Multivariate Machine Learning Techniques Research Articles

Limited generalizability of multivariate brain-based dimensions of child psychiatric symptoms

Multivariate machine learning techniques are a promising set of tools for identifying complex brain-behavior associations. However, failure to replicate results from these methods across samples has hampered their clinical relevance. Here we aimed to delineate dimensions of brain functional connectivity that are associated with child psychiatric symptoms in two large and independent cohorts: the Adolescent Brain Cognitive Development (ABCD) Study and the Generation R Study (total n = 6935). Using sparse canonical correlations analysis, we identified two brain-behavior dimensions in ABCD: attention problems and aggression/rule-breaking behaviors. Importantly, out-of-sample generalizability of these dimensions was consistently observed in ABCD, suggesting robust multivariate brain-behavior associations. Despite this, out-of-study generalizability in Generation R was limited. These results highlight that the degrees of generalizability can vary depending on the external validation methods employed as well as the datasets used, emphasizing that biomarkers will remain elusive until models generalize better in true external settings.

Using Probabilistic Machine Learning Methods to Improve Beef Cattle Price Modeling and Promote Beef Production Efficiency and Sustainability in Canada

Accurate agricultural commodity price models enable efficient allocation of limited natural resources, leading to improved sustainability in agriculture. Because of climate change, price volatility and uncertainty in the sector are expected to increase in the future, increasing the need for improved price modeling. With the emergence of machine learning (ML) algorithms, novel tools are now available to enhance the modeling of agricultural commodity prices. This research explores both univariate and multivariate ML techniques to perform probabilistic price prediction modeling for the Canadian beef industry, taking into account beef production, commodity markets, and international trade features to enhance accuracy. We model Alberta fed steer prices using three multivariate ML algorithms (support vector regression (SVR), random forest (RF), and Adaboost (AB)) and three univariate ML algorithms (autoregressive integrated moving average (ARIMA), seasonal ARIMA (SARIMA), and the seasonal autoregressive integrated moving average with exogenous factors (SARIMAX)). We apply these models to monthly fed steer price data between January 2005 and September 2023 and compare predicted prices with observed prices using several validation metrics. The outcomes indicate that both random forest (RF) and Adaboost (AB) show superior overall performance in accurately predicting Alberta fed steer prices in comparison to other algorithms. To better account for the variance of the best model performance, we subsequently adopted a probabilistic approach by considering uncertainty in our best-selected ML model. The beef industry can use these improved price models to minimize resource waste and inefficiency in the sector and improve the long-term sustainability prospects for beef producers in Canada.

Evaluation of Machine Learning Algorithms and Explainability Techniques to Detect Hearing Loss From a Speech-in-Noise Screening Test.

The aim of this study was to analyze the performance of multivariate machine learning (ML) models applied to a speech-in-noise hearing screening test and investigate the contribution of the measured features toward hearing loss detection using explainability techniques. Seven different ML techniques, including transparent (i.e., decision tree and logistic regression) and opaque (e.g., random forest) models, were trained and evaluated on a data set including 215 tested ears (99 with hearing loss of mild degree or higher and 116 with no hearing loss). Post hoc explainability techniques were applied to highlight the role of each feature in predicting hearing loss. Random forest (accuracy = .85, sensitivity = .86, specificity = .85, precision = .84) performed, on average, better than decision tree (accuracy = .82, sensitivity = .84, specificity = .80, precision = .79). Support vector machine, logistic regression, and gradient boosting had similar performance as random forest. According to post hoc explainability analysis on models generated using random forest, the features with the highest relevance in predicting hearing loss were age, number and percentage of correct responses, and average reaction time, whereas the total test time had the lowest relevance. This study demonstrates that a multivariate approach can help detect hearing loss with satisfactory performance. Further research on a bigger sample and using more complex ML algorithms and explainability techniques is needed to fully investigate the role of input features (including additional features such as risk factors and individual responses to low-/high-frequency stimuli) in predicting hearing loss.

Traffic speed prediction of high‐frequency time series using additively decomposed components as features

AbstractTraffic speed prediction is an integral part of an Intelligent Transportation System (ITS) and the Internet of Vehicles (IoV). Advanced knowledge of average traffic speed can help take proactive preventive steps to avoid impending problems. There have been studies for traffic speed prediction in which data has been decomposed into components using various decomposition techniques such as empirical mode decomposition, wavelets, and seasonal decomposition. As far as the authors are aware, no research has used additively decomposed components as input features. In this study, we used additive decomposition on 21,843 samples of traffic speed data. We implemented two statistical techniques designed for double seasonality (i) Double Seasonal Holt‐Winter, and (ii) Trigonometric seasonality, Box‐Cox transformation, autoregressive integrated moving average errors, trend, and Seasonal components (TBATS), and five machine learning (ML) techniques, (i) Multi‐Layer Perceptron, (ii) Convolutional‐Neural Network, (iii) Long Short‐Term Memory, (iv) Gated Recurrent Unit and (v) Convolutional‐Neural Network‐LSTM. Machine learning techniques are used in univariate mode with raw time series as features and then with decomposed components as features in multivariate mode. This study demonstrates that using decomposed components (trend, seasonal, and residual), as features, improves prediction results for multivariate ML techniques. This becomes a significant advantage when no other features are available.

Editorial: Neuroimaging: How Can Clues from the Brain Help Pain Management?

Editorial: Neuroimaging: How Can Clues from the Brain Help Pain Management?

Pattern of cerebellar grey matter loss associated with ataxia severity in spinocerebellar ataxias type 3: a multi-voxel pattern analysis.

Spinocerebellar ataxias type 3 (SCA3) patients are clinically characterized by progressive cerebellar ataxia combined with degeneration of the cerebellum. Previous neuroimaging studies have indicated ataxia severity associated with cerebellar atrophy using univariate methods. However, whether cerebellar atrophy patterns can be used to quantitatively predict ataxia severity in SCA3 patients at the individual level remains largely unexplored. In this study, a group of 66 SCA3 patients and 58 healthy controls were included. Disease duration and ataxia assessment, including theScale for the Assessment and Rating of Ataxia (SARA) and the International Cooperative Ataxia Rating Scale (ICARS), were collected for SCA3 patients. The high-resolution T1-weighted MRI was obtained, and cerebellar grey matter (GM) was extracted using a spatially unbiased infratentorial template toolbox for all participants. We investigated the association between the pattern of cerebellar grey matter (GM) loss and ataxia assessment in SCA3 by using a multivariate machine learning technique. We found that the application of RVR allowed quantitative prediction of both SARA scores (leave-one-subject-out cross-validation: correlation = 0.56, p-value = 0.001; mean squared error (MSE) = 20.51, p-value = 0.001; ten-fold cross-validation: correlation = 0.52, p-value = 0.001; MSE = 21.00, p-value = 0.001) and ICARS score (leave-one-subject-out cross-validation: correlation = 0.59, p-value = 0.001; MSE = 139.69, p-value = 0.001; ten-fold cross-validation: correlation = 0.57, p-value = 0.001; MSE = 145.371, p-value = 0.001) with statistically significant accuracy. These results provide proof-of-concept that ataxia severity in SCA3 patients can be predicted by the alteration pattern of cerebellar GM using multi-voxel pattern analysis.

Probing the clinical and brain structural boundaries of bipolar and major depressive disorder

Bipolar disorder (BD) and major depressive disorder (MDD) have both common and distinct clinical features, that pose both conceptual challenges in terms of their diagnostic boundaries and practical difficulties in optimizing treatment. Multivariate machine learning techniques offer new avenues for exploring these boundaries based on clinical neuroanatomical features. Brain structural data were obtained at 3 T from a sample of 90 patients with BD, 189 patients with MDD, and 162 healthy individuals. We applied sparse partial least squares discriminant analysis (s-PLS-DA) to identify clinical and brain structural features that may discriminate between the two clinical groups, and heterogeneity through discriminative analysis (HYDRA) to detect patient subgroups with reference to healthy individuals. Two clinical dimensions differentiated BD from MDD (area under the curve: 0.76, P < 0.001); one dimension emphasized disease severity as well as irritability, agitation, anxiety and flight of ideas and the other emphasized mostly elevated mood. Brain structural features could not distinguish between the two disorders. HYDRA classified patients in two clusters that differed in global and regional cortical thickness, the distribution proportion of BD and MDD and positive family history of psychiatric disorders. Clinical features remain the most reliable discriminant attributed of BD and MDD depression. The brain structural findings suggests that biological partitions of patients with mood disorders are likely to lead to the identification of subgroups, that transcend current diagnostic divisions into BD and MDD and are more likely to be aligned with underlying genetic variation. These results set the foundation for future studies to enhance our understanding of brain–behavior relationships in mood disorders.

Identification of specific neural circuit underlying the key cognitive deficit of remitted late-onset depression: A multi-modal MRI and machine learning study

Neuropsychological impairment is a key feature of late-onset depression (LOD), with deficits observed across multiple cognitive domains. And this neuropsychological impairment can persist even after the remission of depressive symptoms. However, none of previous studies have explored the pattern of cognitive deficit in remitted LOD (rLOD), and investigated the specific neural circuit underlying the key cognitive deficit of LOD. 40 rLOD patients and 36 controls underwent comprehensive neuropsychological assessments and magnetic resonance imaging (MRI) scans. The influence of executive function or information processing speed deficit on other cognitive domains was first investigated. We then applied a multivariate machine learning technique known as relevance vector regression to evaluate the potential of multiple-modal MRI (i.e., integrating whole-brain grey-matter [GM] volume and white-matter [WM] tract features) for making accurate predictions about the key cognitive deficit for individual rLOD patient. We revealed that the information processing speed appears to represent a key cognitive deficit in rLOD. Further the machine learning model identified a wide range of GM regions and WM tracts that significantly contributed to the prediction of individual performance on information processing speed (r = 0.50, P < 0.001). The GM regions mainly located in the frontal-subcortical and limbic systems; and the WM tracts mainly located in the frontal-limbic pathway, including the anterior corona radiata, fornix, posterior cingulate bundle, and uncinate fasciculus. This present study provide strongly evidence supporting the concept of rLOD that the core aspect of the cognitive deficits (i.e., information processing speed) is associated with disruption of the frontal-subcortical-limbic pathway.

Evaluating resting-state BOLD variability in relation to biomarkers of preclinical Alzheimer's disease

Recent functional magnetic resonance imaging studies have demonstrated that moment-to-moment variability in the blood oxygen level–dependent (BOLD) signal is related to age differences, cognition, and symptomatic Alzheimer's disease (AD). However, no studies have examined BOLD variability in the context of preclinical AD. We tested relationships between resting-state BOLD variability and biomarkers of amyloidosis, tauopathy, and neurodegeneration in a large (N = 321), well-characterized sample of cognitively normal adults (age = 39–93), using multivariate machine learning techniques. Furthermore, we controlled for cardiovascular health factors, which may contaminate resting-state BOLD variability estimates. BOLD variability, particularly in the default mode network, was related to cerebrospinal fluid (CSF) amyloid-β42 but was not related to CSF phosphorylated tau-181. Furthermore, BOLD variability estimates were also related to markers of neurodegeneration, including CSF neurofilament light protein, hippocampal volume, and a cortical thickness composite. Notably, relationships with hippocampal volume and cortical thickness survived correction for cardiovascular health and also contributed to age-related differences in BOLD variability. Thus, BOLD variability may be sensitive to preclinical pathology, including amyloidosis and neurodegeneration in AD-sensitive areas.

Converging intracortical signatures of two separated processing timescales in human early auditory cortex

Neural oscillations in auditory cortex are argued to support parsing and representing speech constituents at their corresponding temporal scales. Yet, how incoming sensory information interacts with ongoing spontaneous brain activity, what features of the neuronal microcircuitry underlie spontaneous and stimulus-evoked spectral fingerprints, and what these fingerprints entail for stimulus encoding, remain largely open questions. We used a combination of human invasive electrophysiology, computational modeling and decoding techniques to assess the information encoding properties of brain activity and to relate them to a plausible underlying neuronal microarchitecture. We analyzed intracortical auditory EEG activity from 10 patients while they were listening to short sentences. Pre-stimulus neural activity in early auditory cortical regions often exhibited power spectra with a shoulder in the delta range and a small bump in the beta range. Speech decreased power in the beta range, and increased power in the delta-theta and gamma ranges. Using multivariate machine learning techniques, we assessed the spectral profile of information content for two aspects of speech processing: detection and discrimination. We obtained better phase than power information decoding, and a bimodal spectral profile of information content with better decoding at low (delta-theta) and high (gamma) frequencies than at intermediate (beta) frequencies. These experimental data were reproduced by a simple rate model made of two subnetworks with different timescales, each composed of coupled excitatory and inhibitory units, and connected via a negative feedback loop. Modeling and experimental results were similar in terms of pre-stimulus spectral profile (except for the iEEG beta bump), spectral modulations with speech, and spectral profile of information content. Altogether, we provide converging evidence from both univariate spectral analysis and decoding approaches for a dual timescale processing infrastructure in human auditory cortex, and show that it is consistent with the dynamics of a simple rate model.

Quantitative prediction of individual cognitive flexibility using structural MRI.

Cognitive flexibility, a core dimension of executive functions, refers to one's ability to switch between multiple tasks and sets in a quick and flexible manner. However, whether objective neuroimaging can be used to quantitatively predict cognitive flexibility at the individual level remains largely unexplored. High-resolution magnetic resonance imaging data of 100 healthy young participants from the Human Connectome Project (HCP) dataset were used to calculate gray matter volume (GMV). Cognitive flexibility was assessed by the Dimensional Change Card Sort Test (DCCS). Using a multivariate machine learning technique known as relevance vector regression (RVR), we examined the relationship between GMV and cognitive flexibility performance. We found that the application of RVR to GMV allowed quantitative prediction of the DCCS scores with statistically significant accuracy (correlation = 0.41, P = 0.0001; mean squared error = 73.35, P = 0.0001). Accurate prediction was mainly based on GMV in the temporal regions. In addition, a univariate approach also revealed an inverse association between DCCS scores and GMV in the temporal areas. Our findings provide preliminary support to the development of neuroimaging techniques as a useful means to inform the cognitive assessment of individuals. Furthermore, the significant contribution of temporal regions suggests the prominent role of temporal cortex morphology in individual differences in cognitive flexibility.

Sleep in patients with disorders of consciousness characterized by means of machine learning.

Sleep has been proposed to indicate preserved residual brain functioning in patients suffering from disorders of consciousness (DOC) after awakening from coma. However, a reliable characterization of sleep patterns in this clinical population continues to be challenging given severely altered brain oscillations, frequent and extended artifacts in clinical recordings and the absence of established staging criteria. In the present study, we try to address these issues and investigate the usefulness of a multivariate machine learning technique based on permutation entropy, a complexity measure. Specifically, we used long-term polysomnography (PSG), along with video recordings in day and night periods in a sample of 23 DOC; 12 patients were diagnosed as Unresponsive Wakefulness Syndrome (UWS) and 11 were diagnosed as Minimally Conscious State (MCS). Eight hour PSG recordings of healthy sleepers (N = 26) were additionally used for training and setting parameters of supervised and unsupervised model, respectively. In DOC, the supervised classification (wake, N1, N2, N3 or REM) was validated using simultaneous videos which identified periods with prolonged eye opening or eye closure.The supervised classification revealed that out of the 23 subjects, 11 patients (5 MCS and 6 UWS) yielded highly accurate classification with an average F1-score of 0.87 representing high overlap between the classifier predicting sleep (i.e. one of the 4 sleep stages) and closed eyes. Furthermore, the unsupervised approach revealed a more complex pattern of sleep-wake stages during the night period in the MCS group, as evidenced by the presence of several distinct clusters. In contrast, in UWS patients no such clustering was found. Altogether, we present a novel data-driven method, based on machine learning that can be used to gain new and unambiguous insights into sleep organization and residual brain functioning of patients with DOC.

Development and validation of a brain maturation index using longitudinal neuroanatomical scans

BackgroundMajor psychiatric disorders are increasingly being conceptualized as ‘neurodevelopmental’, because they are associated with aberrant brain maturation. Several studies have hypothesized that a brain maturation index integrating patterns of neuroanatomical measurements may reliably identify individual subjects deviating from a normative neurodevelopmental trajectory. However, while recent studies have shown great promise in developing accurate brain maturation indices using neuroimaging data and multivariate machine learning techniques, this approach has not been validated using a large sample of longitudinal data from children and adolescents. MethodsT1-weighted scans from 303 healthy subjects aged 4.88 to 18.35years were acquired from the National Institute of Health (NIH) pediatric repository (http://www.pediatricmri.nih.gov). Out of the 303 subjects, 115 subjects were re-scanned after 2years. The least absolute shrinkage and selection operator algorithm (LASSO) was ‘trained’ to integrate neuroanatomical changes across chronological age and predict each individual's brain maturity. The resulting brain maturation index was developed using first-visit scans only, and was validated using second-visit scans. ResultsWe report a high correlation between the first-visit chronological age and brain maturation index (r=0.82, mean absolute error or MAE=1.69years), and a high correlation between the second-visit chronological age and brain maturation index (r=0.83, MAE=1.71years). The brain maturation index captured neuroanatomical volume changes between the first and second visits with an MAE of 0.27years. ConclusionsThe brain maturation index developed in this study accurately predicted individual subjects' brain maturation longitudinally. Due to its strong clinical potentials in identifying individuals with an abnormal brain maturation trajectory, the brain maturation index may allow timely clinical interventions for individuals at risk for psychiatric disorders.

Identifying neuroanatomical signatures of anorexia nervosa: a multivariate machine learning approach.

There are currently no neuroanatomical biomarkers of anorexia nervosa (AN) available to make clinical inferences at an individual subject level. We present results of a multivariate machine learning (ML) approach utilizing structural neuroanatomical scan data to differentiate AN patients from matched healthy controls at an individual subject level. Structural neuroimaging scans were acquired from 15 female patients with AN (age = 20, s.d. = 4 years) and 15 demographically matched female controls (age = 22, s.d. = 3 years). Neuroanatomical volumes were extracted using the FreeSurfer software and input into the Least Absolute Shrinkage and Selection Operator (LASSO) multivariate ML algorithm. LASSO was 'trained' to identify 'novel' individual subjects as either AN patients or healthy controls. Furthermore, the model estimated the probability that an individual subject belonged to the AN group based on an individual scan. The model correctly predicted 25 out of 30 subjects, translating into 83.3% accuracy (sensitivity 86.7%, specificity 80.0%) (p < 0.001; χ 2 test). Six neuroanatomical regions (cerebellum white matter, choroid plexus, putamen, accumbens, the diencephalon and the third ventricle) were found to be relevant in distinguishing individual AN patients from healthy controls. The predicted probabilities showed a linear relationship with drive for thinness clinical scores (r = 0.52, p < 0.005) and with body mass index (BMI) (r = -0.45, p = 0.01). The model achieved a good predictive accuracy and drive for thinness showed a strong neuroanatomical signature. These results indicate that neuroimaging scans coupled with ML techniques have the potential to provide information at an individual subject level that might be relevant to clinical outcomes.

Multivariate decoding of cerebral blood flow measures in a clinical model of on-going postsurgical pain.

Recent reports of multivariate machine learning (ML) techniques have highlighted their potential use to detect prognostic and diagnostic markers of pain. However, applications to date have focussed on acute experimental nociceptive stimuli rather than clinically relevant pain states. These reports have coincided with others describing the application of arterial spin labeling (ASL) to detect changes in regional cerebral blood flow (rCBF) in patients with on‐going clinical pain. We combined these acquisition and analysis methodologies in a well‐characterized postsurgical pain model. The principal aims were (1) to assess the classification accuracy of rCBF indices acquired prior to and following surgical intervention and (2) to optimise the amount of data required to maintain accurate classification. Twenty male volunteers, requiring bilateral, lower jaw third molar extraction (TME), underwent ASL examination prior to and following individual left and right TME, representing presurgical and postsurgical states, respectively. Six ASL time points were acquired at each exam. Each ASL image was preceded by visual analogue scale assessments of alertness and subjective pain experiences. Using all data from all sessions, an independent Gaussian Process binary classifier successfully discriminated postsurgical from presurgical states with 94.73% accuracy; over 80% accuracy could be achieved using half of the data (equivalent to 15 min scan time). This work demonstrates the concept and feasibility of time‐efficient, probabilistic prediction of clinically relevant pain at the individual level. We discuss the potential of ML techniques to impact on the search for novel approaches to diagnosis, management, and treatment to complement conventional patient self‐reporting. Hum Brain Mapp 36:633–642, 2015. © 2014 The Authors. Human Brain Mapping Published by Wiley Periodicals, Inc.

Decoding the representation of numerical values from brain activation patterns

Human neuroimaging studies have increasingly converged on the possibility that the neural representation of specific numbers may be decodable from brain activity, particularly in parietal cortex. Multivariate machine learning techniques have recently demonstrated that the neural representation of individual concrete nouns can be decoded from fMRI patterns, and that some patterns are general over people. Here we use these techniques to investigate whether the neural codes for quantities of objects can be accurately decoded. The pictorial mode (nonsymbolic) depicted a set of objects pictorially (e.g., a picture of three tomatoes), whereas the digit-object mode depicted quantities as combination of a digit (e.g., 3) with a picture of a single object. The study demonstrated that quantities of objects were decodable from neural activation patterns, in parietal regions. These brain activation patterns corresponding to a given quantity were common across objects and across participants in the pictorial mode. Other important findings included better identification of individual numbers in the pictorial mode, partial commonality of neural patterns across the two modes, and hemispheric asymmetry with pictorially-depicted numbers represented bilaterally and numbers in the digit-object mode represented primarily in the left parietal regions. The findings demonstrate the ability to identify individual quantities of objects based on neural patterns, indicating the presence of stable neural representations of numbers. Additionally, they indicate a predominance of neural representation of pictorially depicted numbers over the digit-object mode.

Decision Processes in Human Performance Monitoring

The ability to detect and compensate for errors is crucial in producing effective, goal-directed behavior. Human error processing is reflected in two event-related brain potential components, the error-related negativity (Ne/ERN) and error positivity (Pe), but the functional significance of both components remains unclear. Our approach was to consider error detection as a decision process involving an evaluation of available evidence that an error has occurred against an internal criterion. This framework distinguishes two fundamental stages of error detection--accumulating evidence (input), and reaching a decision (output)--that should be differentially affected by changes in internal criterion. Predictions from this model were tested in a brightness discrimination task that required human participants to signal their errors, with incentives varied to encourage participants to adopt a high or low criterion for signaling their errors. Whereas the Ne/ERN was unaffected by this manipulation, the Pe varied consistently with criterion: A higher criterion was associated with larger Pe amplitude for signaled errors, suggesting that the Pe reflects the strength of accumulated evidence. Across participants, Pe amplitude was predictive of changes in behavioral criterion as estimated through signal detection theory analysis. Within participants, Pe amplitude could be estimated robustly with multivariate machine learning techniques and used to predict error signaling behavior both at the level of error signaling frequencies and at the level of individual signaling responses. These results suggest that the Pe, rather than the Ne/ERN, is closely related to error detection, and specifically reflects the accumulated evidence that an error has been committed.

Early functional magnetic resonance imaging activations predict language outcome after stroke

An accurate prediction of system-specific recovery after stroke is essential to provide rehabilitation therapy based on the individual needs. We explored the usefulness of functional magnetic resonance imaging scans from an auditory language comprehension experiment to predict individual language recovery in 21 aphasic stroke patients. Subjects with an at least moderate language impairment received extensive language testing 2 weeks and 6 months after left-hemispheric stroke. A multivariate machine learning technique was used to predict language outcome 6 months after stroke. In addition, we aimed to predict the degree of language improvement over 6 months. 76% of patients were correctly separated into those with good and bad language performance 6 months after stroke when based on functional magnetic resonance imaging data from language relevant areas. Accuracy further improved (86% correct assignments) when age and language score were entered alongside functional magnetic resonance imaging data into the fully automatic classifier. A similar accuracy was reached when predicting the degree of language improvement based on imaging, age and language performance. No prediction better than chance level was achieved when exploring the usefulness of diffusion weighted imaging as well as functional magnetic resonance imaging acquired two days after stroke. This study demonstrates the high potential of current machine learning techniques to predict system-specific clinical outcome even for a disease as heterogeneous as stroke. Best prediction of language recovery is achieved when the brain activation potential after system-specific stimulation is assessed in the second week post stroke. More intensive early rehabilitation could be provided for those with a predicted poor recovery and the extension to other systems, for example, motor and attention seems feasible.

Prognosis and sleep disordered breathing in heart failure

The ability to detect and compensate for errors is crucial in producing effective, goal-directed behavior. Human error processing is reflected in two event-related brain potential components, the error-related negativity (Ne/ERN)...