Predicting brain age with deep learning from raw imaging data results in a reliable and heritable biomarker

Transfer learning improves resting-state functional connectivity pattern analysis using convolutional neural networks

Background

Deep learning is gaining importance in the prediction of cognitive states and brain pathology based on neuroimaging data. Including multiple hidden layers in artificial neural networks enables unprecedented predictive power; however, the proper training of deep neural networks requires thousands of exemplars. Collecting this amount of data is not feasible in typical neuroimaging experiments. A handy solution to this problem, which has largely fallen outside the scope of deep learning applications in neuroimaging, is to repurpose deep networks that have already been trained on large datasets by fine-tuning them to target datasets/tasks with fewer exemplars. Here, we investigated how this method, called transfer learning, can aid age category classification and regression based on brain functional connectivity patterns derived from resting-state functional magnetic resonance imaging. We trained a connectome-convolutional neural network on a larger public dataset and then examined how the knowledge learned can be used effectively to perform these tasks on smaller target datasets collected with a different type of scanner and/or imaging protocol and pre-processing pipeline.

Results

Age classification on the target datasets benefitted from transfer learning. Significant improvement (∼9%–13% increase in accuracy) was observed when the convolutional layers’ weights were initialized based on the values learned on the public dataset and then fine-tuned to the target datasets. Transfer learning also appeared promising in improving the otherwise poor prediction of chronological age.

Conclusions

Transfer learning is a plausible solution to adapt convolutional neural networks to neuroimaging data with few exemplars and different data acquisition and pre-processing protocols.

Assessing distinct patterns of cognitive aging using tissue-specific brain age prediction based on diffusion tensor imaging and brain morphometry

Multimodal imaging enables sensitive measures of the architecture and integrity of the human brain, but the high-dimensional nature of advanced brain imaging features poses inherent challenges for the analyses and interpretations. Multivariate age prediction reduces the dimensionality to one biologically informative summary measure with potential for assessing deviations from normal lifespan trajectories. A number of studies documented remarkably accurate age prediction, but the differential age trajectories and the cognitive sensitivity of distinct brain tissue classes have yet to be adequately characterized. Exploring differential brain age models driven by tissue-specific classifiers provides a hitherto unexplored opportunity to disentangle independent sources of heterogeneity in brain biology. We trained machine-learning models to estimate brain age using various combinations of FreeSurfer based morphometry and diffusion tensor imaging based indices of white matter microstructure in 612 healthy controls aged 18-87 years. To compare the tissue-specific brain ages and their cognitive sensitivity, we applied each of the 11 models in an independent and cognitively well-characterized sample (n = 265, 20-88 years). Correlations between true and estimated age and mean absolute error (MAE) in our test sample were highest for the most comprehensive brain morphometry (r = 0.83, CI:0.78-0.86, MAE = 6.76 years) and white matter microstructure (r = 0.79, CI:0.74-0.83, MAE = 7.28 years) models, confirming sensitivity and generalizability. The deviance from the chronological age were sensitive to performance on several cognitive tests for various models, including spatial Stroop and symbol coding, indicating poorer performance in individuals with an over-estimated age. Tissue-specific brain age models provide sensitive measures of brain integrity, with implications for the study of a range of brain disorders.

Active Acquisition for multimodal neuroimaging

In many clinical and scientific situations the optimal neuroimaging sequence may not be known prior to scanning and may differ for each individual being scanned, depending on the exact nature and location of abnormalities. Despite this, the standard approach to data acquisition, in such situations, is to specify the sequence of neuroimaging scans prior to data acquisition and to apply the same scans to all individuals. In this paper, we propose and illustrate an alternative approach, in which data would be analysed as it is acquired and used to choose the future scanning sequence: Active Acquisition. We propose three Active Acquisition scenarios based around multiple MRI modalities. In Scenario 1, we propose a simple use of near-real time analysis to decide whether to acquire more or higher resolution data, or acquire data with a different field -of -view. In Scenario 2, we simulate how multimodal MR data could be actively acquired and combined with a decision tree to classify a known outcome variable (in the simple example here, age). In Scenario 3, we simulate using Bayesian optimisation to actively search across multiple MRI modalities to find those which are most abnormal. These simulations suggest that by actively acquiring data, the scanning sequence can be adapted to each individual. We also consider the many outstanding practical and technical challenges involving normative data acquisition, MR physics, statistical modelling and clinical relevance. Despite these, we argue that  Active Acquisition allows for potentially far more powerful, sensitive or rapid data acquisition, and may open up different perspectives on individual differences, clinical conditions, and biomarker discovery.

Brain age from the electroencephalogram of sleep

The human electroencephalogram (EEG) of sleep undergoes profound changes with age. These changes can be conceptualized as "brain age (BA)," which can be compared to chronological age to reflect the degree of deviation from normal aging. Here, we develop an interpretable machine learning model to predict BA based on 2 large sleep EEG data sets: the Massachusetts General Hospital (MGH) sleep lab data set (N = 2532; ages 18-80); and the Sleep Heart Health Study (SHHS, N = 1974; ages 40-80). The model obtains a mean absolute deviation of 7.6 years between BA and chronological age (CA) in healthy participants in the MGH data set. As validation, a subset of SHHS containing longitudinal EEGs 5.2 years apart shows an average of 5.4 years increase in BA. Participants with significant neurological or psychiatric disease exhibit a mean excess BA, or "brain age index" (BAI = BA-CA) of 4 years relative to healthy controls. Participants with hypertension and diabetes have a mean excess BA of 3.5 years. The findings raise the prospect of using the sleep EEG as a potential biomarker for healthy brain aging.

The application of artificial intelligence to understand the pathophysiological basis of psychogenic nonepileptic seizures

Psychogenic nonepileptic seizures (PNES) are episodes of paroxysmal impairment associated with a range of motor, sensory, and mental manifestations, which perfectly mimic epileptic seizures. Several patterns of neural abnormalities have been described without identifying a definite neurobiological substrate. In this multicenter cross-sectional study, we applied a multivariate classification algorithm on morphological brain imaging metrics to extract reliable biomarkers useful to distinguish patients from controls at an individual level. Twenty-three patients with PNES and 21 demographically matched healthy controls (HC) underwent an extensive neuropsychiatric/neuropsychological and neuroimaging assessment. One hundred and fifty morphological brain metrics were used for training a random forest (RF) machine-learning (ML) algorithm. A typical complex psychopathological construct was observed in PNES. Similarly, univariate neuroimaging analysis revealed widespread neuroanatomical changes affecting patients with PNES. Machine-learning approach, after feature selection, was able to perform an individual classification of PNES from controls with a mean accuracy of 74.5%, revealing that brain regions influencing classification accuracy were mainly localized within the limbic (posterior cingulate and insula) and motor inhibition systems (the right inferior frontal cortex (IFC)). This study provides Class II evidence that the considerable clinical and neurobiological heterogeneity observed in individuals with PNES might be overcome by ML algorithms trained on surface-based magnetic resonance imaging (MRI) data.

Use of Machine Learning to Determine Deviance in Neuroanatomical Maturity Associated With Future Psychosis in Youths at Clinically High Risk

IMPORTANCE

Altered neurodevelopmental trajectories are thought to reflect heterogeneity in the pathophysiologic characteristics of schizophrenia, but whether neural indicators of these trajectories are associated with future psychosis is unclear.

OBJECTIVE

To investigate distinct neuroanatomical markers that can differentiate aberrant neurodevelopmental trajectories among clinically high-risk (CHR) individuals.

DESIGN, SETTING, AND PARTICIPANTS

In this prospective longitudinal multicenter study, a neuroanatomical-based age prediction model was developed using a supervised machine learning technique with T1-weighted magnetic resonance imaging scans of 953 healthy controls 3 to 21 years of age from the Pediatric Imaging, Neurocognition, and Genetics (PING) study and then applied to scans of 275 CHR individuals (including 39 who developed psychosis) and 109 healthy controls 12 to 21 years of age from the North American Prodrome Longitudinal Study 2 (NAPLS 2) for external validation and clinical application. Scans from NAPLS 2 were collected from January 15, 2010, to April 30, 2012.

MAIN OUTCOMES AND MEASURES

Discrepancy between neuroanatomical-based predicted age (hereafter referred to as brain age) and chronological age.

RESULTS

The PING-derived model (460 females and 493 males; age range, 3-21 years) accurately estimated the chronological ages of the 109 healthy controls in the NAPLS 2 (43 females and 66 males; age range, 12-21 years), providing evidence of independent external validation. The 275 CHR individuals in the NAPLS 2 (111 females and 164 males; age range, 12-21 years) showed a significantly greater mean (SD) gap between model-predicted age and chronological age (0.64 [2.16] years) compared with healthy controls (P = .008). This outcome was significantly moderated by chronological age, with brain age systematically overestimating the ages of CHR individuals who developed psychosis at ages 12 to 17 years but not the brain ages of those aged 18 to 21 years. Greater brain age deviation was associated with a higher risk for developing psychosis (F = 3.70; P = .01) and a pattern of stably poor functioning over time, but only among younger CHR adolescents. Previously reported evidence of accelerated reduction in cortical thickness among CHR individuals who developed psychosis was found to apply only to those who were 18 years of age or older.

CONCLUSIONS AND RELEVANCE

These results are consistent with the view that neuroanatomical markers of schizophrenia may help to explain some of the heterogeneity of this disorder, particularly with respect to early vs later age of onset of psychosis, with younger and older individuals having differing intercepts and trajectories in structural brain parameters as a function of age. The results also suggest that baseline neuroanatomical measures are likely to be useful in estimating onset of psychosis, especially (or only) among CHR individuals with an earlier age of onset of prodromal symptoms.

Biological Brain Age Prediction Using Cortical Thickness Data: A Large Scale Cohort Study

Brain age estimation from anatomical features has been attracting more attention in recent years. This interest in brain age estimation is motivated by the importance of biological age prediction in health informatics, with an application to early prediction of neurocognitive disorders. It is well-known that normal brain aging follows a specific pattern, which enables researchers and practitioners to predict the age of a human's brain from its degeneration. In this paper, we model brain age predicted by cortical thickness data gathered from large cohort brain images. We collected 2,911 cognitively normal subjects (age 45-91 years) at a single medical center and acquired their brain magnetic resonance (MR) images. All images were acquired using the same scanner with the same protocol. We propose to first apply Sparse Group Lasso (SGL) for feature selection by utilizing the brain's anatomical grouping. Once the features are selected, a non-parametric non-linear regression using the Gaussian Process Regression (GPR) algorithm is applied to fit the final age prediction model. Experimental results demonstrate that the proposed method achieves the mean absolute error of 4.05 years, which is comparable with or superior to several recent methods. Our method can also be a critical tool for clinicians to differentiate patients with neurodegenerative brain disease by extracting a cortical thinning pattern associated with normal aging.

Structure-function multi-scale connectomics reveals a major role of the fronto-striato-thalamic circuit in brain aging

Physiological aging affects brain structure and function impacting morphology, connectivity, and performance. However, whether some brain connectivity metrics might reflect the age of an individual is still unclear. Here, we collected brain images from healthy participants (N = 155) ranging from 10 to 80 years to build functional (resting state) and structural (tractography) connectivity matrices, both data sets combined to obtain different connectivity features. We then calculated the brain connectome age-an age estimator resulting from a multi-scale methodology applied to the structure-function connectome, and compared it to the chronological age (ChA). Our results were twofold. First, we found that aging widely affects the connectivity of multiple structures, such as anterior cingulate and medial prefrontal cortices, basal ganglia, thalamus, insula, cingulum, hippocampus, parahippocampus, occipital cortex, fusiform, precuneus, and temporal pole. Second, we found that the connectivity between basal ganglia and thalamus to frontal areas, also known as the fronto-striato-thalamic (FST) circuit, makes the major contribution to age estimation. In conclusion, our results highlight the key role played by the FST circuit in the process of healthy aging. Notably, the same methodology can be generally applied to identify the structural-functional connectivity patterns correlating to other biomarkers than ChA.

Predicting Age From Brain EEG Signals-A Machine Learning Approach

Objective: The brain age gap estimate (BrainAGE) is the difference between the estimated age and the individual chronological age. BrainAGE was studied primarily using MRI techniques. EEG signals in combination with machine learning (ML) approaches were not commonly used for the human age prediction, and BrainAGE. We investigated whether age-related changes are affecting brain EEG signals, and whether we can predict the chronological age and obtain BrainAGE estimates using a rigorous ML framework with a novel and extensive EEG features extraction. Methods: EEG data were obtained from 468 healthy, mood/anxiety, eating and substance use disorder participants (297 females) from the Tulsa-1000, a naturalistic longitudinal study based on Research Domain Criteria framework. Five sets of preprocessed EEG features across channels and frequency bands were used with different ML methods to predict age. Using a nested-cross-validation (NCV) approach and stack-ensemble learning from EEG features, the predicted age was estimated. The important features and their spatial distributions were deduced. Results: The stack-ensemble age prediction model achieved R2 = 0.37 (0.06), Mean Absolute Error (MAE) = 6.87(0.69) and RMSE = 8.46(0.59) in years. The age and predicted age correlation was r = 0.6. The feature importance revealed that age predictors are spread out across different feature types. The NCV approach produced a reliable age estimation, with features consistent behavior across different folds. Conclusion: Our rigorous ML framework and extensive EEG signal features allow a reliable estimation of chronological age, and BrainAGE. This general framework can be extended to test EEG association with and to predict/study other physiological relevant responses.

Neuroimaging Studies Illustrate the Commonalities Between Ageing and Brain Diseases

The lack of specificity in neuroimaging studies of neurological and psychiatric diseases suggests that these different diseases have more in common than is generally considered. Potentially, features that are secondary effects of different pathological processes may share common neurobiological underpinnings. Intriguingly, many of these mechanisms are also observed in studies of normal (i.e., non-pathological) brain ageing. Different brain diseases may be causing premature or accelerated ageing to the brain, an idea that is supported by a line of "brain ageing" research that combines neuroimaging data with machine learning analysis. In reviewing this field, I conclude that such observations could have important implications, suggesting that we should shift experimental paradigm: away from characterizing the average case-control brain differences resulting from a disease toward methods that place individuals in their age-appropriate context. This will also lead naturally to clinical applications, whereby neuroimaging can contribute to a personalized-medicine approach to improve brain health.

Evaluation of non-negative matrix factorization of grey matter in age prediction

The relationship between grey matter volume (GMV) patterns and age can be captured by multivariate pattern analysis, allowing prediction of individuals' age based on structural imaging. Raw data, voxel-wise GMV and non-sparse factorization (with Principal Component Analysis, PCA) show good performance but do not promote relatively localized brain components for post-hoc examinations. Here we evaluated a non-negative matrix factorization (NNMF) approach to provide a reduced, but also interpretable representation of GMV data in age prediction frameworks in healthy and clinical populations. This examination was performed using three datasets: a multi-site cohort of life-span healthy adults, a single site cohort of older adults and clinical samples from the ADNI dataset with healthy subjects, participants with Mild Cognitive Impairment and patients with Alzheimer's disease (AD) subsamples. T1-weighted images were preprocessed with VBM8 standard settings to compute GMV values after normalization, segmentation and modulation for non-linear transformations only. Non-negative matrix factorization was computed on the GM voxel-wise values for a range of granularities (50-690 components) and LASSO (Least Absolute Shrinkage and Selection Operator) regression were used for age prediction. First, we compared the performance of our data compression procedure (i.e., NNMF) to various other approaches (i.e., uncompressed VBM data, PCA-based factorization and parcellation-based compression). We then investigated the impact of the granularity on the accuracy of age prediction, as well as the transferability of the factorization and model generalization across datasets. We finally validated our framework by examining age prediction in ADNI samples. Our results showed that our framework favorably compares with other approaches. They also demonstrated that the NNMF based factorization derived from one dataset could be efficiently applied to compress VBM data of another dataset and that granularities between 300 and 500 components give an optimal representation for age prediction. In addition to the good performance in healthy subjects our framework provided relatively localized brain regions as the features contributing to the prediction, thereby offering further insights into structural changes due to brain aging. Finally, our validation in clinical populations showed that our framework is sensitive to deviance from normal structural variations in pathological aging.

BRAIN AGE PREDICTION BASED ON RESTING-STATE FUNCTIONAL CONNECTIVITY PATTERNS USING CONVOLUTIONAL NEURAL NETWORKS

Brain age prediction based on neuroimaging data could help characterize both the typical brain development and neuropsychiatric disorders. Pattern recognition models built upon functional connectivity (FC) measures derived from resting state fMRI (rsfMRI) data have been successfully used to predict the brain age. However, most existing studies focus on coarse-grained FC measures between brain regions or intrinsic connectivity networks (ICNs), which may sacrifice fine-grained FC information of the rsfMRI data. Whole brain voxel-wise FC measures could provide fine-grained FC information of the brain and may improve the prediction performance. In this study, we develop a deep learning method to use convolutional neural networks (CNNs) to learn informative features from the fine-grained whole brain FC measures for the brain age prediction. Experimental results on a large dataset of resting-state fMRI demonstrate that the deep learning model with fine-grained FC measures could better predict the brain age.

Bayesian Optimization for Neuroimaging Pre-processing in Brain Age Classification and Prediction

Neuroimaging-based age prediction using machine learning is proposed as a biomarker of brain aging, relating to cognitive performance, health outcomes and progression of neurodegenerative disease. However, even leading age-prediction algorithms contain measurement error, motivating efforts to improve experimental pipelines. T1-weighted MRI is commonly used for age prediction, and the pre-processing of these scans involves normalization to a common template and resampling to a common voxel size, followed by spatial smoothing. Resampling parameters are often selected arbitrarily. Here, we sought to improve brain-age prediction accuracy by optimizing resampling parameters using Bayesian optimization. Using data on N = 2003 healthy individuals (aged 16-90 years) we trained support vector machines to (i) distinguish between young (<22 years) and old (>50 years) brains (classification) and (ii) predict chronological age (regression). We also evaluated generalisability of the age-regression model to an independent dataset (CamCAN, N = 648, aged 18-88 years). Bayesian optimization was used to identify optimal voxel size and smoothing kernel size for each task. This procedure adaptively samples the parameter space to evaluate accuracy across a range of possible parameters, using independent sub-samples to iteratively assess different parameter combinations to arrive at optimal values. When distinguishing between young and old brains a classification accuracy of 88.1% was achieved, (optimal voxel size = 11.5 mm³, smoothing kernel = 2.3 mm). For predicting chronological age, a mean absolute error (MAE) of 5.08 years was achieved, (optimal voxel size = 3.73 mm³, smoothing kernel = 3.68 mm). This was compared to performance using default values of 1.5 mm³ and 4mm respectively, resulting in MAE = 5.48 years, though this 7.3% improvement was not statistically significant. When assessing generalisability, best performance was achieved when applying the entire Bayesian optimization framework to the new dataset, out-performing the parameters optimized for the initial training dataset. Our study outlines the proof-of-principle that neuroimaging models for brain-age prediction can use Bayesian optimization to derive case-specific pre-processing parameters. Our results suggest that different pre-processing parameters are selected when optimization is conducted in specific contexts. This potentially motivates use of optimization techniques at many different points during the experimental process, which may improve statistical sensitivity and reduce opportunities for experimenter-led bias.

Modelling neuroanatomical variation during childhood and adolescence with neighbourhood-preserving embedding

Brain development is a dynamic process with tissue-specific alterations that reflect complex and ongoing biological processes taking place during childhood and adolescence. Accurate identification and modelling of these anatomical processes in vivo with MRI may provide clinically useful imaging markers of individual variability in development. In this study, we use manifold learning to build a model of age- and sex-related anatomical variation using multiple magnetic resonance imaging metrics. Using publicly available data from a large paediatric cohort (n = 768), we apply a multi-metric machine learning approach combining measures of tissue volume, cortical area and cortical thickness into a low-dimensional data representation. We find that neuroanatomical variation due to age and sex can be captured by two orthogonal patterns of brain development and we use this model to simultaneously predict age with a mean error of 1.5-1.6 years and sex with an accuracy of 81%. We validate this model in an independent developmental cohort. We present a framework for modelling anatomical development during childhood using manifold embedding. This model accurately predicts age and sex based on image-derived markers of cerebral morphology and generalises well to independent populations.

Computational biology: deep learning

Deep learning is the trendiest tool in a computational biologist's toolbox. This exciting class of methods, based on artificial neural networks, quickly became popular due to its competitive performance in prediction problems. In pioneering early work, applying simple network architectures to abundant data already provided gains over traditional counterparts in functional genomics, image analysis, and medical diagnostics. Now, ideas for constructing and training networks and even off-the-shelf models have been adapted from the rapidly developing machine learning subfield to improve performance in a range of computational biology tasks. Here, we review some of these advances in the last 2 years.