Kernel regression for fMRI pattern prediction

Stacked regressions and structured variance partitioning for interpretable brain maps

Relating brain activity associated with a complex stimulus to different properties of that stimulus is a powerful approach for constructing functional brain maps. However, when stimuli are naturalistic, their properties are often correlated (e.g., visual and semantic features of natural images, or different layers of a convolutional neural network that are used as features of images). Correlated properties can act as confounders for each other and complicate the interpretability of brain maps, and can impact the robustness of statistical estimators. Here, we present an approach for brain mapping based on two proposed methods: stacking different encoding models and structured variance partitioning. Our stacking algorithm combines encoding models that each uses as input a feature space that describes a different stimulus attribute. The algorithm learns to predict the activity of a voxel as a linear combination of the outputs of different encoding models. We show that the resulting combined model can predict held-out brain activity better or at least as well as the individual encoding models. Further, the weights of the linear combination are readily interpretable; they show the importance of each feature space for predicting a voxel. We then build on our stacking models to introduce structured variance partitioning, a new type of variance partitioning that takes into account the known relationships between features. Our approach constrains the size of the hypothesis space and allows us to ask targeted questions about the similarity between feature spaces and brain regions even in the presence of correlations between the feature spaces. We validate our approach in simulation, showcase its brain mapping potential on fMRI data, and release a Python package. Our methods can be useful for researchers interested in aligning brain activity with different layers of a neural network, or with other types of correlated feature spaces.

Postoperative delirium prediction after cardiac surgery using machine learning models

OBJECTIVE

Postoperative delirium (POD) is a common postoperative complication in elderly patients, especially those undergoing cardiac surgery, which seriously affects the short- and long-term prognosis of patients. Early identification of risk factors for the development of POD can help improve the perioperative management of surgical patients. In the present study, five machine learning models were developed to predict patients at high risk of delirium after cardiac surgery and their performance was compared.

METHODS

A total of 367 patients who underwent cardiac surgery were retrospectively included in this study. Using single-factor analysis, 21 risk factors for POD were selected for inclusion in machine learning. The dataset was divided using 10-fold cross-validation for model training and testing. Five machine learning models (random forest (RF), support vector machine (SVM), radial based kernel neural network (RBFNN), K-nearest neighbour (KNN), and Kernel ridge regression (KRR)) were compared using area under the receiver operating characteristic curve (AUC-ROC), accuracy (ACC), sensitivity (SN), specificity (SPE), and Matthews coefficient (MCC).

RESULTS

Among 367 patients, 105 patients developed POD, the incidence of delirium was 28.6 %. Among the five ML models, RF had the best performance in ACC (87.99 %), SN (69.27 %), SPE (95.38 %), MCC (70.00 %) and AUC (0.9202), which was far superior to the other four models.

CONCLUSION

Delirium is common in patients after cardiac surgery. This analysis confirms the importance of the computational ML models in predicting the occurrence of delirium after cardiac surgery, especially the outstanding performance of the RF model, which has practical clinical applications for early identification of patients at risk of developing POD.

Predicting narcissistic personality traits from brain and psychological features: A supervised machine learning approach

Narcissism is a multifaceted construct often linked to pathological conditions whose neural correlates are still poorly understood. Previous studies have reported inconsistent findings related to the neural underpinnings of narcissism, probably due to methodological limitations such as the low number of participants or the use of mass univariate methods. The present study aimed to overcome the previous methodological limitations and to build a predictive model of narcissistic traits based on neural and psychological features. In this respect, two machine learning-based methods (Kernel Ridge Regression and Support Vector Regression) were used to predict narcissistic traits from brain structural organization and from other relevant normal and abnormal personality features. Results showed that a circuit including the lateral and middle frontal gyri, the angular gyrus, Rolandic operculum, and Heschl's gyrus successfully predicted narcissistic personality traits (p < 0.003). Moreover, narcissistic traits were predicted by normal (openness, agreeableness, conscientiousness) and abnormal (borderline, antisocial, insecure, addicted, negativistic, machiavellianism) personality traits. This study is the first to predict narcissistic personality traits via a supervised machine learning approach. As such, these results may expand the possibility of deriving personality traits from neural and psychological features.

Borderline shades: Morphometric features predict borderline personality traits but not histrionic traits

Borderline personality disorder (BPD) is one of the most diagnosed disorders in clinical settings. Besides the fully diagnosed disorder, borderline personality traits (BPT) are quite common in the general population. Prior studies have investigated the neural correlates of BPD but not of BPT. This paper investigates the neural correlates of BPT in a subclinical population using a supervised machine learning method known as Kernel Ridge Regression (KRR) to build predictive models. Additionally, we want to determine whether the same brain areas involved in BPD are also involved in subclinical BPT. Recent attempts to characterize the specific role of resting state-derived macro networks in BPD have highlighted the role of the default mode network. However, it is not known if this extends to the subclinical population. Finally, we wanted to test the hypothesis that the same circuitry that predicts BPT can also predict histrionic personality traits. Histrionic personality is sometimes considered a milder form of BPD, and making a differential diagnosis between the two may be difficult. For the first time KRR was applied to structural images of 135 individuals to predict BPT, based on the whole brain, on a circuit previously found to correctly classify BPD, and on the five macro-networks. At a whole brain level, results show that frontal and parietal regions, as well as the Heschl's area, the thalamus, the cingulum, and the insula, are able to predict borderline traits. BPT predictions increase when considering only the regions limited to the brain circuit derived from a study on BPD, confirming a certain overlap in brain structure between subclinical and clinical samples. Of all the five macro networks, only the DMN successfully predicts BPD, confirming previous observations on its role in the BPD. Histrionic traits could not be predicted by the BPT circuit. The results have implications for the diagnosis of BPD and a dimensional model of personality.

Stacked regressions and structured variance partitioning for interpretable brain maps

Relating brain activity associated with a complex stimulus to different properties of that stimulus is a powerful approach for constructing functional brain maps. However, when stimuli are naturalistic, their properties are often correlated (e.g., visual and semantic features of natural images, or different layers of a convolutional neural network that are used as features of images). Correlated properties can act as confounders for each other and complicate the interpretability of brain maps, and can impact the robustness of statistical estimators. Here, we present an approach for brain mapping based on two proposed methods: stacking different encoding models and structured variance partitioning. Our stacking algorithm combines encoding models that each use as input a feature space that describes a different stimulus attribute. The algorithm learns to predict the activity of a voxel as a linear combination of the outputs of different encoding models. We show that the resulting combined model can predict held-out brain activity better or at least as well as the individual encoding models. Further, the weights of the linear combination are readily interpretable; they show the importance of each feature space for predicting a voxel. We then build on our stacking models to introduce structured variance partitioning, a new type of variance partitioning that takes into account the known relationships between features. Our approach constrains the size of the hypothesis space and allows us to ask targeted questions about the similarity between feature spaces and brain regions even in the presence of correlations between the feature spaces. We validate our approach in simulation, showcase its brain mapping potential on fMRI data, and release a Python package. Our methods can be useful for researchers interested in aligning brain activity with different layers of a neural network, or with other types of correlated feature spaces.

Neuroanatomical predictors of real‐time fMRI ‐based anterior insula regulation. A supervised machine learning study

Increasing evidence showed that learned control of metabolic activity in selected brain regions can support emotion regulation. Notably, a number of studies demonstrated that neurofeedback-based regulation of fMRI activity in several emotion-related areas leads to modifications of emotional behavior along with changes of neural activity in local and distributed networks, in both healthy individuals and individuals with emotional disorders. However, the current understanding of the neural mechanisms underlying self-regulation of the emotional brain, as well as their relationship with other emotion regulation strategies, is still limited. In this study, we attempted to delineate neuroanatomical regions mediating real-time fMRI-based emotion regulation by exploring whole brain GM and WM features predictive of self-regulation of anterior insula (AI) activity, a neuromodulation procedure that can successfully support emotional brain regulation in healthy individuals and patients. To this aim, we employed a multivariate kernel ridge regression model to assess brain volumetric features, at regional and network level, predictive of real-time fMRI-based AI regulation. Our results showed that several GM regions including fronto-occipital and medial temporal areas and the basal ganglia as well as WM regions including the fronto-occipital fasciculus, tapetum and fornix significantly predicted learned AI regulation. Remarkably, we observed a substantial contribution of the cerebellum in relation to both the most effective regulation run and average neurofeedback performance. Overall, our findings highlighted specific neurostructural features contributing to individual differences of AI-guided emotion regulation. Notably, such neuroanatomical topography partially overlaps with the neurofunctional network associated with cognitive emotion regulation strategies, suggesting common neural mechanisms.

Predicting the severity of internet gaming disorder with resting-state brain features: A multi-voxel pattern analysis

BACKGROUND

Internet gaming disorder (IGD) has become a worldwide mental health concern; however, the neural mechanism underlying this disorder remains unclear. Multivoxel pattern analysis (MVPA), a newly developed data-driven approach, can be used to investigate the neural features of IGD based on massive neural data.

METHODS

Resting-state fMRI data from four hundred and two participants with varying levels of IGD severity were recruited. Regional homogeneity (ReHo) and the amplitude of low-frequency fluctuation (ALFF) were calculated and subsequently decoded by applying MVPA. The highly weighted regions in both predictive models were selected as regions of interest for further graph theory and Granger causality analysis (GCA) to explore how they affect IGD severity.

RESULTS

The results revealed that the neural patterns of ReHo and ALFF can independently and significantly predict IGD severity. The highly weighted regions that contributed to both predictive models were the right precentral gyrus and left postcentral gyrus. Moreover, topological properties of the right precentral gyrus were significantly correlated with IGD severity; further GCA revealed effective connectivity from the right precentral gyrus to left precentral gyrus and dorsal anterior cingulate cortex, both of which were significantly associated with IGD severity.

CONCLUSIONS

The present study demonstrated that IGD has distinctive neural patterns, and this pattern could be found by machine learning. In addition, the neural features in the right precentral gyrus play a key role in predicting IGD severity. The current study revealed the neural features of IGD and provided a potential target for IGD interventions using brain modulation.

On the encoding of natural music in computational models and human brains

This article discusses recent developments and advances in the neuroscience of music to understand the nature of musical emotion. In particular, it highlights how system identification techniques and computational models of music have advanced our understanding of how the human brain processes the textures and structures of music and how the processed information evokes emotions. Musical models relate physical properties of stimuli to internal representations called features, and predictive models relate features to neural or behavioral responses and test their predictions against independent unseen data. The new frameworks do not require orthogonalized stimuli in controlled experiments to establish reproducible knowledge, which has opened up a new wave of naturalistic neuroscience. The current review focuses on how this trend has transformed the domain of the neuroscience of music.

Forest yield prediction under different climate change scenarios using data intelligent models in Pakistan

This study aimed to develop and evaluate data driven models for prediction of forest yield under different climate change scenarios in the Gallies forest division of district Abbottabad, Pakistan. The Random Forest (RF) and Kernel Ridge Regression (KRR) models were developed and evaluated using yield data of two species (Blue pine and Silver fir) as an objective variable and climate data (temperature, humidity, rainfall and wind speed) as predictive variables. Prediction accuracy of both the models were assessed by means of root mean squared error (RMSE), mean absolute error (MAE), correlation coefficient (r), relative root mean squared error (RRMSE), Legates-McCabe's (LM), Willmott's index (WI) and Nash-Sutcliffe (NSE) metrics. Overall, the RF model outperformed the KRR model due to its higher accuracy in forecasting of forest yield. The study strongly recommends that RF model should be applied in other regions of the country for prediction of forest growth and yield, which may help in the management and future planning of forest productivity in Pakistan.

Neuroimaging-based Individualized Prediction of Cognition and Behavior for Mental Disorders and Health: Methods and Promises

The neuroimaging community has witnessed a paradigm shift in biomarker discovery from using traditional univariate brain mapping approaches to multivariate predictive models, allowing the field to move toward a translational neuroscience era. Regression-based multivariate models (hereafter "predictive modeling") provide a powerful and widely used approach to predict human behavior with neuroimaging features. These studies maintain a focus on decoding individual differences in a continuously behavioral phenotype from neuroimaging data, opening up an exciting opportunity to describe the human brain at the single-subject level. In this survey, we provide an overview of recent studies that utilize machine learning approaches to identify neuroimaging predictors over the past decade. We first review regression-based approaches and highlight connectome-based predictive modeling, which has grown in popularity in recent years. Next, we systematically describe recent representative studies using these tools in the context of cognitive function, symptom severity, personality traits, and emotion processing. Finally, we highlight a few challenges related to combining multimodal data, longitudinal prediction, external validations, and the employment of deep learning methods that have emerged from our review of the existing literature, as well as present some promising and challenging future directions.

Denouements of machine learning and multimodal diagnostic classification of Alzheimer's disease

Alzheimer's disease (AD) is the most common type of dementia. The exact cause and treatment of the disease are still unknown. Different neuroimaging modalities, such as magnetic resonance imaging (MRI), positron emission tomography, and single-photon emission computed tomography, have played a significant role in the study of AD. However, the effective diagnosis of AD, as well as mild cognitive impairment (MCI), has recently drawn large attention. Various technological advancements, such as robots, global positioning system technology, sensors, and machine learning (ML) algorithms, have helped improve the diagnostic process of AD. This study aimed to determine the influence of implementing different ML classifiers in MRI and analyze the use of support vector machines with various multimodal scans for classifying patients with AD/MCI and healthy controls. Conclusions have been drawn in terms of employing different classifier techniques and presenting the optimal multimodal paradigm for the classification of AD.

Towards Mapping Images to Text Using Deep-Learning Architectures

Images and text represent types of content that are used together for conveying a message. The process of mapping images to text can provide very useful information and can be included in many applications from the medical domain, applications for blind people, social networking, etc. In this paper, we investigate an approach for mapping images to text using a Kernel Ridge Regression model. We considered two types of features: simple RGB pixel-value features and image features extracted with deep-learning approaches. We investigated several neural network architectures for image feature extraction: VGG16, Inception V3, ResNet50, Xception. The experimental evaluation was performed on three data sets from different domains. The texts associated with images represent objective descriptions for two of the three data sets and subjective descriptions for the other data set. The experimental results show that the more complex deep-learning approaches that were used for feature extraction perform better than simple RGB pixel-value approaches. Moreover, the ResNet50 network architecture performs best in comparison to the other three deep network architectures considered for extracting image features. The model error obtained using the ResNet50 network is less by approx. 0.30 than other neural network architectures. We extracted natural language descriptors of images and we made a comparison between original and generated descriptive words. Furthermore, we investigated if there is a difference in performance between the type of text associated with the images: subjective or objective. The proposed model generated more similar descriptions to the original ones for the data set containing objective descriptions whose vocabulary is simpler, bigger and clearer.

Predicting differential diagnosis between bipolar and unipolar depression with multiple kernel learning on multimodal structural neuroimaging

One of the greatest challenges in providing early effective treatment in mood disorders is the early differential diagnosis between major depression (MDD) and bipolar disorder (BD). A remarkable need exists to identify reliable biomarkers for these disorders. We integrate structural neuroimaging techniques (i.e. Tract-based Spatial Statistics, TBSS, and Voxel-based morphometry) in a multiple kernel learning procedure in order to define a predictive function of BD against MDD diagnosis in a sample of 148 patients. We achieved a balanced accuracy of 73.65% with a sensitivity for BD of 74.32% and specificity for MDD of 72.97%. Mass-univariates analyses showed reduced grey matter volume in right hippocampus, amygdala, parahippocampal, fusiform gyrus, insula, rolandic and frontal operculum and cerebellum, in BD compared to MDD. Volumes in these regions and in anterior cingulate cortex were also reduced in BD compared to healthy controls (n = 74). TBSS analyses revealed widespread significant effects of diagnosis on fractional anisotropy, axial, radial, and mean diffusivity in several white matter tracts, suggesting disruption of white matter microstructure in depressed patients compared to healthy controls, with worse pattern for MDD. To best of our knowledge, this is the first study combining grey matter and diffusion tensor imaging in predicting BD and MDD diagnosis. Our results prompt brain quantitative biomarkers and multiple kernel learning as promising tool for personalized treatment in mood disorders.

Neural representations of transitive relations predict current and future math calculation skills in children

A large body of evidence suggests that math learning in children is built upon innate mechanisms for representing numerical quantities in the intraparietal sulcus (IPS). Learning math, however, is about more than processing quantitative information. It is also about understanding relations between quantities and making inferences based on these relations. Consistent with this idea, recent behavioral studies suggest that the ability to process transitive relations (A > B, B > C, therefore A > C) may contribute to math skills in children. Here we used fMRI coupled with a longitudinal design to determine whether the neural processing of transitive relations in children could predict their current and future math skills. At baseline (T₁), children (n = 31) processed transitive relations in an MRI scanner. Math skills were measured at T₁ and again 1.5 years later (T₂). Using a machine learning approach with cross-validation, we found that activity associated with the representation of transitive relations in the IPS predicted math calculation skills at both T₁ and T₂. Our study highlights the potential of neurobiological measures of transitive reasoning for forecasting math skills in children, providing additional evidence for a link between this type of reasoning and math learning.

Voxelwise encoding models with non-spherical multivariate normal priors

Predictive models for neural or fMRI data are often fit using regression methods that employ priors on the model parameters. One widely used method is ridge regression, which employs a spherical multivariate normal prior that assumes equal and independent variance for all parameters. However, a spherical prior is not always optimal or appropriate. There are many cases where expert knowledge or hypotheses about the structure of the model parameters could be used to construct a better prior. In these cases, non-spherical multivariate normal priors can be employed using a generalized form of ridge known as Tikhonov regression. Yet Tikhonov regression is only rarely used in neuroscience. In this paper we discuss the theoretical basis for Tikhonov regression, demonstrate a computationally efficient method for its application, and show several examples of how Tikhonov regression can improve predictive models for fMRI data. We also show that many earlier studies have implicitly used Tikhonov regression by linearly transforming the regressors before performing ridge regression.

Connectome-based model predicts individual differences in propensity to trust

Trust constitutes a fundamental basis of human society and plays a pivotal role in almost every aspect of human relationships. Although enormous interest exists in determining the neuropsychological underpinnings of a person's propensity to trust utilizing task-based fMRI; however, little progress has been made in predicting its variations by task-free fMRI based on whole-brain resting-state functional connectivity (RSFC). Here, we combined a one-shot trust game with a connectome-based predictive modeling approach to predict propensity to trust from whole-brain RSFC. We demonstrated that individual variations in the propensity to trust were primarily predicted by RSFC rooted in the functional integration of distributed key nodes-caudate, amygdala, lateral prefrontal cortex, temporal-parietal junction, and the temporal pole-which are part of domain-general large-scale networks essential for the motivational, affective, and cognitive aspects of trust. We showed, further, that the identified brain-behavior associations were only evident for trust but not altruistic preferences and that propensity to trust (and its underlying neural underpinnings) were modulated according to the extent to which a person emphasizes general social preferences (i.e., horizontal collectivism) rather than general risk preferences (i.e., trait impulsiveness). In conclusion, the employed data-driven approach enables to predict propensity to trust from RSFC and highlights its potential use as an objective neuromarker of trust impairment in mental disorders.

A comparison of various MRI feature types for characterizing whole brain anatomical differences using linear pattern recognition methods

There is a widespread interest in applying pattern recognition methods to anatomical neuroimaging data, but so far, there has been relatively little investigation into how best to derive image features in order to make the most accurate predictions. In this work, a Gaussian Process machine learning approach was used for predicting age, gender and body mass index (BMI) of subjects in the IXI dataset, as well as age, gender and diagnostic status using the ABIDE and COBRE datasets. MRI data were segmented and aligned using SPM12, and a variety of feature representations were derived from this preprocessing. We compared classification and regression accuracy using the different sorts of features, and with various degrees of spatial smoothing. Results suggested that feature sets that did not ignore the implicit background tissue class, tended to result in better overall performance, whereas some of the most commonly used feature sets performed relatively poorly.

Sparse Ordinal Logistic Regression and Its Application to Brain Decoding

Brain decoding with multivariate classification and regression has provided a powerful framework for characterizing information encoded in population neural activity. Classification and regression models are respectively used to predict discrete and continuous variables of interest. However, cognitive and behavioral parameters that we wish to decode are often ordinal variables whose values are discrete but ordered, such as subjective ratings. To date, there is no established method of predicting ordinal variables in brain decoding. In this study, we present a new algorithm, sparse ordinal logistic regression (SOLR), that combines ordinal logistic regression with Bayesian sparse weight estimation. We found that, in both simulation and analyses using real functional magnetic resonance imaging (fMRI) data, SOLR outperformed ordinal logistic regression with non-sparse regularization, indicating that sparseness leads to better decoding performance. SOLR also outperformed classification and linear regression models with the same type of sparseness, indicating the advantage of the modeling tailored to ordinal outputs. Our results suggest that SOLR provides a principled and effective method of decoding ordinal variables.

The effect of machine learning regression algorithms and sample size on individualized behavioral prediction with functional connectivity features

Individualized behavioral/cognitive prediction using machine learning (ML) regression approaches is becoming increasingly applied. The specific ML regression algorithm and sample size are two key factors that non-trivially influence prediction accuracies. However, the effects of the ML regression algorithm and sample size on individualized behavioral/cognitive prediction performance have not been comprehensively assessed. To address this issue, the present study included six commonly used ML regression algorithms: ordinary least squares (OLS) regression, least absolute shrinkage and selection operator (LASSO) regression, ridge regression, elastic-net regression, linear support vector regression (LSVR), and relevance vector regression (RVR), to perform specific behavioral/cognitive predictions based on different sample sizes. Specifically, the publicly available resting-state functional MRI (rs-fMRI) dataset from the Human Connectome Project (HCP) was used, and whole-brain resting-state functional connectivity (rsFC) or rsFC strength (rsFCS) were extracted as prediction features. Twenty-five sample sizes (ranged from 20 to 700) were studied by sub-sampling from the entire HCP cohort. The analyses showed that rsFC-based LASSO regression performed remarkably worse than the other algorithms, and rsFCS-based OLS regression performed markedly worse than the other algorithms. Regardless of the algorithm and feature type, both the prediction accuracy and its stability exponentially increased with increasing sample size. The specific patterns of the observed algorithm and sample size effects were well replicated in the prediction using re-testing fMRI data, data processed by different imaging preprocessing schemes, and different behavioral/cognitive scores, thus indicating excellent robustness/generalization of the effects. The current findings provide critical insight into how the selected ML regression algorithm and sample size influence individualized predictions of behavior/cognition and offer important guidance for choosing the ML regression algorithm or sample size in relevant investigations.

Kernel-Based Relevance Analysis with Enhanced Interpretability for Detection of Brain Activity Patterns

We introduce Enhanced Kernel-based Relevance Analysis (EKRA) that aims to support the automatic identification of brain activity patterns using electroencephalographic recordings. EKRA is a data-driven strategy that incorporates two kernel functions to take advantage of the available joint information, associating neural responses to a given stimulus condition. Regarding this, a Centered Kernel Alignment functional is adjusted to learning the linear projection that best discriminates the input feature set, optimizing the required free parameters automatically. Our approach is carried out in two scenarios: (i) feature selection by computing a relevance vector from extracted neural features to facilitating the physiological interpretation of a given brain activity task, and (ii) enhanced feature selection to perform an additional transformation of relevant features aiming to improve the overall identification accuracy. Accordingly, we provide an alternative feature relevance analysis strategy that allows improving the system performance while favoring the data interpretability. For the validation purpose, EKRA is tested in two well-known tasks of brain activity: motor imagery discrimination and epileptic seizure detection. The obtained results show that the EKRA approach estimates a relevant representation space extracted from the provided supervised information, emphasizing the salient input features. As a result, our proposal outperforms the state-of-the-art methods regarding brain activity discrimination accuracy with the benefit of enhanced physiological interpretation about the task at hand.

Support vector machine-based classification of neuroimages in Alzheimer's disease: direct comparison of FDG-PET, rCBF-SPECT and MRI data acquired from the same individuals

Objective:

To conduct the first support vector machine (SVM)-based study comparing the diagnostic accuracy of T1-weighted magnetic resonance imaging (T1-MRI), F-fluorodeoxyglucose-positron emission tomography (FDG-PET) and regional cerebral blood flow single-photon emission computed tomography (rCBF-SPECT) in Alzheimer’s disease (AD).

Method:

Brain T1-MRI, FDG-PET and rCBF-SPECT scans were acquired from a sample of mild AD patients (n=20) and healthy elderly controls (n=18). SVM-based diagnostic accuracy indices were calculated using whole-brain information and leave-one-out cross-validation.

Results:

The accuracy obtained using PET and SPECT data were similar. PET accuracy was 68∼71% and area under curve (AUC) 0.77∼0.81; SPECT accuracy was 68∼74% and AUC 0.75∼0.79, and both had better performance than analysis with T1-MRI data (accuracy of 58%, AUC 0.67). The addition of PET or SPECT to MRI produced higher accuracy indices (68∼74%; AUC: 0.74∼0.82) than T1-MRI alone, but these were not clearly superior to the isolated neurofunctional modalities.

Conclusion:

In line with previous evidence, FDG-PET and rCBF-SPECT more accurately identified patients with AD than T1-MRI, and the addition of either PET or SPECT to T1-MRI data yielded increased accuracy. The comparable SPECT and PET performances, directly demonstrated for the first time in the present study, support the view that rCBF-SPECT still has a role to play in AD diagnosis.

A review on neuroimaging-based classification studies and associated feature extraction methods for Alzheimer's disease and its prodromal stages

Neuroimaging has made it possible to measure pathological brain changes associated with Alzheimer's disease (AD) in vivo. Over the past decade, these measures have been increasingly integrated into imaging signatures of AD by means of classification frameworks, offering promising tools for individualized diagnosis and prognosis. We reviewed neuroimaging-based studies for AD and mild cognitive impairment classification, selected after online database searches in Google Scholar and PubMed (January, 1985-June, 2016). We categorized these studies based on the following neuroimaging modalities (and sub-categorized based on features extracted as a post-processing step from these modalities): i) structural magnetic resonance imaging [MRI] (tissue density, cortical surface, and hippocampal measurements), ii) functional MRI (functional coherence of different brain regions, and the strength of the functional connectivity), iii) diffusion tensor imaging (patterns along the white matter fibers), iv) fluorodeoxyglucose positron emission tomography (FDG-PET) (metabolic rate of cerebral glucose), and v) amyloid-PET (amyloid burden). The studies reviewed indicate that the classification frameworks formulated on the basis of these features show promise for individualized diagnosis and prediction of clinical progression. Finally, we provided a detailed account of AD classification challenges and addressed some future research directions.

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.

Brain responses to biological motion predict treatment outcome in young adults with autism receiving Virtual Reality Social Cognition Training: Preliminary findings

Autism Spectrum Disorder (ASD) is characterized by remarkable heterogeneity in social, communication, and behavioral deficits, creating a major barrier in identifying effective treatments for a given individual with ASD. To facilitate precision medicine in ASD, we utilized a well-validated biological motion neuroimaging task to identify pretreatment biomarkers that can accurately forecast the response to an evidence-based behavioral treatment, Virtual Reality-Social Cognition Training (VR-SCT). In a preliminary sample of 17 young adults with high-functioning ASD, we identified neural predictors of change in emotion recognition after VR-SCT. The predictors were characterized by the pretreatment brain activations to biological vs. scrambled motion in the neural circuits that support (a) language comprehension and interpretation of incongruent auditory emotions and prosody, and (b) processing socio-emotional experience and interpersonal affective information, as well as emotional regulation. The predictive value of the findings for individual adults with ASD was supported by regression-based multivariate pattern analyses with cross validation. To our knowledge, this is the first pilot study that shows neuroimaging-based predictive biomarkers for treatment effectiveness in adults with ASD. The findings have potentially far-reaching implications for developing more precise and effective treatments for ASD.

Distinct Frontoparietal Networks Underlying Attentional Effort and Cognitive Control

We investigated the brain activity patterns associated with stabilizing performance during challenges to attention. Our findings revealed distinct patterns of frontoparietal activity and functional connectivity associated with increased attentional effort versus preserved performance during challenged attention. Participants performed a visual signal detection task with and without presentation of a perceptual-attention challenge (changing background). The challenge condition increased activation in frontoparietal regions including right mid-dorsal/dorsolateral PFC (RPFC), approximating Brodmann's area 9, and superior parietal cortex. We found that greater behavioral impact of the challenge condition was correlated with greater RPFC activation, suggesting that increased engagement of cognitive control regions is not always sufficient to maintain high levels of performance. Functional connectivity between RPFC and ACC increased during the challenge condition and was also associated with performance declines, suggesting that the level of synchronized engagement of these regions reflects individual differences in attentional effort. Pretask, resting-state RPFC-ACC connectivity did not predict subsequent performance, suggesting that RPFC-ACC connectivity increased dynamically during task performance in response to performance decrement and error feedback. In contrast, functional connectivity between RPFC and superior parietal cortex not only during the task but also during pretask rest was associated with preserved performance in the challenge condition. Together, these data suggest that resting frontoparietal connectivity predicts performance on attention tasks that rely on those same cognitive control networks and that, under challenging conditions, other control regions dynamically couple with this network to initiate the engagement of cognitive control.

Brain responses to biological motion predict treatment outcome in young children with autism

Autism spectrum disorders (ASDs) are common yet complex neurodevelopmental disorders, characterized by social, communication and behavioral deficits. Behavioral interventions have shown favorable results-however, the promise of precision medicine in ASD is hampered by a lack of sensitive, objective neurobiological markers (neurobiomarkers) to identify subgroups of young children likely to respond to specific treatments. Such neurobiomarkers are essential because early childhood provides a sensitive window of opportunity for intervention, while unsuccessful intervention is costly to children, families and society. In young children with ASD, we show that functional magnetic resonance imaging-based stratification neurobiomarkers accurately predict responses to an evidence-based behavioral treatment-pivotal response treatment. Neural predictors were identified in the pretreatment levels of activity in response to biological vs scrambled motion in the neural circuits that support social information processing (superior temporal sulcus, fusiform gyrus, amygdala, inferior parietal cortex and superior parietal lobule) and social motivation/reward (orbitofrontal cortex, insula, putamen, pallidum and ventral striatum). The predictive value of our findings for individual children with ASD was supported by a multivariate pattern analysis with cross validation. Predicting who will respond to a particular treatment for ASD, we believe the current findings mark the very first evidence of prediction/stratification biomarkers in young children with ASD. The implications of the findings are far reaching and should greatly accelerate progress toward more precise and effective treatments for core deficits in ASD.

Can Emotional and Behavioral Dysregulation in Youth Be Decoded from Functional Neuroimaging?

INTRODUCTION

High comorbidity among pediatric disorders characterized by behavioral and emotional dysregulation poses problems for diagnosis and treatment, and suggests that these disorders may be better conceptualized as dimensions of abnormal behaviors. Furthermore, identifying neuroimaging biomarkers related to dimensional measures of behavior may provide targets to guide individualized treatment. We aimed to use functional neuroimaging and pattern regression techniques to determine whether patterns of brain activity could accurately decode individual-level severity on a dimensional scale measuring behavioural and emotional dysregulation at two different time points.

METHODS

A sample of fifty-seven youth (mean age: 14.5 years; 32 males) was selected from a multi-site study of youth with parent-reported behavioral and emotional dysregulation. Participants performed a block-design reward paradigm during functional Magnetic Resonance Imaging (fMRI). Pattern regression analyses consisted of Relevance Vector Regression (RVR) and two cross-validation strategies implemented in the Pattern Recognition for Neuroimaging toolbox (PRoNTo). Medication was treated as a binary confounding variable. Decoded and actual clinical scores were compared using Pearson's correlation coefficient (r) and mean squared error (MSE) to evaluate the models. Permutation test was applied to estimate significance levels.

RESULTS

Relevance Vector Regression identified patterns of neural activity associated with symptoms of behavioral and emotional dysregulation at the initial study screen and close to the fMRI scanning session. The correlation and the mean squared error between actual and decoded symptoms were significant at the initial study screen and close to the fMRI scanning session. However, after controlling for potential medication effects, results remained significant only for decoding symptoms at the initial study screen. Neural regions with the highest contribution to the pattern regression model included cerebellum, sensory-motor and fronto-limbic areas.

CONCLUSIONS

The combination of pattern regression models and neuroimaging can help to determine the severity of behavioral and emotional dysregulation in youth at different time points.

Evolving spatio-temporal data machines based on the NeuCube neuromorphic framework: Design methodology and selected applications

The paper describes a new type of evolving connectionist systems (ECOS) called evolving spatio-temporal data machines based on neuromorphic, brain-like information processing principles (eSTDM). These are multi-modular computer systems designed to deal with large and fast spatio/spectro temporal data using spiking neural networks (SNN) as major processing modules. ECOS and eSTDM in particular can learn incrementally from data streams, can include 'on the fly' new input variables, new output class labels or regression outputs, can continuously adapt their structure and functionality, can be visualised and interpreted for new knowledge discovery and for a better understanding of the data and the processes that generated it. eSTDM can be used for early event prediction due to the ability of the SNN to spike early, before whole input vectors (they were trained on) are presented. A framework for building eSTDM called NeuCube along with a design methodology for building eSTDM using this is presented. The implementation of this framework in MATLAB, Java, and PyNN (Python) is presented. The latter facilitates the use of neuromorphic hardware platforms to run the eSTDM. Selected examples are given of eSTDM for pattern recognition and early event prediction on EEG data, fMRI data, multisensory seismic data, ecological data, climate data, audio-visual data. Future directions are discussed, including extension of the NeuCube framework for building neurogenetic eSTDM and also new applications of eSTDM.

A Primer on Pattern-Based Approaches to fMRI: Principles, Pitfalls, and Perspectives

Human fMRI signals exhibit a spatial patterning that contains detailed information about a person's mental states. Using classifiers it is possible to access this information and study brain processes at the level of individual mental representations. The precise link between fMRI signals and neural population signals still needs to be unraveled. Also, the interpretation of classification studies needs to be handled with care. Nonetheless, pattern-based analyses make it possible to investigate human representational spaces in unprecedented ways, especially when combined with computational modeling.

Automatic supporting system for regionalization of ventricular tachycardia exit site in implantable defibrillators

Electrograms stored in Implantable Cardioverter Defibrillators (ICD-EGM) have been proven to convey useful information for roughly determining the anatomical location of the Left Ventricular Tachycardia exit site (LVTES). Our aim here was to evaluate the possibilities from a machine learning system intended to provide an estimation of the LVTES anatomical region with the use of ICD-EGM in the situation where 12-lead electrocardiogram of ventricular tachycardia are not available. Several machine learning techniques were specifically designed and benchmarked, both from classification (such as Neural Networks (NN), and Support Vector Machines (SVM)) and regression (Kernel Ridge Regression) problem statements. Classifiers were evaluated by using accuracy rates for LVTES identification in a controlled number of anatomical regions, and the regression approach quality was studied in terms of the spatial resolution. We analyzed the ICD-EGM of 23 patients (18±10 EGM per patient) during left ventricular pacing and simultaneous recording of the spatial coordinates of the pacing electrode with a navigation system. Several feature sets extracted from ICD-EGM (consisting of times and voltages) were shown to convey more discriminative information than the raw waveform. Among classifiers, the SVM performed slightly better than NN. In accordance with previous clinical works, the average spatial resolution for the LVTES was about 3 cm, as in our system, which allows it to support the faster determination of the LVTES in ablation procedures. The proposed approach also provides with a framework suitable for driving the design of improved performance future systems.

Investigating the use of support vector machine classification on structural brain images of preterm-born teenagers as a biological marker

Preterm birth has been shown to induce an altered developmental trajectory of brain structure and function. With the aid support vector machine (SVM) classification methods we aimed to investigate whether MRI data, collected in adolescence, could be used to predict whether an individual had been born preterm or at term. To this end we collected T1-weighted anatomical MRI data from 143 individuals (69 controls, mean age 14.6y). The inclusion criteria for those born preterm were birth weight ≤ 1500g and gestational age < 37w. A linear SVM was trained on the grey matter segment of MR images in two different ways. First, all the individuals were used for training and classification was performed by the leave-one-out method, yielding 93% correct classification (sensitivity = 0.905, specificity = 0.942). Separately, a random half of the available data were used for training twice and each time the other, unseen, half of the data was classified, resulting 86% and 91% accurate classifications. Both gestational age (R = -0.24, p<0.04) and birth weight (R = -0.51, p < 0.001) correlated with the distance to decision boundary within the group of individuals born preterm. Statistically significant correlations were also found between IQ (R = -0.30, p < 0.001) and the distance to decision boundary. Those born small for gestational age did not form a separate subgroup in these analyses. The high rate of correct classification by the SVM motivates further investigation. The long-term goal is to automatically and non-invasively predict the outcome of preterm-born individuals on an individual basis using as early a scan as possible.

Framingham Coronary Heart Disease Risk Score Can be Predicted from Structural Brain Images in Elderly Subjects

Recent literature has presented evidence that cardiovascular risk factors (CVRF) play an important role on cognitive performance in elderly individuals, both those who are asymptomatic and those who suffer from symptoms of neurodegenerative disorders. Findings from studies applying neuroimaging methods have increasingly reinforced such notion. Studies addressing the impact of CVRF on brain anatomy changes have gained increasing importance, as recent papers have reported gray matter loss predominantly in regions traditionally affected in Alzheimer's disease (AD) and vascular dementia in the presence of a high degree of cardiovascular risk. In the present paper, we explore the association between CVRF and brain changes using pattern recognition techniques applied to structural MRI and the Framingham score (a composite measure of cardiovascular risk largely used in epidemiological studies) in a sample of healthy elderly individuals. We aim to answer the following questions: is it possible to decode (i.e., to learn information regarding cardiovascular risk from structural brain images) enabling individual predictions? Among clinical measures comprising the Framingham score, are there particular risk factors that stand as more predictable from patterns of brain changes? Our main findings are threefold: (i) we verified that structural changes in spatially distributed patterns in the brain enable statistically significant prediction of Framingham scores. This result is still significant when controlling for the presence of the APOE 4 allele (an important genetic risk factor for both AD and cardiovascular disease). (ii) When considering each risk factor singly, we found different levels of correlation between real and predicted factors; however, single factors were not significantly predictable from brain images when considering APOE4 allele presence as covariate. (iii) We found important gender differences, and the possible causes of that finding are discussed.

Neural coding for effective rehabilitation

Successful neurological rehabilitation depends on accurate diagnosis, effective treatment, and quantitative evaluation. Neural coding, a technology for interpretation of functional and structural information of the nervous system, has contributed to the advancements in neuroimaging, brain-machine interface (BMI), and design of training devices for rehabilitation purposes. In this review, we summarized the latest breakthroughs in neuroimaging from microscale to macroscale levels with potential diagnostic applications for rehabilitation. We also reviewed the achievements in electrocorticography (ECoG) coding with both animal models and human beings for BMI design, electromyography (EMG) interpretation for interaction with external robotic systems, and robot-assisted quantitative evaluation on the progress of rehabilitation programs. Future rehabilitation would be more home-based, automatic, and self-served by patients. Further investigations and breakthroughs are mainly needed in aspects of improving the computational efficiency in neuroimaging and multichannel ECoG by selection of localized neuroinformatics, validation of the effectiveness in BMI guided rehabilitation programs, and simplification of the system operation in training devices.

Predicting dissolved oxygen concentration using kernel regression modeling approaches with nonlinear hydro-chemical data

Kernel function-based regression models were constructed and applied to a nonlinear hydro-chemical dataset pertaining to surface water for predicting the dissolved oxygen levels. Initial features were selected using nonlinear approach. Nonlinearity in the data was tested using BDS statistics, which revealed the data with nonlinear structure. Kernel ridge regression, kernel principal component regression, kernel partial least squares regression, and support vector regression models were developed using the Gaussian kernel function and their generalization and predictive abilities were compared in terms of several statistical parameters. Model parameters were optimized using the cross-validation procedure. The proposed kernel regression methods successfully captured the nonlinear features of the original data by transforming it to a high dimensional feature space using the kernel function. Performance of all the kernel-based modeling methods used here were comparable both in terms of predictive and generalization abilities. Values of the performance criteria parameters suggested for the adequacy of the constructed models to fit the nonlinear data and their good predictive capabilities.

PRoNTo: pattern recognition for neuroimaging toolbox

In the past years, mass univariate statistical analyses of neuroimaging data have been complemented by the use of multivariate pattern analyses, especially based on machine learning models. While these allow an increased sensitivity for the detection of spatially distributed effects compared to univariate techniques, they lack an established and accessible software framework. The goal of this work was to build a toolbox comprising all the necessary functionalities for multivariate analyses of neuroimaging data, based on machine learning models. The “Pattern Recognition for Neuroimaging Toolbox” (PRoNTo) is open-source, cross-platform, MATLAB-based and SPM compatible, therefore being suitable for both cognitive and clinical neuroscience research. In addition, it is designed to facilitate novel contributions from developers, aiming to improve the interaction between the neuroimaging and machine learning communities. Here, we introduce PRoNTo by presenting examples of possible research questions that can be addressed with the machine learning framework implemented in PRoNTo, and cannot be easily investigated with mass univariate statistical analysis.

Brain Imaging Analysis

The increasing availability of brain imaging technologies has led to intense neuroscientific inquiry into the human brain. Studies often investigate brain function related to emotion, cognition, language, memory, and numerous other externally induced stimuli as well as resting-state brain function. Studies also use brain imaging in an attempt to determine the functional or structural basis for psychiatric or neurological disorders and, with respect to brain function, to further examine the responses of these disorders to treatment. Neuroimaging is a highly interdisciplinary field, and statistics plays a critical role in establishing rigorous methods to extract information and to quantify evidence for formal inferences. Neuroimaging data present numerous challenges for statistical analysis, including the vast amounts of data collected from each individual and the complex temporal and spatial dependence present. We briefly provide background on various types of neuroimaging data and analysis objectives that are commonly targeted in the field. We present a survey of existing methods targeting these objectives and identify particular areas offering opportunities for future statistical contribution.

Fractal analysis of spontaneous fluctuations of the BOLD signal in the human brain networks

PURPOSE

To investigate what extent brain regions are continuously interacting during resting-state, independent component analyses (ICA) was applied to analyze resting-state functional MRI (RS-fMRI) data. According to the analyzed results, it was surprisingly found that low frequency fluctuations (LFFs), which belong to the 1/f signal (a signal with power spectrum whose power spectral density is inversely proportional to the frequency), have been classified into groups using ICA; furthermore, the spatial distributions of these groups within the brain were found to resemble the spatial distributions of different networks, which manifests that the signal characteristics of RS LFFs are distinct across networks. In our work, we applied the 1/f model in the fractal analyses to further investigate this distinction.

MATERIALS AND METHODS

Twenty healthy participants got involved in this study. They were scanned to acquire the RS-fMRI data. The acquired data were first processed with ICA to obtain the networks of the resting brain. Afterward, the blood-oxygenation level dependent (BOLD) signals of these networks were processed with the fractal analyses for obtaining the fractal parameter α.

RESULTS

α was found to significantly vary across networks, which reveals that the fractal characteristic of LFFs differs across networks. According to prior literatures, this difference could be brought by the discrepancy of hemodynamic response amplitude (HRA) between networks. Hence, in our work, we also performed the computational simulation to discover the relationship between α and HRA. Based on the simulation results, HRA is highly linear-correlated with the fractal characteristics of LFFs which is revealed by α.

CONCLUSION

Our results support that the origin of RS-fMRI signals contains arterial fluctuations. Hence, in addition to the commonly used method such as synchrony analysis and power spectral analysis, another approach, the fractal analysis, is suggested for acquiring the information of hemodynamic responses by means of RS-fMRI data.

Multivariate linear regression of high-dimensional fMRI data with multiple target variables

Multivariate regression is increasingly used to study the relation between fMRI spatial activation patterns and experimental stimuli or behavioral ratings. With linear models, informative brain locations are identified by mapping the model coefficients. This is a central aspect in neuroimaging, as it provides the sought-after link between the activity of neuronal populations and subject's perception, cognition or behavior. Here, we show that mapping of informative brain locations using multivariate linear regression (MLR) may lead to incorrect conclusions and interpretations. MLR algorithms for high dimensional data are designed to deal with targets (stimuli or behavioral ratings, in fMRI) separately, and the predictive map of a model integrates information deriving from both neural activity patterns and experimental design. Not accounting explicitly for the presence of other targets whose associated activity spatially overlaps with the one of interest may lead to predictive maps of troublesome interpretation. We propose a new model that can correctly identify the spatial patterns associated with a target while achieving good generalization. For each target, the training is based on an augmented dataset, which includes all remaining targets. The estimation on such datasets produces both maps and interaction coefficients, which are then used to generalize. The proposed formulation is independent of the regression algorithm employed. We validate this model on simulated fMRI data and on a publicly available dataset. Results indicate that our method achieves high spatial sensitivity and good generalization and that it helps disentangle specific neural effects from interaction with predictive maps associated with other targets.

Proof of concept study: relating infarct location to stroke disability in the NINDS rt-PA trial

BACKGROUND

The summed Alberta Stroke Program Early CT Score (ASPECTS) for noncontrast head CT scan represents the extent of early brain ischemia and has been shown to be useful for predicting stroke outcome. The ASPECTS template contains information on anatomical location which so far has not been used in analysis. This may not have been done because adjacent brain regions have related functions and share vascular territory. The task of relating neurological deficit to infarct localization requires brain imaging analysis tools which deal with this issue of relatedness or collinearity. We have previously used partial least squares with penalized logistic regression (PLR) to handle this problem of collinearity. A disadvantage of this method is that it cannot be performed at the bedside and requires processing and analysis in the imaging laboratory. PLR is a simpler analytic tool compared to partial least squares with PLR for dealing with this issue of relatedness (collinearity). It provides results in terms of β coefficients related to specific infarct locations in a manner that is intuitively understood by clinicians. In this exploratory analysis, we hypothesized that infarct location as represented by the individual ASPECTS region may be independently related to disability.

METHODS

ASPECTS from CT scans of patients in the National Institute of Neurological Disorders and Stroke (NINDS) recombinant tissue plasminogen activator (rt-PA) Study were obtained. Due to the collinearity between the ASPECTS regions, we used PLR to determine the independent associations of exposures (rt-PA), demographic variables (age and sex), and imaging (ASPECTS location) with poor outcome as defined by a modified Rankin Scale score of >2.

RESULTS

In 607/624 subjects with ASPECTS readings, variables significantly associated with poor outcome included: interactions between ASPECTS M6 region (primary motor cortex/parietal lobe) and age (p = 0.004), lentiform nucleus and age (p = 0.007), and blood sugar level and age (p = 0.01). The model suggested that older age or involvement of either M6 or lentiform nucleus slightly increased the odds of disability. However, the predominant effect was driven by rt-PA which reduced the odds of poor disability (OR 0.597, 95% CI 0.425-0.838, p = 0.003). This may potentially explain why certain patients have smaller gains from rt-PA treatments.

CONCLUSION

At an older age, specific infarct locations may be associated with a poorer outcome in this exploratory re-analysis of the NINDS rt-PA Study.

Predicting intrinsic brain activity

Multivariate supervised learning methods exhibit a remarkable ability to decode externally driven sensory, behavioral, and cognitive states from functional neuroimaging data. Although they are typically applied to task-based analyses, supervised learning methods are equally applicable to intrinsic effective and functional connectivity analyses. The obtained models of connectivity incorporate the multivariate interactions between all brain regions simultaneously, which will result in a more accurate representation of the connectome than the ones available with standard bivariate methods. Additionally the models can be applied to decode or predict the time series of intrinsic brain activity of a region from an independent dataset. The obtained prediction accuracy provides a measure of the integration between a brain region and other regions in its network, as well as a method for evaluating acquisition and preprocessing pipelines for resting state fMRI data. This article describes a method for learning multivariate models of connectivity. The method is applied in the non-parametric prediction accuracy, influence, and reproducibility-resampling (NPAIRS) framework, to study the regional variation of prediction accuracy and reproducibility (Strother et al., 2002). The resulting spatial distribution of these metrics is consistent with the functional hierarchy proposed by Mesulam (1998). Additionally we illustrate the utility of the multivariate regression connectivity modeling method for optimizing experimental parameters and assessing the quality of functional neuroimaging data.

Generative-discriminative basis learning for medical imaging

This paper presents a novel dimensionality reduction method for classification in medical imaging. The goal is to transform very high-dimensional input (typically, millions of voxels) to a low-dimensional representation (small number of constructed features) that preserves discriminative signal and is clinically interpretable. We formulate the task as a constrained optimization problem that combines generative and discriminative objectives and show how to extend it to the semi-supervised learning (SSL) setting. We propose a novel large-scale algorithm to solve the resulting optimization problem. In the fully supervised case, we demonstrate accuracy rates that are better than or comparable to state-of-the-art algorithms on several datasets while producing a representation of the group difference that is consistent with prior clinical reports. Effectiveness of the proposed algorithm for SSL is evaluated with both benchmark and medical imaging datasets. In the benchmark datasets, the results are better than or comparable to the state-of-the-art methods for SSL. For evaluation of the SSL setting in medical datasets, we use images of subjects with mild cognitive impairment (MCI), which is believed to be a precursor to Alzheimer's disease (AD), as unlabeled data. AD subjects and normal control (NC) subjects are used as labeled data, and we try to predict conversion from MCI to AD on follow-up. The semi-supervised extension of this method not only improves the generalization accuracy for the labeled data (AD/NC) slightly but is also able to predict subjects which are likely to converge to AD.

Does feature selection improve classification accuracy? Impact of sample size and feature selection on classification using anatomical magnetic resonance images

There are growing numbers of studies using machine learning approaches to characterize patterns of anatomical difference discernible from neuroimaging data. The high-dimensionality of image data often raises a concern that feature selection is needed to obtain optimal accuracy. Among previous studies, mostly using fixed sample sizes, some show greater predictive accuracies with feature selection, whereas others do not. In this study, we compared four common feature selection methods. 1) Pre-selected region of interests (ROIs) that are based on prior knowledge. 2) Univariate t-test filtering. 3) Recursive feature elimination (RFE), and 4) t-test filtering constrained by ROIs. The predictive accuracies achieved from different sample sizes, with and without feature selection, were compared statistically. To demonstrate the effect, we used grey matter segmented from the T1-weighted anatomical scans collected by the Alzheimer's disease Neuroimaging Initiative (ADNI) as the input features to a linear support vector machine classifier. The objective was to characterize the patterns of difference between Alzheimer's disease (AD) patients and cognitively normal subjects, and also to characterize the difference between mild cognitive impairment (MCI) patients and normal subjects. In addition, we also compared the classification accuracies between MCI patients who converted to AD and MCI patients who did not convert within the period of 12 months. Predictive accuracies from two data-driven feature selection methods (t-test filtering and RFE) were no better than those achieved using whole brain data. We showed that we could achieve the most accurate characterizations by using prior knowledge of where to expect neurodegeneration (hippocampus and parahippocampal gyrus). Therefore, feature selection does improve the classification accuracies, but it depends on the method adopted. In general, larger sample sizes yielded higher accuracies with less advantage obtained by using knowledge from the existing literature.

Prediction of illness severity in patients with major depression using structural MR brain scans

PURPOSE

To develop a model for the prediction of Major Depressive Disorder (MDD) illness severity ratings from individual structural MRI brain scans.

MATERIALS AND METHODS

Structural T1-weighted MRI scans were obtained from 30 patients with MDD recruited from two different scanning centers. Self-rated (Beck Depression Inventory; BDI), and clinician-rated (Hamilton Rating Scale for Depression, HRSD), syndrome-specific illness severity ratings were obtained just before scanning. Relevance vector regression (RVR) was used to predict the scores (BDI, HRSD) from T1-weighted MRI scans.

RESULTS

It was possible to predict the BDI score (correlation between actual score and RVR predicted scores r = 0.694; P < 0.0001), but not the HRSD scores (r = 0.34; P = 0.068) from individual subjects. BDI scores from the most ill patients were predicted more accurately than those from patients who were least ill (standard deviation of difference between predicted and actual scores 2.5 versus 7.4, respectively).

CONCLUSION

These data suggest that T1-weighted MRI scans contain sufficient information about neurobiological change in patients with MDD to permit accurate predictions about illness severity, on an individual subject basis, particularly for the most ill patients.

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).

Total variation regularization for fMRI-based prediction of behavior

While medical imaging typically provides massive amounts of data, the extraction of relevant information for predictive diagnosis remains a difficult challenge. Functional magnetic resonance imaging (fMRI) data, that provide an indirect measure of task-related or spontaneous neuronal activity, are classically analyzed in a mass-univariate procedure yielding statistical parametric maps. This analysis framework disregards some important principles of brain organization: population coding, distributed and overlapping representations. Multivariate pattern analysis, i.e., the prediction of behavioral variables from brain activation patterns better captures this structure. To cope with the high dimensionality of the data, the learning method has to be regularized. However, the spatial structure of the image is not taken into account in standard regularization methods, so that the extracted features are often hard to interpret. More informative and interpretable results can be obtained with the l(1) norm of the image gradient, also known as its total variation (TV), as regularization. We apply for the first time this method to fMRI data, and show that TV regularization is well suited to the purpose of brain mapping while being a powerful tool for brain decoding. Moreover, this article presents the first use of TV regularization for classification.

Generative embedding for model-based classification of fMRI data

Decoding models, such as those underlying multivariate classification algorithms, have been increasingly used to infer cognitive or clinical brain states from measures of brain activity obtained by functional magnetic resonance imaging (fMRI). The practicality of current classifiers, however, is restricted by two major challenges. First, due to the high data dimensionality and low sample size, algorithms struggle to separate informative from uninformative features, resulting in poor generalization performance. Second, popular discriminative methods such as support vector machines (SVMs) rarely afford mechanistic interpretability. In this paper, we address these issues by proposing a novel generative-embedding approach that incorporates neurobiologically interpretable generative models into discriminative classifiers. Our approach extends previous work on trial-by-trial classification for electrophysiological recordings to subject-by-subject classification for fMRI and offers two key advantages over conventional methods: it may provide more accurate predictions by exploiting discriminative information encoded in ‘hidden’ physiological quantities such as synaptic connection strengths; and it affords mechanistic interpretability of clinical classifications. Here, we introduce generative embedding for fMRI using a combination of dynamic causal models (DCMs) and SVMs. We propose a general procedure of DCM-based generative embedding for subject-wise classification, provide a concrete implementation, and suggest good-practice guidelines for unbiased application of generative embedding in the context of fMRI. We illustrate the utility of our approach by a clinical example in which we classify moderately aphasic patients and healthy controls using a DCM of thalamo-temporal regions during speech processing. Generative embedding achieves a near-perfect balanced classification accuracy of 98% and significantly outperforms conventional activation-based and correlation-based methods. This example demonstrates how disease states can be detected with very high accuracy and, at the same time, be interpreted mechanistically in terms of abnormalities in connectivity. We envisage that future applications of generative embedding may provide crucial advances in dissecting spectrum disorders into physiologically more well-defined subgroups.

Neurological and psychiatric spectrum disorders are typically defined in terms of particular symptom sets, despite increasing evidence that the same symptom may be caused by very different pathologies. Pathophysiological classification and effective treatment of such disorders will increasingly require a mechanistic understanding of inter-individual differences and clinical tools for making accurate diagnostic inference in individual patients. Previous classification studies have shown that functional magnetic resonance imaging (fMRI) can be used to differentiate between healthy controls and neurological or psychiatric patients. However, these studies are typically based on descriptive patterns and indirect measures of neural activity, and they rarely afford mechanistic insights into the underlying condition. In this paper, we address this challenge by proposing a classification approach that rests on a model of brain function and exploits the rich discriminative information encoded in directed interregional connection strengths. Based on an fMRI dataset acquired from moderately aphasic patients and healthy controls, we illustrate that our approach enables more accurate classification and deeper mechanistic insights about disease processes than conventional classification methods.

Multiclass Sparse Bayesian Regression for fMRI-Based Prediction

Inverse inference has recently become a popular approach for analyzing neuroimaging data, by quantifying the amount of information contained in brain images on perceptual, cognitive, and behavioral parameters. As it outlines brain regions that convey information for an accurate prediction of the parameter of interest, it allows to understand how the corresponding information is encoded in the brain. However, it relies on a prediction function that is plagued by the curse of dimensionality, as there are far more features (voxels) than samples (images), and dimension reduction is thus a mandatory step. We introduce in this paper a new model, called Multiclass Sparse Bayesian Regression (MCBR), that, unlike classical alternatives, automatically adapts the amount of regularization to the available data. MCBR consists in grouping features into several classes and then regularizing each class differently in order to apply an adaptive and efficient regularization. We detail these framework and validate our algorithm on simulated and real neuroimaging data sets, showing that it performs better than reference methods while yielding interpretable clusters of features.