaining overly optimistic results due to overfitting. As one example, lack of replication has been reported in relation to imaging-derived cluster biotypes [91] but addressed in a subsequent re-analysis addressing overfitting [92] and further in a large-single site study [41]. This example highlights one path forward to testing the stability of biotypes both within and across datasets. Another approach is to focus on regions defined from theory, meta-analyses, and anatomy. We have demonstrated that robust cluster biotypes can be replicated with held-out data using a tractable set of brain circuits defined by anatomically located regions of activation and functional connectivity, both at rest and during tasks [43].

There is also a need to establish the reliability of fMRI measures, internally and over time. Returning to our cardiology analogy, a goal might be to establish fMRI metrics that have equivalent reliability as those used for heart applications, and to systematically build the evidence base for reliability using common imaging protocols. In cardiac contexts, reliability has been found to differ with imaging modality. For example, cortical gray matter perfusion has relatively higher between-session variability (around 0.70) while fractional isotropy and, in particularly gray matter volume, are less variable (greater than 0.80) [93]. Multi-site reproducibility for brain MRI measures of 4D neurovascular flow has also been established (>0.75) in 10 subjects tested twice at each site [94]. Studies focused on evaluating fMRI reliability for psychiatry applications report low- to- modest reliability when focusing on between-individual variation. However, within-individual task-based activations show higher reliability for both cognitive and emotion tasks [95, 96]. Reliability estimates range from fair- to- excellent for within-individual reliability, particularly when activations are derived from anatomically defined regions of interest [95]. Reliability may also be more robust for peak activation as indicated in data from a precision targeting context [97]. Similarly, within-subject reliability estimates are also within the good- to- excellent range for resting fMRI [95], whereas between-subjects estimates for voxel-wise connectivity tend to be low [95, 98]. Relative to task fMRI, higher between-subject test-retest reliability has been observed for local activity measures of resting state fMRI, such as ALFF, fALFF and ReHo [95], Future studies could focus on within-subject estimates and within-individual variation as an indicator of amount of information carried by both task and resting fMRI. To advance clinical translation it will also be important to take into account amount of activation as a first reliability estimate in task fMRI, systematically evaluate consistency at different statistical thresholds for defining activation, and focus on tasks and clinical populations most relevant to the targeted clinical application [99].

In cardiology and in fields within psychiatry, such as neuropsychology, the clinical utility of objective measures is enhanced by reference to healthy and/or clinical norms. A method for using healthy referencing of fMRI data has been demonstrated [57] and used in the clinical illustration summarized in section 4 [100]. In this critical review, we presented an illustration using ROC to demonstrate how standardized and normed patient-level circuit measures could be used to optimize remission outcomes for commonly used antidepressant medications. This approach can be expanded by comparing the prediction for different classes of antidepressants, and by incorporating additional treatments. It would also be worthwhile to investigate the interaction between circuits and/or the convergence of multiple circuit dysfunctions. Here, we have focused on circuit dysfunction distinguished from the healthy range by a standard deviation threshold. Future analyses are also warranted to how circuit variation within the healthy range might combine with circuit dysfunctions outside the healthy range.

There is a need for large sets of norms acquired and harmonized across sites using standard sequences intended for clinical applications. To facilitate decisions around standard sequences, experts using fMRI could develop consensus guidelines about the choice of resting conditions, tasks, and quantification approaches. For task fMRI, norms could be used to clarify directions of dysfunction, in both depression subtypes and in predicting treatment response. For example, blunted amygdala activation observed during explicit emotion processing [76] may represent avoidance of salient emotions, particularly if there is sufficient time for processing the emotional content of stimuli whereas hyper-activity to subliminal negative emotion [69, 77] might reflect the priming of automatic biases.

Neuromodulation and emerging treatments

The availability of psychometrics and norms for fMRI metrics could be an important foundation for rapidly expanding discovery of predictors for other treatment modalities. Although discussion of these other modalities, including neuromodulation, selective and psychedelic approaches, is beyond the scope of this review, the common thread is the need for circuit predictors that apply across treatment modalities. Trials of these new approaches are occurring largely without the use of baseline predictors to stratify heterogeneity and identify individuals who are not responding and why. When imaging is used, the sequences and/or tasks vary considerably across trials, even though each trial is elegant within its own design. As a field, we may be missing an opportunity for exponential advances without the use of some common methods. Both resting and task-evoked fMRI measures demonstrated utility for many emerging new treatment approaches, spanning neuromodulation, selective compounds and rapid-acting interventions [101], as well as behavioral approaches such as mindfulness [102–104] and exercise [105, 106].

Novel precision imaging and personalized designs

The availability of predictive circuit metrics opens up several opportunities for novel treatment designs. Prospective designs will help move the needle on new strategies for finding the right treatment for the right person at the right time, beyond assuming that one treatment will be effective for all patients within a disorder class. Prospective trials can now enrich for samples that have the circuit dysfunction of interest and stratify patients based on fMRI measures of circuit dysfunction. Such designs are required to evaluate circuit dysfunction as biomarker for predicting which individual patients respond to treatments selective for this such dysfunction, and as a response biomarker of outcomes. Another novel treatment direction may be referred to as “personalized therapeutics” in which intrinsic networks and network interactions are identified in an individualized manner and subsequently targeted with interventions adapted for the individual. For example, mindfulness meditation can be used to train patients how to down regulate their individually identified default mode network relative to their frontoparietal network via real-time fMRI neurofeedback [104, 107–109]. Training patients how to modulate neural dynamics on a network level with neurofeedback may be a more personalized method of neural regulation than neuromodulation involving a single region.

Future prospective designs could also assess outcomes over a longer time frame than the acute time frame of approximately 8 to 12 weeks used in most prior imaging trials. To date, the emphasis of imaging studies focused on prediction of treatment outcomes has been on acute response and remission outcomes. Complementing this approach is the emergence of approaches focused on modulation of short-term dynamic fluctuations in mood and mental state by coupling real-time fMRI with online experience sampling (ES) (e.g., [110–112]). This methodology can be applied to develop novel predictive models of different clinical features which can be subsequently used to inform personalized, adaptive interventions, such as neuromodulation, which are applied in a ‘just in time’ manner to target immediate changes these clinical features.

Multi-modal predictors

A common repository of psychometric criteria and norms would also be a strong foundation for expanding to multi-modal imaging predictors. Already, evidence from social anxiety disorder highlights the promise of multi-modal baseline imaging measures for predicting outcomes with CBT [17, 18]. For example, a combination of theoretically-informed quantification of resting state connectivity seeded from the amygdala, a multi-voxel pattern analysis of whole-brain resting state connectivity and probabilistic diffusion tractography of the right inferior longitudinal fasciculus has been found to accurately predict improvement in social anxiety after CBT [18]. Each connectomics measure improved the prediction of individuals’ treatment outcomes significantly better than a clinical measure of initial severity, and combining the multimodal connectomics yielded a fivefold improvement in predicting treatment response. Generalization of the findings was supported by leave-one-out cross-validation. In considering future clinical translation, the cost-effectiveness case for fMRI can be expanded to consider combining multiple sequences within one baseline session. Further, if fMRI scans were part of a standard clinical protocol, then other measures, such as genetic variants [113] and health records could also be included in the predictive modeling of treatment outcome.

Trajectories

A natural expansion of research identifying predictors of treatment is to identify predictors of risk, with the goal of ultimately enhancing preventative strategies even prior to overt illness. This goal is compelling given that approximately 25% of adolescents in the United States are diagnosed with MDD, the leading cause of disability and the second-leading cause of death in this age group. In a normative pediatric sample it has been shown that resting state network metrics based on theoretically informed regions of interest predict the subsequent trajectory and progression of internalizing depression and anxiety symptoms over four years, significantly better than baseline clinical measures [114]. Specifically, resting functional connectivity of the subgenual anterior cingulate cortex (sgACC) and dLPFC at age seven predicted worsening of symptoms at age eleven. Distributed cortical and subcortical profiles of functional connectivity also predict symptom severity in currently healthy children as well as in depressed and anxious children and adolescents [114, 115].

Characterizing circuit metrics for identifying high-risk youth who become more vulnerable to depression, versus those who remain resilient, will increase our understanding of how to optimize targets for novel prevention and early treatment approaches [116]. Returning to our cardiac analogy, we can consider the development of a neuroimaging-based mental health metric, equivalent to cardiovascular fitness, for quantifying risk-resilience status. VO2 max, or maximal oxygen consumption, refers to the maximum amount of oxygen that an individual can utilize during intense or maximal exercise and is often considered the best measure of cardiovascular fitness and aerobic endurance. Similarly, one could construct an analogous measure for brain fitness and mental endurance. Such an analogous metric might reflect brain network flexibility as measured by adaptive switching from states of resting connectivity to task-evoked engagement, and down- regulating or up- regulating specific connections and activations as required by a context or task.

Toward consensus guidelines

To facilitate translating fMRI circuit metrics to clinical use, we consider the need for consensus clinical guidelines. In cardiology, published research papers regularly cite the level of evidence obtained by each new finding [94]. This practice could be adopted in psychiatry, to make imaging findings suited to clinical guidelines. A recent systematic review has highlighted the opportunities and challenges for real-world implementation of precision psychiatry [117]. Here, we consider the roadmap for translating circuit tools into practice within the context of the Summary of Grading of Recommendations Assessment, Development and Evaluation (GRADE) criteria used by the American Psychiatric Association [118] (Table 2).

Table 2.

Summary of grading of recommendations assessment, development and evaluation (GRADE) criteria used by the American Psychiatric Association, applicable for fMRI and other measures used in precision psychiatry.

Factor and criteria*	How the factor influences the direction and strength of a recommendation
Problem This factor can be integrated with the balance of the benefits and harms and burden.	The problem is determined by the importance and frequency of the health care issue that is addressed (burden of disease, prevalence, or baseline risk). If the problem is of great importance, a strong recommendation is more likely.
Values and preferences	This describes how important health outcomes are to those affected, how variable they are and if there is uncertainty about this. The less variability or uncertainty there is about values and preferences for the critical or important outcomes, the more likely is a strong recommendation.
Quality of the evidence	The confidence in any estimate of the criteria determining the direction and strength of the recommendation will determine if a strong or conditional recommendation is offered. However, the overall quality that is assigned to the recommendation is that of the evidence about effects on population-important outcomes. The higher the quality of evidence the more likely is a strong recommendation.
Benefits and harms and burden	This requires an evaluation of the absolute effects of both the benefits and harms and their importance. The greater the net benefit or net harm, the more likely is a strong recommendation for or against the option.
Resource implications	This describes how resource- intense an option is if it is cost-effective and if there is incremental benefit. The more advantageous or clearly disadvantageous these resource implications are, the more likely is a strong recommendation.
Equity This factor is often addressed under values preferences, and frequently also includes resource considerations.	The greater the likelihood to reduce inequities or increase equity and the more accessible an option is, the more likely is a strong recommendation.
Acceptability This factor can be integrated with the balance of the benefits and harms and burden.	The greater the acceptability of an option to all or most stakeholders, the more likely is a strong recommendation.
Feasibility This factor includes considerations about values and preferences, and resource implications.	The greater the acceptability of an option to all or most stakeholders, the more likely is a strong recommendation.

A major GRADE component is the quality of available evidence. As summarized in Table 1, a number of existing studies meet criteria for high quality because of their well-controlled designs. Nonetheless, accelerating translation will require large scale clinical trials with matched reference datasets acquired with standard, clinically applicable sequences, across samples and sites. As highlighted above, we also need normative reference samples with sociodemographic representation, to facilitate the clinical interpretation of data similarly to neuropsychological testing [119, 120]. Regarding an evaluation of both benefits and harms, we are not aware of a formal evaluation of these factors for fMRI tools in depression treatment prediction. Although fMRI technology is used safely in routine practice, a formal evaluation might be needed.

Regarding first GRADE criterion, ‘the problem’, we believe that improving major depression outcomes is an important problem [4, 121, 122]. According to the CDC, suicide was the second leading cause of death among people aged 22–44 years in 2020 [5]. For two-thirds of individuals, depression is the primary cause of suicide [123]. This represents 15,951 young people aged 25–44 years who die by suicide each year, comparable to the 19,553 in this age band who die by heart disease [5, 124]. Functional neuroimaging tools are an opportunity to reduce these fatalities, similar to how imaging reduced deaths and morbidity due to heart disease [125].

Regarding the burden of illness criterion, for every employee experiencing depression, an average $15,000 per year is lost [90]. Depression costs the US economy $210.5 billion per year in absenteeism, reduced productivity, and medical cost [126]. As highlighted in our clinical illustration in Section 4, use of circuit predictors has the potential to improve remission outcomes, and thus reduce the cost of lost productivity and talent. Regarding values and preferences, major depression is also associated with substantial public stigma that is commonly internalized [127]. A survey of providers and patients suggests that brain scans could help alleviate the effects of stigma and self-blame [128]. We made similar observations in our Stanford discovery clinic for neuroscience-informed precision psychiatry. Spontaneous patient feedback indicates that “seeing their own brain” helps resolve stigma around their experience of major depression with potential benefits for patient engagement [129]. Although there are clearly other approaches to reducing stigma, these results suggest that a validated imaging technology has the potential to mitigate the social burden of depression.

GRADE criteria also encompass resource implications. Current models used for other prevalent chronic conditions, such as migraine, could inform considerations for future clinical use. Regarding equity, one key consideration is availability of MRI technology per capita. Figure 3 provides a summary of the availability of 3 T MRI scanners per 1,000,000 inhabitants available by country [130]. A related challenge will be the management of wait time. Although there are relatively few investigations of wait time, a Norwegian register study reported MRI wait times of 8–12 weeks [131] and in Canada a wait time information program reported 15 weeks [132]. These wait times may also impact the overall psychiatry service wait times which have been reported as 8 weeks or longer for an in-person visit [133].

Fig. 3. Availability of magnetic resonance imaging scanners.

Overview of the availability of 3 T magnetic resonance imaging (MRI) scanners that could be utilized for circuit assessment in depression and in precision psychiatry more broadly, by country. These values are drawn from OECD data [130] and the figure is reproduced in accordance with EOCD permission criteria.

The available evidence for acceptability of fMRI tools adds some weight to the consideration of benefits versus risks. Work done through a clinical neuroethics perspective suggests that there is high receptivity to brain scans for treatment tailoring in major depression [128]. Regarding the operational aspects of feasibility, we need HIPAA compliant fMRI tools that are integrated with the radiological picture archiving and communication system (PACS). There is also a need to equip busy practitioners with terminology and training that is suited to integration with current clinical workflows.

Conclusion

In this critical review we highlight how the growing body of fMRI evidence from the last two decades show that fMRI measurements can help select treatments to improve outcomes for major mental diseases. Given the enormous burden due to diseases such as depression, there is an urgent need for tools that help identify the most effective treatment for individuals more rapidly. To close the gap between discovery and delivery into practice, there is a need for pragmatic approaches and translational clinics that evaluate fMRI tools for implementation within clinical care settings.

Author contributions

LMW contributed to the conceptualization of the review and to the writing and editing. SWG contributed to the conceptualization of the review and to the writing and editing.

Funding

This work was supported by the National Institutes of Health [grant numbers R01MH101496 (LMW; NCT02220309), U01MH109985 (LMW)] and R61/R33 MH132072-01 (SWG).

Competing interests

LMW declares US Pants. App. 10/034,645 and 15/820,338: Systems and methods for detecting complex networks in MRI image data. SWG has nothing to declare.