Data-driven approaches to neuroimaging biomarkers for neurological and psychiatric disorders: emerging approaches and examples

Abstract

Purpose:

The ‘holy grail’ of clinical applications of neuroimaging to neurological and psychiatric disorders via personalized biomarkers has remained mostly elusive, despite considerable effort. However, there are many reasons to continue to be hopeful, as the field has made remarkable advances over the past few years, fueled by a variety of converging technical and data developments.

Recent findings:

We discuss a number of advances that are accelerating the push for neuroimaging biomarkers including the advent of the ‘neuroscience big data’ era, biomarker data competitions, the development of more sophisticated algorithms including ‘guided’ data-driven approaches which facilitate automation of network-based analyses, dynamic connectivity, and deep learning. Another key advance includes multimodal data fusion approaches which can provide convergent and complementary evidence pointing to possible mechanisms as well as increase predictive accuracy.

Summary:

The search for clinically relevant neuroimaging biomarkers for neurological and psychiatric disorders is rapidly accelerating. Here we highlight some of these aspects, provide recent examples from studies in our group, and link to other ongoing work in the field. It is critical that access and use of these advanced approaches becomes mainstream, this will help propel the community forward and facilitate the production of robust and replicable neuroimaging biomarkers.

Introduction

The quest for neuroimaging biomarkers of brain disorders has been ongoing for several decades now. The advent of noninvasive neuroimaging of the human brain with magnetic resonance imaging (MRI) including functional MRI (fMRI), was a major leap forward, allowing us to visualize brain function and structure in humans completely noninvasively. And while MRI has proven highly useful clinically in many areas including cancer, stroke, and visualization of structural and vessel damage, the use of MRI to study various neurological and neuropsychiatric disorders is still largely a research question, due in part to the complexity of the brain itself, the disorders and a lack of precision in diagnostic categorizations¹. For example, most psychiatric disorders are still diagnosed on the basis of clinical symptoms and progression rather than on quantitative biological measures, as in the rest of medicine. Additionally, disorders such as schizophrenia can exhibit significant heterogeneity, making individual-level diagnosis from neuroimaging biomarkers especially challenging^2,3. Finally, promising results derived from one sample often replicate poorly in analyses of subsequent independent samples⁴. Because of these complexities, while we have learned much about how such disorders manifest in the human brain, this knowledge has not yet led to clinical translation into clinically useful brain MRI-based markers of brain disorders⁵, such as Alzheimer’s disease, mild traumatic brain injury, schizophrenia, and depression. However, despite this there is considerable reason to be optimistic about the future, and we contend that in large part this will be driven by advances in data-driven approaches⁶. The field has advanced rapidly in recent years on multiple fronts. Large datasets are much more available than ever before, analytic advances allow for visualization of more complex, yet precise, estimates of the impact of illness on the brain. In this paper we focus on the use of data-driven approaches to develop neuroimaging-based biomarkers. We highlight a few specific categories which have seen remarkable growth in recent years, including: 1) ‘automated’ data-driven approaches, 2) dynamic connectivity, 3) deep learning, and 4) multimodal data fusion predictive approaches. Finally, we discuss conclusions and future directions for the field.

There has been a large growth in the number of research articles focused on individualized classification (i.e., categorical⁷) and prediction (i.e., continuous/regression^8,9 based approaches) of brain disorders from neuroimaging data. While accuracies are still quite variable, particularly at an individual level, (Figure 1, left), it is also clear that 1) the number of studies continues to grow rapidly, 2) sample sizes are increasing, and 3) multimodal-imaging typically yields higher accuracies (Figure 1, right). In addition, the field is starting to accelerate various ‘virtuous cycles’ such as data-driven vs hypothesis-driven and individual effects vs group (canonical) effects.

Figure 1:

left: prediction accuracy of neuroimaging studies across a range of brain disorders (adapted from⁷), right: comparison of unimodal/multimodal and patient/control accuracies in the context of studies which predict various scores/symptoms (adapted from ⁸).

Competitions as bias control:

For anyone who works on biomarker studies, it becomes clear that it is depressingly easy to produce a solution which is biased, despite sometimes herculean attempts to avoid this via extensive cross-validation and replication. One way that can mitigate this problem is to host competitions in which the ground truth is completely hidden from the algorithm developer until winners are announced. An increasing number of such data competitions have emerged, focused on classification/prediction from neuroimaging data^10–16. Supplemental Figure 1 highlights a couple of such competitions, one from a machine learning community (top row), that hosted easily accessible structural and functional neuroimaging data on schizophrenia patients and controls, and received wide participation from non-brain imagers¹⁶. Interestingly, post-analysis showed that by combining multiple results together, performance can be further increased beyond the best single answer. The bottom row highlights the first ever UK Biobank (UKB) competition, which was a partnership between the Organization for Human Brain Mapping, the Institute for Electrical and Electronic Engineers (IEEE) Signal Processing Community and the UKB, and included over 11,000 fMRI and sMRI datasets¹⁰. Over 14,000 entries were received, further highlighting both the popularity and the potential of these approaches. Importantly, even in the well-controlled environment of a Kaggle competition, it can be challenging to replicate the resulting code, in the UKB competition for example, it took almost a month to validate the final solutions.

Automating data-driven approaches

One of the challenges of using data-driven approaches for biomarker development is that the results can be difficult to compare across individuals, and thus hard to automate. One way to address this is to use a hybrid or constrained approach. One such example is a fully automated independent component analysis (ICA) framework called NeuroMark¹⁷ which can be applied to resting fMRI data. The NeuroMark framework is based on the use of spatial networks derived from multiple large datasets and combines the power of ICA analysis with a fully automated approach that is robust to artifacts¹⁸. First, reproducible network templates are constructed from different groups of large-sample (N>800 each) healthy controls (HCs). Next, the network templates are used as spatial network priors for an ICA analysis¹⁹ to estimate subject-specific functional networks and associated time-courses (TCs). Finally, different functional network features such as the interaction between intrinsic connectivity networks (ICNs) are computed and evaluated.

This approach has a number of benefits in that it retains the benefits of data-driven strategies by adapting to individual subjects, and thus mitigating the pitfalls of fixed ROIs which may not correspond well to the underlying data (e.g., voxels are not coherent), and simplifies the data driven approach by providing an additional constraint that enables us to fully automate the approach. Finally, it provides a normalized measure, that is, correlation values. Figure 2 shows 53 reproducible network templates that are common between the Human Connectome Project (HCP)^20,21 and the brain genomics superstruct project (GSP)²² data, arranged into seven function domains according to their functional and anatomical features²³. These domains include the sub-cortical (SC: 5 ICNs), auditory (AU: 2 ICNs), sensorimotor (SM: 9 ICNs), visual (VI: 9 ICNs), cognitive control (CC: 17 ICNs), default mode (DM: 7 ICNs) and cerebellar (CB: 4 ICNs). In addition, we can compute the between-network connectivity, called functional network connectivity (FNC) that is the cross-correlation among component time courses (right of Figure 2). On the top right, we see that the connectivity patterns are highly consistent in different datasets, and even small group differences show good replicability. This type of approach is also more robust to artifacts in scanner data due to the spatial constraints and the use of ICA. The bottom right shows highly consistent differences between schizophrenia patients and controls.

Figure 2:

left: Spatial maps (SMs) of 53 identified intrinsic connectivity networks (ICNs), replicated across independent analysis of the HCP and GSP data, which are divided into seven different functional domains based on their anatomical and functional properties. Each color in the composite maps corresponds to a different ICN. These ICNs were used as network templates in our framework. Right: FNC matrices showing highly replicable main effects (top) and schizophrenia vs control differences (bottom). Adapted from ¹⁷.

Auto-labeling:

Another strategy is to use machine learning approaches to automatically label components in a fully data-driven framework. While this has been used previously for artifacts versus noise^24,25, it is also important to be able to identify and group the components of interest, as this impacts the modular structure of the FNC matrix. On the left of Supplemental Figure 2 we see an example of an ICA connectivity matrix (i.e., FNC) or correlation among ICA time courses for resting fMRI data which is unsorted²⁶. On the right of Supplemental Figure 2 the same FNC values are shown after automatically being sorted into ‘noise’ and intrinsic connectivity networks (top left block). This type of visualization is helpful as the expected modular structure is visible after sorting. In addition, by visualizing the intrinsic networks and artifacts together, we can also quickly see that there exists anticorrelation between the sensorimotor block (red) and some of the noise components. If we visualize these component maps it becomes clear these are predominantly white matter components. This, usually ignored, information can help inform research into white matter and/or the impact of global signal and other factors.

Example: Prediction of medication class response:

One of the challenges of mood disorders related to individuals who report to the clinician in the baseline euthymic state between illness episodes. In can be very challenging to diagnose and, if warranted, prescribe a medication class (e.g., mood stabilizers or antidepressants) that will work best for them. Some promising work has shown that using resting fMRI data it is possible to accurately predict medication class response to either mood stabilizers or antidepressants²⁷. The patients were followed up to 1–2 years after their initial scan to determine the medication class that worked best for them. A spatially constrained ICA approach was used to generate ICN, and classification of the medication class response was computed via an SVM analysis²⁷ performed on the individual subspace similarity of ICNs²⁸. Such an analysis enables us to assess the accuracy, positive and negative predictive values, sensitivity, and specificity of the classification algorithm and to determine which scan features contribute meaningfully to those predictions. Results showed accuracy at approximately 95%, including on patients that had initial ‘unknown’ diagnoses. The top five networks comprised the default mode, bilateral insula/auditory, fronto-cingulate, subcortical/thalamic, and bilateral frontoparietal networks (see Figure 3).

Figure 3: Prediction of medication class response to antidepressants or mood stabilizers.

An accuracy of 95% was achieved on a dataset of over 100 individuals following a rigorous assessment protocol which involved monitoring of patients over an extended period of time to obtain the medication class to which they were responsive (adapted from ²⁶).

Dynamic connectivity

Mood/psychosis classification:

Most functional connectivity work has focused on average resting fMRI connectivity, however more recently approaches have emphasized the importance of characterizing transient connectivity patterns (i.e., functional ‘states’)^29–31. A number of studies have suggested that dynamic functional connectivity metrics show improved sensitivity to brain disorders^32–34, especially in rest fMRI data which is expected to show considerably variability over time and is a mostly uncontrolled task³⁵. For example, a three-way classification of schizophrenia, bipolar disorder, and controls showed considerable improvement for dynamic measures (84% accuracy) vs static connectivity (59%)(Figure 4)³⁶. Approaches that allow for changes over time provide a more intuitively natural way to analyze functional connectivity, and studies have shown dynamic connectivity to be both highly replicable³⁷ and also more sensitive³⁸ than static connectivity approaches.

Figure 4: Three-way classification of schizophrenia, bipolar disorder, and controls.

A direct comparison of static and dynamic functional network connectivity approaches reveals significantly higher accuracy in the dynamic case for a three-way classification between schizophrenia, bipolar disorder, and control subjects using resting fMRI data (adapted from ³⁶).

Dynamic attractors predict traumatic brain injury recovery:

It is also possible to model functional brain states as well as information flow into and out of such states by using an approach that captures both the mean and the first temporal derivative³⁹. Using such an approach, it was observed that the resulting connectivity patterns could be grouped into pairs which exhibited similar average state values but opposite derivative values (Figure 5; top left). These patterns can be modeled as attractors which exhibit information flow in and out or the corresponding states (Figure 5; top left). Interestingly, these attractors were predictive of stages of recovery for traumatic brain injury⁴⁰ (Figure 5; bottom). We are still only scratching the surface of modeling the dynamics of connectivity and its potential for the development of robust neuroimaging biomarkers.

Figure 5: Attractor models.

(top left) by modeling the mean and the temporal derivative in estimating functional states we can identify patterns with similar magnitude but oppositely signed derivatives. These pairs can be modeled as dynamics attractors with information flowing in and out of the pairs (top right). This model can then be used to characterize stages of recovery from traumatic brain injury (bottom; adapted from ⁴⁰).

Deep learning approaches

Deep learning (DL) approaches applied to neuroimaging data have grown rapidly since their initial introduction to the field^41,42. Since then, there have been various discussions about what additional value such models are adding at the cost of increased complexity. While much is made of the black box nature of deep learning, we would argue that the models are actually “white box”, in that all the parameters are available and exposed. The challenge is in developing approaches to visualize the results. One of the strengths of DL is its flexibility, allowing us to capture unanticipated relationships in the data. In addition, by leveraging more of the raw data (e.g., the 3D information in the case of structural MRI data) rather than highlight distilled and engineered features, we can accrue additional benefits. Supplemental Figure 3 at the left is the cross-validated prediction accuracy of male vs female and age groups from over 17,000 UK Biobank structural MRI data sets. The prediction values are robust to scanner site, whether we include them or not and the DL approaches consistently outperform standard machine learning (SML) models such as support vector machines (SVM), logistic regression (LR), or linear discriminant analysis (LDA) models. A visualization of the results in a 2D space using the T-distributed stochastic nonlinear embedding (tSNE) approach⁴³ is show on the right of Supplemental Figure 3 (males blue, females red, age reflected as darker vs lighter shading). By rendering the weights within the model, we can visualize the relevant brain regions for each classification task as shown at the bottom of Supplemental Figure 3.

Deep Learning: AD and stable vs progressive MCI:

Deep learning for neuroimaging biomarker development is just getting started, but is rapidly growing. There have been a few studies using deep learning to predict Alzheimer’s disease (AD)^44–49. One of the more relevant applications for biomarkers is to predict disease progression⁵⁰. One recent study from our group used a deep residual network to predict Alzheimer’s disease versus controls, as well as progressive versus stable cognitive impairment⁵¹. Results showed high accuracy for prediction of Alzheimer’s disease (93%), but even more importantly, prediction of stable versus progressive mild cognitive impairment (MCI) achieved an accuracy of 81% (Figure 6). Using layer-wise relevance propagation⁵² to visualize the most relevant brain regions highlights medial temporal lobe as well as left occipital and parietal regions. The advancement of 3D+ DL models will undoubtedly have a large impact in advancing neuroimaging applications to biomarker development^53,54.

Figure 6: Deep learning prediction of cognitive impairment.

(top) correct identification of controls verses Alzheimer’s disease (93% accuracy) and stable versus progressive mild cognitive impairment (81%) with visualized brain regions shown on bottom row (adapted from ⁵¹).

Multimodal fusion

Multimodal data fusion can be placed on a spectrum that extends from separately analyzing two modalities in the context of a similar study, all the way to full symmetric analysis of multimodalities to extract the joint information which is often otherwise hidden⁵⁵. There are currently numerous approaches to multimodal fusion, and many are now beginning to be applied to systematically evaluate neuroimaging biomarkers of brain disorders^56–58. For example one approach is called multiset canonical correlation analysis with reference plus joint ICA (mCCAR+jICA). Although this sounds complicated, this procedure essentially jointly decomposes data into shared components that are also linked to a behavioral or symptom variable⁵⁹. As such it falls into the category of a hybrid approach, part data-driven and part hypothesis-driven.

Novelty seeking as a general multimodal predictor:

One of the benefits of the hybrid approach in this context is the ability to steer the results to the variance of interest, while still benefitting from data-driven methods. In one recent study we focused on novelty seeking (NS) which assesses preference for seeking novel experiences, and has been linked to sensitivity to reward environmental cues and has also been shown to be dysfunctional in multiple brain disorders. Based on the longitudinal adolescent-neuroimaging cohort, we identified multimodal biomarkers (based on task fMRI and structural MRI) associated with high-NS scores in adolescents at age 14. The identified brain regions were able to longitudinally predict five different symptom/risk scales including alcohol drinking, smoking, hyperactivity, depression, and psychosis for the same subjects as well as independent subjects at year 19, and also predict the corresponding symptom scores of five types of patient cohorts, including drinking, smoking, attention-deficit/hyperactivity disorder, major depressive disorders and schizophrenia, using resting fMRI and structural MRI data (see (Figure 7; top)⁶⁰. By leveraging the initial data to develop a predictive template, we can classify among attention deficit hyperactivity disorder (ADHD), major depressive disorder (MDD) and schizophrenia (SZ) with an accuracy of 87.2% (Figure 7; bottom right). The identified multimodal reward-circuit (Figure 7; bottom left) shows promise as a transdiagnostic neuroimaging biomarker to predict disease severity as well as to classify among hyperactivity, depression and schizophrenia. Collectively, by jointly analyzing four MRI features in a high-NS guided multimodal fusion model via supervised learning, the present study showed a contribution of personality (novelty seeking), neural response (activation in reward, emotion and inhibition) and structural (gray matter volume) covariation that capture a generalized multimodal reward circuit shared over alcohol drinking, smoking, ADHD, MDD and SZ, which may be differentially mediated by personality of NS and neural responses in a reward circuit.

Figure 7: Novelty-seeking related multimodal predictive network.

(top) Using task fMRI and structural MRI data to extract multimodal features linked to novelty seeking, we identify a set of regions that were predictive of novelty seeking in the same individuals 5 years later as well as predictive of symptoms of various brain disorders including substance use, ADHD, MDD, and SZ (adapted from ⁶⁰).

Conclusions and future directions

The combination of big data, algorithmic advances, and neuroinformatics solutions has propelled the neuroimaging field forward rapidly. Undoubtedly, employment of neuroimaging biomarkers will move from single data types to incorporating all relevant available data, in essence developing multimodal fingerprints or predictomes⁶¹ (Supplemental Figure 4). In addition, recent approaches have started to address sub-groups within neuropsychiatric syndromes^62,63, which can further improve prediction^64–66, improve validity and potentially help to redefine both diagnostic criteria^67–69 and to improve individual-level diagnosis^70,71. While brain-based classification has proven challenging, there has been considerable progress made in recent years and we look forward to seeing the full potential of the brain-based predictome realized¹.

Beyond resolving scientific and technical questions, adapting neuroimaging as a standardized clinical tool will require addressing such practical matters as developing a workflow, designing clinical reports, and applying for reimbursement. Another key question will be how to benefit from access to large amounts of data while protecting privacy. New data management infrastructure (i.e., neuroinformatics) to support the use of neuroimaging will need to incorporate strategies to safeguard sensitive information. Decentralized⁷² or federated approaches⁷³ may provide a way forward here, combining intermittent neuroimaging with the regular capture of information from a mobile device in a way that preserves privacy (http://coinstac.trendscenter.org/), that, for example, has considerable potential to treat disorders such as depression or cognitive decline by detecting mental health patterns.

In sum, while we have not advanced as quickly as we might have hoped in explicating the mysteries of neurological and mental disorders, there is considerable reason to be optimistic about the not-too-distant future. The brain is extremely complex, and this makes a focus on brain disorders even more challenging⁷⁴. The rapid progress of data availability, algorithms, and tools continue to advance the field. Additional approaches such as network lesioning studies^75–77 and brain stimulation^78,79 will also play a key role in helping establish the causal impact of particular brain networks or regions.

Supplementary Material

1

Supplemental Figure 1: left: prediction accuracy of neuroimaging studies across a range of brain disorders (adapted from ⁷), right: comparison of unimodal/multimodal and patient/control accuracies in the context of studies which predict various scores/symptoms (adapted from ⁸).

Supplemental Figure 2: Auto-labeling of ICA components. (left) unsorted FNC matrix, (right) FNC matrix after sorting into noise and intrinsic connectivity networks (ICNs) and also sorting ICNs based on existing atlases and machine learning. Results reveal the expected FNC modular structure (top left block) and also interesting relationships between the ‘noise’ components and the ICNs (adapted from ²⁶).

Supplemental Figure 3: Deep learning outperforms standard machine learning. (top left) performance of various standard machine learning (SML) approaches versus two deep learning (DL) approaches to classify five age groups and two gender groups (10 way). In all cases the DL approaches significantly outperform the SML approaches. A 2D tSNE visualization is show on the top right, the age groups show clear gradient from young to old and the gender groups are well separated. Visualization of the most relevant features in brain space for age and gender is show in the bottom panel (adapted from ⁴²).

Supplemental Figure 4: Multimodal predictome pipeline: An overview of a neuroimaging-based predictome pipeline. (a) Neuroimaging modalities, (b) Current approaches for feature selection. Feature extraction can include (i) voxel-based (ii) network-based, (iii) data-driven approaches (e.g., ICA, clustering), or (iv) jointly estimated features from multiple modalities (e.g., fMRI and genomics) and can include automatic or expert selection approaches. (c) Model validation can be performed using either a test-validation setup or using a k-fold cross validation scheme. (d) Data-driven sub-type identification can also be performed for homogeneous groups, subgroups, or continuous measures (adapted from ⁶¹).

Key summary points.

• The development of neuroimaging biomarkers for psychiatric and neurological disorders has been relatively underwhelming today, however despite this there are many reasons to believe the field is in a phase of rapid acceleration towards this goal.

• Some key aspects which are propelling the field are access to large data sets, development of sophisticated algorithms which combine accurate prediction and visualization approaches.

• Key areas of grown include network-based analyses, dynamic connectivity, and deep learning as well as multimodal data fusion.

• While a dizzying number of processing strategies have been proposed over the years, this is only the tip of the iceberg, we need to continue to balance a focus on groups, subgroups, and individual characteristics in order to facilitate the production of robust and replicable neuroimaging biomarkers.