A hierarchical Bayesian model to find brain-behaviour associations in incomplete data sets.

Canonical Correlation Analysis (CCA) and its regularised versions have been widely used in the neuroimaging community to uncover multivariate associations between two data modalities (e.g., brain imaging and behaviour). However, these methods have inherent limitations: (1) statistical inferences about the associations are often not robust; (2) the associations within each data modality are not modelled; (3) missing values need to be imputed or removed. Group Factor Analysis (GFA) is a hierarchical model that addresses the first two limitations by providing Bayesian inference and modelling modality-specific associations. Here, we propose an extension of GFA that handles missing data, and highlight that GFA can be used as a predictive model. We applied GFA to synthetic and real data consisting of brain connectivity and non-imaging measures from the Human Connectome Project (HCP). In synthetic data, GFA uncovered the underlying shared and specific factors and predicted correctly the non-observed data modalities in complete and incomplete data sets. In the HCP data, we identified four relevant shared factors, capturing associations between mood, alcohol and drug use, cognition, demographics and psychopathological measures and the default mode, frontoparietal control, dorsal and ventral networks and insula, as well as two factors describing associations within brain connectivity. In addition, GFA predicted a set of non-imaging measures from brain connectivity. These findings were consistent in complete and incomplete data sets, and replicated previous findings in the literature. GFA is a promising tool that can be used to uncover associations between and within multiple data modalities in benchmark datasets (such as, HCP), and easily extended to more complex models to solve more challenging tasks.

Will artificial intelligence eventually replace psychiatrists?

SUMMARY: The dystopian scenario of an 'artificial intelligence takeover' imagines artificial intelligence (AI) becoming the dominant form of intelligence on Earth, rendering humans redundant. As a society we have become increasingly familiar with AI and robots replacing humans in many tasks, certain jobs and even some areas of medicine, but surely this is not the fate of psychiatry?Here a computational neuroscientist (Janaina Mourão-Miranda) and psychiatrist (Justin Taylor Baker) suggest that psychiatry as a profession is relatively safe, whereas psychiatrists Christian Brown and Giles William Story predict that robots will be taking over the asylum.

Finding the needle in a high-dimensional haystack: Canonical correlation analysis for neuroscientists.

The 21st century marks the emergence of "big data" with a rapid increase in the availability of datasets with multiple measurements. In neuroscience, brain-imaging datasets are more commonly accompanied by dozens or hundreds of phenotypic subject descriptors on the behavioral, neural, and genomic level. The complexity of such "big data" repositories offer new opportunities and pose new challenges for systems neuroscience. Canonical correlation analysis (CCA) is a prototypical family of methods that is useful in identifying the links between variable sets from different modalities. Importantly, CCA is well suited to describing relationships across multiple sets of data, such as in recently available big biomedical datasets. Our primer discusses the rationale, promises, and pitfalls of CCA.

Combining heterogeneous data sources for neuroimaging based diagnosis: re-weighting and selecting what is important.

Combining neuroimaging and clinical information for diagnosis, as for example behavioral tasks and genetics characteristics, is potentially beneficial but presents challenges in terms of finding the best data representation for the different sources of information. Their simple combination usually does not provide an improvement if compared with using the best source alone. In this paper, we proposed a framework based on a recent multiple kernel learning algorithm called EasyMKL and we investigated the benefits of this approach for diagnosing two different mental health diseases. The well known Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset tackling the Alzheimer Disease (AD) patients versus healthy controls classification task, and a second dataset tackling the task of classifying an heterogeneous group of depressed patients versus healthy controls. We used EasyMKL to combine a huge amount of basic kernels alongside a feature selection methodology, pursuing an optimal and sparse solution to facilitate interpretability. Our results show that the proposed approach, called EasyMKLFS, outperforms baselines (e.g. SVM and SimpleMKL), state-of-the-art random forests (RF) and feature selection (FS) methods.

Predictive modelling using neuroimaging data in the presence of confounds.

When training predictive models from neuroimaging data, we typically have available non-imaging variables such as age and gender that affect the imaging data but which we may be uninterested in from a clinical perspective. Such variables are commonly referred to as 'confounds'. In this work, we firstly give a working definition for confound in the context of training predictive models from samples of neuroimaging data. We define a confound as a variable which affects the imaging data and has an association with the target variable in the sample that differs from that in the population-of-interest, i.e., the population over which we intend to apply the estimated predictive model. The focus of this paper is the scenario in which the confound and target variable are independent in the population-of-interest, but the training sample is biased due to a sample association between the target and confound. We then discuss standard approaches for dealing with confounds in predictive modelling such as image adjustment and including the confound as a predictor, before deriving and motivating an Instance Weighting scheme that attempts to account for confounds by focusing model training so that it is optimal for the population-of-interest. We evaluate the standard approaches and Instance Weighting in two regression problems with neuroimaging data in which we train models in the presence of confounding, and predict samples that are representative of the population-of-interest. For comparison, these models are also evaluated when there is no confounding present. In the first experiment we predict the MMSE score using structural MRI from the ADNI database with gender as the confound, while in the second we predict age using structural MRI from the IXI database with acquisition site as the confound. Considered over both datasets we find that none of the methods for dealing with confounding gives more accurate predictions than a baseline model which ignores confounding, although including the confound as a predictor gives models that are less accurate than the baseline model. We do find, however, that different methods appear to focus their predictions on specific subsets of the population-of-interest, and that predictive accuracy is greater when there is no confounding present. We conclude with a discussion comparing the advantages and disadvantages of each approach, and the implications of our evaluation for building predictive models that can be used in clinical practice.

Decoding negative affect personality trait from patterns of brain activation to threat stimuli.

INTRODUCTION: Pattern recognition analysis (PRA) applied to functional magnetic resonance imaging (fMRI) has been used to decode cognitive processes and identify possible biomarkers for mental illness. In the present study, we investigated whether the positive affect (PA) or negative affect (NA) personality traits could be decoded from patterns of brain activation in response to a human threat using a healthy sample. METHODS: fMRI data from 34 volunteers (15 women) were acquired during a simple motor task while the volunteers viewed a set of threat stimuli that were directed either toward them or away from them and matched neutral pictures. For each participant, contrast images from a General Linear Model (GLM) between the threat versus neutral stimuli defined the spatial patterns used as input to the regression model. We applied a multiple kernel learning (MKL) regression combining information from different brain regions hierarchically in a whole brain model to decode the NA and PA from patterns of brain activation in response to threat stimuli. RESULTS: The MKL model was able to decode NA but not PA from the contrast images between threat stimuli directed away versus neutral with a significance above chance. The correlation and the mean squared error (MSE) between predicted and actual NA were 0.52 (p-value=0.01) and 24.43 (p-value=0.01), respectively. The MKL pattern regression model identified a network with 37 regions that contributed to the predictions. Some of the regions were related to perception (e.g., occipital and temporal regions) while others were related to emotional evaluation (e.g., caudate and prefrontal regions). CONCLUSION: These results suggest that there was an interaction between the individuals' NA and the brain response to the threat stimuli directed away, which enabled the MKL model to decode NA from the brain patterns. To our knowledge, this is the first evidence that PRA can be used to decode a personality trait from patterns of brain activation during emotional contexts.

Multi-center MRI prediction models: Predicting sex and illness course in first episode psychosis patients.

Structural Magnetic Resonance Imaging (MRI) studies have attempted to use brain measures obtained at the first-episode of psychosis to predict subsequent outcome, with inconsistent results. Thus, there is a real need to validate the utility of brain measures in the prediction of outcome using large datasets, from independent samples, obtained with different protocols and from different MRI scanners. This study had three main aims: 1) to investigate whether structural MRI data from multiple centers can be combined to create a machine-learning model able to predict a strong biological variable like sex; 2) to replicate our previous finding that an MRI scan obtained at first episode significantly predicts subsequent illness course in other independent datasets; and finally, 3) to test whether these datasets can be combined to generate multicenter models with better accuracy in the prediction of illness course. The multi-center sample included brain structural MRI scans from 256 males and 133 females patients with first episode psychosis, acquired in five centers: University Medical Center Utrecht (The Netherlands) (n=67); Institute of Psychiatry, Psychology and Neuroscience, London (United Kingdom) (n=97); University of São Paulo (Brazil) (n=64); University of Cantabria, Santander (Spain) (n=107); and University of Melbourne (Australia) (n=54). All images were acquired on 1.5-Tesla scanners and all centers provided information on illness course during a follow-up period ranging 3 to 7years. We only included in the analyses of outcome prediction patients for whom illness course was categorized as either "continuous" (n=94) or "remitting" (n=118). Using structural brain scans from all centers, sex was predicted with significant accuracy (89%; p<0.001). In the single- or multi-center models, illness course could not be predicted with significant accuracy. However, when reducing heterogeneity by restricting the analyses to male patients only, classification accuracy improved in some samples. This study provides proof of concept that combining multi-center MRI data to create a well performing classification model is possible. However, to create complex multi-center models that perform accurately, each center should contribute a sample either large or homogeneous enough to first allow accurate classification within the single-center.

Sparse network-based models for patient classification using fMRI.

Pattern recognition applied to whole-brain neuroimaging data, such as functional Magnetic Resonance Imaging (fMRI), has proved successful at discriminating psychiatric patients from healthy participants. However, predictive patterns obtained from whole-brain voxel-based features are difficult to interpret in terms of the underlying neurobiology. Many psychiatric disorders, such as depression and schizophrenia, are thought to be brain connectivity disorders. Therefore, pattern recognition based on network models might provide deeper insights and potentially more powerful predictions than whole-brain voxel-based approaches. Here, we build a novel sparse network-based discriminative modeling framework, based on Gaussian graphical models and L1-norm regularized linear Support Vector Machines (SVM). In addition, the proposed framework is optimized in terms of both predictive power and reproducibility/stability of the patterns. Our approach aims to provide better pattern interpretation than voxel-based whole-brain approaches by yielding stable brain connectivity patterns that underlie discriminative changes in brain function between the groups. We illustrate our technique by classifying patients with major depressive disorder (MDD) and healthy participants, in two (event- and block-related) fMRI datasets acquired while participants performed a gender discrimination and emotional task, respectively, during the visualization of emotional valent faces.

Voxelwise Multivariate Analysis of Brain-Psychosocial Associations in Adolescents Reveals 6 Latent Dimensions of Cognition and Psychopathology.

BACKGROUND: Adolescence heralds the onset of considerable psychopathology, which may be conceptualized as an emergence of altered covariation between symptoms and brain measures. Multivariate methods can detect such modes of covariation or latent dimensions, but none specifically relating to psychopathology have yet been found using population-level structural brain data. Using voxelwise (instead of parcellated) brain data may strengthen latent dimensions' brain-psychosocial relationships, but this creates computational challenges. METHODS: We obtained voxelwise gray matter density and psychosocial variables from the baseline (ages 9-10 years) Adolescent Brain Cognitive Development (ABCD) Study cohort (N = 11,288) and employed a state-of-the-art segmentation method, sparse partial least squares, and a rigorous machine learning framework to prevent overfitting. RESULTS: We found 6 latent dimensions, 4 of which pertain specifically to mental health. The mental health dimensions were related to overeating, anorexia/internalizing, oppositional symptoms (all ps < .002) and attention-deficit/hyperactivity disorder symptoms (p = .03). Attention-deficit/hyperactivity disorder was related to increased and internalizing symptoms related to decreased gray matter density in dopaminergic and serotonergic midbrain areas, whereas oppositional symptoms were related to increased gray matter in a noradrenergic nucleus. Internalizing symptoms were related to increased and oppositional symptoms to reduced gray matter density in the insular, cingulate, and auditory cortices. Striatal regions featured strongly, with reduced caudate nucleus gray matter in attention-deficit/hyperactivity disorder and reduced putamen gray matter in oppositional/conduct problems. Voxelwise gray matter density generated stronger brain-psychosocial correlations than brain parcellations. CONCLUSIONS: Voxelwise brain data strengthen latent dimensions of brain-psychosocial covariation, and sparse multivariate methods increase their psychopathological specificity. Internalizing and externalizing symptoms are associated with opposite gray matter changes in similar cortical and subcortical areas.

A novel approach to probabilistic biomarker-based classification using functional near-infrared spectroscopy.

Pattern recognition approaches to the analysis of neuroimaging data have brought new applications such as the classification of patients and healthy controls within reach. In our view, the reliance on expensive neuroimaging techniques which are not well tolerated by many patient groups and the inability of most current biomarker algorithms to accommodate information about prior class frequencies (such as a disorder's prevalence in the general population) are key factors limiting practical application. To overcome both limitations, we propose a probabilistic pattern recognition approach based on cheap and easy-to-use multi-channel near-infrared spectroscopy (fNIRS) measurements. We show the validity of our method by applying it to data from healthy controls (n = 14) enabling differentiation between the conditions of a visual checkerboard task. Second, we show that high-accuracy single subject classification of patients with schizophrenia (n = 40) and healthy controls (n = 40) is possible based on temporal patterns of fNIRS data measured during a working memory task. For classification, we integrate spatial and temporal information at each channel to estimate overall classification accuracy. This yields an overall accuracy of 76% which is comparable to the highest ever achieved in biomarker-based classification of patients with schizophrenia. In summary, the proposed algorithm in combination with fNIRS measurements enables the analysis of sub-second, multivariate temporal patterns of BOLD responses and high-accuracy predictions based on low-cost, easy-to-use fNIRS patterns. In addition, our approach can easily compensate for variable class priors, which is highly advantageous in making predictions in a wide range of clinical neuroimaging applications.

Spectral density-based clustering algorithms for complex networks.

Introduction: Clustering is usually the first exploratory analysis step in empirical data. When the data set comprises graphs, the most common approaches focus on clustering its vertices. In this work, we are interested in grouping networks with similar connectivity structures together instead of grouping vertices of the graph. We could apply this approach to functional brain networks (FBNs) for identifying subgroups of people presenting similar functional connectivity, such as studying a mental disorder. The main problem is that real-world networks present natural fluctuations, which we should consider. Methods: In this context, spectral density is an exciting feature because graphs generated by different models present distinct spectral densities, thus presenting different connectivity structures. We introduce two clustering methods: k-means for graphs of the same size and gCEM, a model-based approach for graphs of different sizes. We evaluated their performance in toy models. Finally, we applied them to FBNs of monkeys under anesthesia and a dataset of chemical compounds. Results: We show that our methods work well in both toy models and real-world data. They present good results for clustering graphs presenting different connectivity structures even when they present the same number of edges, vertices, and degree of centrality. Discussion: We recommend using k-means-based clustering for graphs when graphs present the same number of vertices and the gCEM method when graphs present a different number of vertices.

Diagnostic neuroimaging across diseases.

Fully automated classification algorithms have been successfully applied to diagnose a wide range of neurological and psychiatric diseases. They are sufficiently robust to handle data from different scanners for many applications and in specific cases outperform radiologists. This article provides an overview of current applications taking structural imaging in Alzheimer's disease and schizophrenia as well as functional imaging to diagnose depression as examples. In this context, we also report studies aiming to predict the future course of the disease and the response to treatment for the individual. This has obvious clinical relevance but is also important for the design of treatment studies that may aim to include a cohort with a predicted fast disease progression to be more sensitive to detect treatment effects. In the second part, we present our own opinions on i) the role these classification methods can play in the clinical setting; ii) where their limitations are at the moment and iii) how those can be overcome. Specifically, we discuss strategies to deal with disease heterogeneity, diagnostic uncertainties, a probabilistic framework for classification and multi-class classification approaches.

Pattern recognition analyses of brain activation elicited by happy and neutral faces in unipolar and bipolar depression.

OBJECTIVES: Recently, pattern recognition approaches have been used to classify patterns of brain activity elicited by sensory or cognitive processes. In the clinical context, these approaches have been mainly applied to classify groups of individuals based on structural magnetic resonance imaging (MRI) data. Only a few studies have applied similar methods to functional MRI (fMRI) data. METHODS: We used a novel analytic framework to examine the extent to which unipolar and bipolar depressed individuals differed on discrimination between patterns of neural activity for happy and neutral faces. We used data from 18 currently depressed individuals with bipolar I disorder (BD) and 18 currently depressed individuals with recurrent unipolar depression (UD), matched on depression severity, age, and illness duration, and 18 age- and gender ratio-matched healthy comparison subjects (HC). fMRI data were analyzed using a general linear model and Gaussian process classifiers. RESULTS: The accuracy for discriminating between patterns of neural activity for happy versus neutral faces overall was lower in both patient groups relative to HC. The predictive probabilities for intense and mild happy faces were higher in HC than in BD, and for mild happy faces were higher in HC than UD (all p < 0.001). Interestingly, the predictive probability for intense happy faces was significantly higher in UD than BD (p = 0.03). CONCLUSIONS: These results indicate that patterns of whole-brain neural activity to intense happy faces were significantly less distinct from those for neutral faces in BD than in either HC or UD. These findings indicate that pattern recognition approaches can be used to identify abnormal brain activity patterns in patient populations and have promising clinical utility as techniques that can help to discriminate between patients with different psychiatric illnesses.

Canonical Correlation Analysis and Partial Least Squares for Identifying Brain-Behavior Associations: A Tutorial and a Comparative Study.

Canonical correlation analysis (CCA) and partial least squares (PLS) are powerful multivariate methods for capturing associations across 2 modalities of data (e.g., brain and behavior). However, when the sample size is similar to or smaller than the number of variables in the data, standard CCA and PLS models may overfit, i.e., find spurious associations that generalize poorly to new data. Dimensionality reduction and regularized extensions of CCA and PLS have been proposed to address this problem, yet most studies using these approaches have some limitations. This work gives a theoretical and practical introduction into the most common CCA/PLS models and their regularized variants. We examine the limitations of standard CCA and PLS when the sample size is similar to or smaller than the number of variables. We discuss how dimensionality reduction and regularization techniques address this problem and explain their main advantages and disadvantages. We highlight crucial aspects of the CCA/PLS analysis framework, including optimizing the hyperparameters of the model and testing the identified associations for statistical significance. We apply the described CCA/PLS models to simulated data and real data from the Human Connectome Project and Alzheimer's Disease Neuroimaging Initiative (both of n > 500). We use both low- and high-dimensionality versions of these data (i.e., ratios between sample size and variables in the range of ∼1-10 and ∼0.1-0.01, respectively) to demonstrate the impact of data dimensionality on the models. Finally, we summarize the key lessons of the tutorial.

A cross-cohort replicable and heritable latent dimension linking behaviour to multi-featured brain structure.

Identifying associations between interindividual variability in brain structure and behaviour requires large cohorts, multivariate methods, out-of-sample validation and, ideally, out-of-cohort replication. Moreover, the influence of nature vs nurture on brain-behaviour associations should be analysed. We analysed associations between brain structure (grey matter volume, cortical thickness, and surface area) and behaviour (spanning cognition, emotion, and alertness) using regularized canonical correlation analysis and a machine learning framework that tests the generalisability and stability of such associations. The replicability of brain-behaviour associations was assessed in two large, independent cohorts. The load of genetic factors on these associations was analysed with heritability and genetic correlation. We found one heritable and replicable latent dimension linking cognitive-control/executive-functions and positive affect to brain structural variability in areas typically associated with higher cognitive functions, and with areas typically associated with sensorimotor functions. These results revealed a major axis of interindividual behavioural variability linking to a whole-brain structural pattern.

Describing the brain in autism in five dimensions--magnetic resonance imaging-assisted diagnosis of autism spectrum disorder using a multiparameter classification approach.

Autism spectrum disorder (ASD) is a neurodevelopmental condition with multiple causes, comorbid conditions, and a wide range in the type and severity of symptoms expressed by different individuals. This makes the neuroanatomy of autism inherently difficult to describe. Here, we demonstrate how a multiparameter classification approach can be used to characterize the complex and subtle structural pattern of gray matter anatomy implicated in adults with ASD, and to reveal spatially distributed patterns of discriminating regions for a variety of parameters describing brain anatomy. A set of five morphological parameters including volumetric and geometric features at each spatial location on the cortical surface was used to discriminate between people with ASD and controls using a support vector machine (SVM) analytic approach, and to find a spatially distributed pattern of regions with maximal classification weights. On the basis of these patterns, SVM was able to identify individuals with ASD at a sensitivity and specificity of up to 90% and 80%, respectively. However, the ability of individual cortical features to discriminate between groups was highly variable, and the discriminating patterns of regions varied across parameters. The classification was specific to ASD rather than neurodevelopmental conditions in general (e.g., attention deficit hyperactivity disorder). Our results confirm the hypothesis that the neuroanatomy of autism is truly multidimensional, and affects multiple and most likely independent cortical features. The spatial patterns detected using SVM may help further exploration of the specific genetic and neuropathological underpinnings of ASD, and provide new insights into the most likely multifactorial etiology of the condition.

Utilizing temporal information in fMRI decoding: classifier using kernel regression methods.

This paper describes a general kernel regression approach to predict experimental conditions from activity patterns acquired with functional magnetic resonance image (fMRI). The standard approach is to use classifiers that predict conditions from activity patterns. Our approach involves training different regression machines for each experimental condition, so that a predicted temporal profile is computed for each condition. A decision function is then used to classify the responses from the testing volumes into the corresponding category, by comparing the predicted temporal profile elicited by each event, against a canonical hemodynamic response function. This approach utilizes the temporal information in the fMRI signal and maintains more training samples in order to improve the classification accuracy over an existing strategy. This paper also introduces efficient techniques of temporal compaction, which operate directly on kernel matrices for kernel classification algorithms such as the support vector machine (SVM). Temporal compacting can convert the kernel computed from each fMRI volume directly into the kernel computed from beta-maps, average of volumes or spatial-temporal kernel. The proposed method was applied to three different datasets. The first one is a block-design experiment with three conditions of image stimuli. The method outperformed the SVM classifiers of three different types of temporal compaction in single-subject leave-one-block-out cross-validation. Our method achieved 100% classification accuracy for six of the subjects and an average of 94% accuracy across all 16 subjects, exceeding the best SVM classification result, which was 83% accuracy (p=0.008). The second dataset is also a block-design experiment with two conditions of visual attention (left or right). Our method yielded 96% accuracy and SVM yielded 92% (p=0.005). The third dataset is from a fast event-related experiment with two categories of visual objects. Our method achieved 77% accuracy, compared with 72% using SVM (p=0.0006).

Patient classification as an outlier detection problem: an application of the One-Class Support Vector Machine.

Pattern recognition approaches, such as the Support Vector Machine (SVM), have been successfully used to classify groups of individuals based on their patterns of brain activity or structure. However these approaches focus on finding group differences and are not applicable to situations where one is interested in accessing deviations from a specific class or population. In the present work we propose an application of the one-class SVM (OC-SVM) to investigate if patterns of fMRI response to sad facial expressions in depressed patients would be classified as outliers in relation to patterns of healthy control subjects. We defined features based on whole brain voxels and anatomical regions. In both cases we found a significant correlation between the OC-SVM predictions and the patients' Hamilton Rating Scale for Depression (HRSD), i.e. the more depressed the patients were the more of an outlier they were. In addition the OC-SVM split the patient groups into two subgroups whose membership was associated with future response to treatment. When applied to region-based features the OC-SVM classified 52% of patients as outliers. However among the patients classified as outliers 70% did not respond to treatment and among those classified as non-outliers 89% responded to treatment. In addition 89% of the healthy controls were classified as non-outliers.

Quantitative prediction of subjective pain intensity from whole-brain fMRI data using Gaussian processes.

Supervised machine learning (ML) algorithms are increasingly popular tools for fMRI decoding due to their predictive capability and their ability to capture information encoded by spatially correlated voxels. In addition, an important secondary outcome is a multivariate representation of the pattern underlying the prediction. Despite an impressive array of applications, most fMRI applications are framed as classification problems and predictions are limited to categorical class decisions. For many applications, quantitative predictions are desirable that more accurately represent variability within subject groups and that can be correlated with behavioural variables. We evaluate the predictive capability of Gaussian process (GP) models for two types of quantitative prediction (multivariate regression and probabilistic classification) using whole-brain fMRI volumes. As a proof of concept, we apply GP models to an fMRI experiment investigating subjective responses to thermal pain and show GP models predict subjective pain ratings without requiring anatomical hypotheses about functional localisation of relevant brain processes. Even in the case of pain perception, where strong hypotheses do exist, GP predictions were more accurate than any region previously demonstrated to encode pain intensity. We demonstrate two brain mapping methods suitable for GP models and we show that GP regression models outperform state of the art support vector- and relevance vector regression. For classification, GP models perform categorical prediction as accurately as a support vector machine classifier and furnish probabilistic class predictions.