Brain Structure and Function Predict Adherence to an Exercise Intervention in Older Adults

Abstract

Introduction:

Individual differences in brain structure and function in older adults are potential proxies of brain reserve or maintenance and may provide mechanistic predictions of adherence to exercise. We hypothesized that multimodal neuroimaging features would predict adherence to a six-month randomized controlled trial of exercise in 131 older adults (aged 65.79 (4.65) years, 63 percent female), alone and in combination with psychosocial, cognitive and health measures.

Methods:

Regularized elastic net regression within a nested cross-validation framework was applied to predict adherence to the intervention in three separate models (brain structure and function only, psychosocial, health and demographic data only, and a multimodal model).

Results:

Higher cortical thickness in somatosensory and inferior frontal regions and less surface area in primary visual and inferior frontal regions predicted adherence. Higher nodal functional connectivity (degree count) in default, frontoparietal and attentional networks, and less nodal strength in primary visual and temporoparietal networks predicted exercise adherence (r = 0.24, p = 0.004). Survey and clinical measures of gait and walking self-efficacy, biological sex and perceived stress also predicted adherence (r = 0.17, p = 0.056), however this prediction was not significant when tested against a null test statistic. A combined multimodal model achieved the highest predictive strength (r = 0.28, p = 0.001).

Conclusions:

Our results suggest there is substantial utility of using brain-based measures in future research into precision and individualized exercise interventions older adults.

1. INTRODUCTION

Physical activity and structured exercise have received a lot of attention as potential efficacious interventions to improve or maintain brain health with advancing age [1]. Despite millions of government and private dollars being spent on understanding how physical activity can improve or maintain brain function across the lifespan, over a third of the US population do not engage in sufficient physical activity [2], a statistic which continues to increase [2]. This physical inactivity pandemic [3] is estimated to cost private and public health-care systems $53.8 billion per year and a further $13.7 billion in productivity losses due to physical inactivity-related deaths each year[4].

Previous research on understanding the adoption and maintenance of physical activity has leveraged psychological and psychosocial theories [5]. In experimental studies, several psychosocial, behavioral and demographic measures were shown to correlate with exercise adherence, such as self-efficacy [6], self-regulation [7], social support, perceived benefits and biological sex (males adhered more than females) [8], higher baseline physical activity outside of the intervention [9] as well as depression, fatigue and general perceived health [10]. Additionally, greater cognitive resources, particularly executive functions, known to depend on cortical structural integrity [11], have been associated with engagement in exercise [12,13]. While theory-based programs are generally successful, optimizable, and adaptable precision medicine approaches will improve cost-efficiency of interventions and health outcomes for individuals who show poor adherence to one-size fits all approaches. That is, predicting from a set of variables, who is more or less likely to adhere to an intervention from the outset will allow the practitioner to provide alternative and individualized interventions prior to individuals having to either demonstrate poor adherence or present with deteriorating health.

Within the intervention setting, several recent studies have demonstrated that pre-intervention brain structure, known to be critical in supporting executive functions [11], is associated with individual differences in adherence to exercise interventions. Specifically, regions of the prefrontal, temporal and somatosensory cortices have been associated with adherence to each respective exercise intervention [14,15]. Relatedly, several reviews and observational studies have suggested that the relationship between exercise and the brain is bi-directional [13,16]. This hypothesis leverages the concepts of cognitive reserve and brain maintenance [17], such that there is a circular nature between greater cognitive and brain resources and higher participation in complex exercise behaviors, which in turn helps with the maintenance and upkeep of cognitive and brain health. While rarely measured directly, proxy measures of brain reserve and brain maintenance include brain structure and brain functional connectivity (for a review see: Stern et al., 2019). In terms of function, numerous applications of functional and anatomical connectivity have led to the observation that the brain is organized into large-scale functional networks [18]. A number of these networks, such as the frontoparietal control network (FPCN), the default mode network (DMN) and the dorsal attention network (DAN) change with age [19] and are thought to be particularly important in age-related cognitive decline [20]. These networks also subserve internally and externally directed cognition [21], including broadly defined executive function. Given the heterogeneity of these brain networks and their importance in higher-order cognition, individual patterns of functional connectivity in older adults may be predictive of exercise adherence. Indeed, prior work on mindfulness and meditation have demonstrated that functional connectivity within the default mode network and in frontal and temporal nodes are strong predictors of adherence [22].

The term “prediction” is used in a number of ways in the literature and can commonly refer to either the correlation of one variable in a group at one point in time with another variable in that same group at another point in time (within-sample correlation) or it can refer to a generalizable model that makes predictions on out-of-sample participants [23]. Prior studies on psychosocial and behavioral “predictions” of exercise adherence have typically used correlation and therefore their results are likely overly optimistic [24]. This overestimation and poor generalizability is compounded further when one considers they only explain relatively small amount of variance (<20%) in adherence [25,26]. Consequently, calls have been made to introduce prediction statistics into psychological and cognitive neuroscience research to improve the generalizability of the results, specifically when using brain-based metrics [23]. In the case of exercise interventions for example, being able to successfully predict adherence to exercise interventions with generalizable results prior to engagement in the intervention would allow for optimization or individualization of an alternative intervention for those who are predicted to poorly adhere. For example, by providing those who are predicted to adhere poorly with a health coach [27] or just-in-time messaging [28], paving the way towards efficacious precision medicine approaches in health-based settings, when time can rarely be lost, and cost-efficiency is of upmost importance. This is also important as engagement in effective exercise interventions may lead to successful maintenance of physical activity after cessation of the intervention [29] and so predicting adherence to the exercise intervention may also provide insights into sustained exercise behaviors. Moreover, if brain-based metrics are shown to predict adherence, they can then provide mechanistic understanding of this complex behavior. Under the NIH stage model of intervention development, such information could then be leveraged to design novel interventions that improve exercise adherence by targeting certain predictive mechanisms and testing whether their modulation improves adherence.

In a secondary analysis of data from a randomized control trial of exercise in older adults, the objectives of this study were to predict adherence to a 6-month exercise intervention and an active control in a structured and supervised group-based intervention. We hypothesized that multiple metrics of brain structure and function in frontoparietal control, default mode and dorsal attention networks and regions (prefrontal, temporal, parietal) would predict adherence to the exercise intervention and that these measures would augment traditional behavioral and psychosocial measures. To test this, we used a whole brain, data-driven machine learning approach.

2. METHODS

2.1. Participants

This study is a secondary analysis of data from participants who participated in a six-month randomized controlled exercise trial (clinical study identifier: NCT01472744, November 16, 2011). The study procedures were approved by the University of Illinois Institutional Review Board and written informed consent was obtained from all participants prior to any research activities. Healthy but low active older adults were recruited in Champaign County, Illinois. Two hundred and forty-seven (169 women) low-active older adults met inclusion criteria for the initial clinical trial. Of which, 165 underwent structural and functional magnetic resonance imaging (MRI) at baseline, prior to any involvement in intervention sessions and 131 had complete data across all variables and were included in this analysis (see further exclusion criteria based on adherence below). Participants in the initial trial were randomized to one of four intervention conditions; a walking intervention, the same walking intervention plus a dietary supplement designed to enhance lean muscle mass, a dancing intervention and an active control consisting of a stretching and toning intervention. All intervention groups met for approximately one hour, three times per week for six months and for the purpose of this analysis we analyzed our main outcome (adherence) with all four conditions combined. A one-way analysis of variance revealed no significant differences in adherence between the conditions (F_1,129 =1.11, p = 0.294) and no significant differences in baseline characteristics between conditions existed (see Table S1, Supplemental Digital Content, SDC 1, Demographics stratified by group assignment). Further we replicate the pattern results in the most common intervention condition (walking) only as a sensitivity analysis with a reduced sample size (n=59) to ensure our results were not affected by condition assignment (see Table S2, Supplemental Digital Content, SDC 2, Prediction models in the walking and walking + group only). Initial inclusion criteria included being between the ages of 60 and 80 years old, free from psychiatric and neurological illness including no history of stroke, transient ischemic attack, or head trauma, scored > 23 on the Mini-Mental State Exam, > 21 on a Telephone Interview of Cognitive Status questionnaire, < 10 on the Geriatric Depression Scale, at least 75% right-handed based on the Edinburgh Handedness Questionnaire (a criterion related to functional magnetic resonance imaging (MRI) analyses), demonstrated normal or corrected-to-normal vision of at least 20/40 and no color blindness, screened for safe participation in an MRI environment (e.g., no metallic implants that could interfere with the magnetic field or cause injury and no claustrophobia) and reported to have participated in no more than two bouts of moderate exercise per week within the past six months (with the goal of recruiting low active older adults). Table 1 contains complete characterization of the study participants included in this analysis. All methods were carried out in accordance with the Declaration of Helsinki.

Table 1.

Participant Characteristics

N	131
Age (mean (SD))	65.79 (4.65)
Gender: Female (%)	63 (71.6)
Race (%)
White	79 (89.7)
African American or black	7 (8)
Asian	2 (2.3)
Years of education (mean (SD))	15.89 (2.66)
Adherence (%)	81 (12)

2.2. Adherence

Our primary outcome, adherence, was modelled as percentage attendance to the weekly intervention sessions. These supervised sessions were scheduled three times per week and lasted approximately for one hour. The number and frequency of the sessions were consistent across all four conditions. For this analysis, participants (N=34) who dropped-out of the study (i.e., did not complete the intervention) were excluded, as were those who failed to attend at least half of the intervention sessions (n= 12) as our aim was to capture and predict variations in adherence to the entire 6-month intervention.

2.3. Psychosocial, physical function and activity, health, and cognitive features

A comprehensive battery of psychosocial, physical, and cognitive assessments was completed by each participant pre-intervention. A complete table of these assessments is found in Table S3 (see Supplemental Digital Content, SDC 3, Complete list of behavioral variables). Eighty-four assessments were initially included in this analysis. In brief, these assessments included self-reported psychosocial questionnaires gauging participants self-efficacy (global, exercise and gait self-efficacy), leisure time activity, perceived sleep quality, anxiety, depression and self-worth/esteem, barriers to exercise, self-regulation, stress, loneliness, and subjective memory. All psychosocial measures were taken at week one or in some instances repeated at week three. A set of physical function tests were also collected at week one, which included a stair climb test, arm curl, sit and reach and back scratch. Seven days of accelerometry capturing objective measures of time spent sedentary, time spent in light or moderate-to-vigorous aerobic physical activity and average daily step counts were also collected. The procedures to capture, preprocess and validate these measures can be found in a previous publication [30]. A measure of cardiorespiratory fitness from a complete cardiopulmonary exercise test was also included as were measures of body composition (see our previous work [31] for a detailed description of the methodology of these measures). Finally, a battery of neuropsychological tasks were completed in week one which included numerous assessments of vocabulary, abstract, inductive and visuospatial reasoning, memory and perceptual speed, taken from the Virginia Cognitive Aging Project [32]. Intervention condition assignment was also included as a feature.

2.4. Neuroimaging features

2.4.1. Magnetic resonance imaging: acquisition

Participants undertook an MRI scanning session in a three Tesla Siemens Trio Tim system with a 12-channel head coil prior to the intervention. High-resolution structural MRI scans were acquired using 3D MPRAGE T1-weighted sequences (TR = 1900 ms; TE = 2.32 ms; TI: 900 ms; flip angle = 9°; matrix = 256 × 256; FOV = 230 mm; 192 slices; resolution = 0.9 × 0.9 × 0.9 mm; GRAPPA acceleration factor 2). T2*-weighted resting state echoplanar imaging (EPI) data was obtained with the following parameters: (6min,TR=2s,TE=25ms,flipangle=80°, 3.4 × 3.4 mm² in-plane resolution, 35 4 mm-thick slices acquired in ascending order, Grappa acceleration factor = 2, 64 × 64 matrix). Structural and resting state functional images were acquired with these scanning parameters and the preprocessing and analyses of each respective modality are outlined in the following two sections.

2.4.2. Structural MRI preprocessing and analyses

Cortical reconstruction and image segmentation and estimation of the cortical surface models was performed using the freely available FreeSurfer software v.5.3 (http://surfer-nmr.mgh.harvard.edu/). For preprocessing of the cortex a 3-dimensional surface model was created using the “recon-all” surface-based stream. Automated Talairach transformation and intensity normalization were followed by nonbrain tissue removal, tessellation of the gray and white matter boundary and automated topology correction. Finally, surface deformation enabled the detection of tissue boundaries; grey–white and grey–CSF borders. The cortical surfaces were then inflated and registered to a spherical atlas that used individual cortical folding patterns to match cortical geometry across participants. Cortical thickness was calculated at each vertex in the cortex as a measure of the distance between the white and pial surfaces and cortical surface area was calculated by averaging the area of all faces that meet at a given vertex on the white matter surface. We chose to analyze cortical thickness and surface area separately given their genetic independence and sensitivity to clinical and aging outcomes [33]. Automatic labelling per the Desikan-Killany cortical parcellation scheme was performed and average cortical thickness and cortical surface area were calculated within each parcellation resulting in 136 structural features (68 features per modality) to be used for feature selection.

2.4.3. Resting state functional connectivity preprocessing and analyses

Preprocessing of the functional resting state data was performed using the CONN-toolbox v.19c [34], relying upon SPM v.12 (Wellcome Department of Imaging Neuroscience, UCL, London, UK) in MATLAB R2019a (The MathWorks Inc, Natick, MA, USA). The default preprocessing pipeline implemented in CONN was performed which consists of the following steps: functional realignment and unwarping, slice timing correction, outlier identification, segmentation (into grey matter, white matter and cerebrospinal fluid tissue) and normalization into standard Montreal Neurologic Institute (MNI) space with 2mm isotropic voxels for functional data and 1mm for anatomical data, using 4^th order spline interpolation. Finally, functional scans were spatially smoothed using a 6mm Gaussian kernel. During the outlier detection step, acquisitions with framewise displacement above 0.9mm (per several prior publications in studies with older adults who are more susceptible to movement within the scanner [35-37] or global BOLD signal changes above five standard deviations were flagged as potential outliers using the Artefact Detection Tools (www.nitrc.org/projects/artifact_detect). Two participants were removed from the final analyses for having >40 volumes flagged. This cut off was determined based on preserving at least five minutes of scanning time [38]. Additionally, mean framewise displacement was calculated via the Jenkinson method and regressed out of the final analysis (see statistical analysis section below). This was done to be over-conservative given previous studies have shown high degree of motion-behavior correlations [39], despite the fact that no motion variable was significantly correlated with adherence outcomes in our study (all p > 0.1). Denoising of the functional data was performed using a component-based correction method, CompCor [40] and temporal band-pass filtering (0.01-0.1Hz) to remove physiological, subject-motion and outlier artefacts. Linear regression was used to remove the effects of these artifacts on the BOLD time series for each voxel and each subject considering noise components from voxels within white matter and cerebrospinalfluid, estimated subject-motion parameters (three rotation and three translation parameters and six other parameters representing their first order time derivatives), scrubbing and constant and first-order linear session effects.

To prepare the functional data for feature selection, we parcellated the functional scans into a medium resolution 300 region (node) atlas with a 17-network parcellation scheme [41]. A medium resolution parcellation was chosen to strike a balance between having sufficient dimensionality to capture predictive information within the functional connectivity matrix and having too many dimensions that would create a very large search space for the subsequent model building described below. Mean BOLD activity was calculated within each node and transformed into a 300 by 300 correlation matrix where the timeseries at each node was correlated with that of every other node. This matrix then underwent Fisher’s Z transformation. Functional brain connectivity was summarized using degree count, a graph theory metric which represents whole brain connection density for each node, per a previous publication [42]. This metric reduces the dimensions of the functional connectivity data into a sparse matrix that better represents real world graphs [43]. Degree count for each participant was calculated at several thresholds using a fixed network cost (keeping the strongest 15%, 20% and 25% of connections) and a final threshold was chosen through cross-validation (see statistical analysis section). Degree count at each node here represents a single measure of the sum of connections between that node and every other node in the cortex characterizing its degree of connectedness within the cortical gray matter.

2.5. Power analysis

We performed a power analysis (in R using the “pwr” package) on our sample size to ensure sufficient power was gained to detect a true effect. Based on our sample size (N = 131) and an assumed type I error rate of 0.05, we calculated an estimated 96% power to detect an effect size of 0.22 from a general linear model with 38 covariates (multimodal model). In the reduced sample size of N=59 for the sensitivity analysis (see Table S2, Supplemental Digital Content, SDC 2, Prediction models in the walking and walking + group only) this power fell to 60% for the same effect size.

2.6. Statistical analysis

To predict adherence to the exercise interventions we used Elastic net regression, a regularized (penalized) regression method within a nested cross-validation procedure. Elastic net is a data-driven machine learning regression method that applies a penalty term to each variable (feature) coefficient in the model, resulting in the shrinkage of that coefficient’s value. This penalty term is a function of the overall weight of a given feature in a model and serves to improve the generalizability of the model results when applied to new data. Elastic net aims to avoid overfitting (a problem in statistics that leads to poor generalizability of study results) by producing less-complex models via applying both a L1 and L2 norm penalty to calculate the coefficients of each feature. Elastic net is a linear combination of both Ridge regression (L2 norm) and least absolute selection and shrinkage operate (Lasso) regression. The L1 penalty shrinks coefficients towards zero (Ridge). The L2 penalty, in the case of features that have no predictive value, shrinks their coefficients to exactly zero (Lasso), relative to the maximum likelihood estimates, effectivity removing them from the model. The predictors with non-zero coefficients are therefore interpreted as those that contained predictive information and contributed to the final model of predicting the outcome. This model is particularly useful in cases where correlated features are present (i.e., neuroimaging features and some psychosocial measures) as the combination of both the L1 and L2 penalties will maintain groups of correlated features in the model (whereas the L1 penalty (Lasso) alone would remove all but one of the correlated features). The amount of shrinkage is determined via tuning of two hyperparameters λ₁ and λ₂, whereby via a grid search approach, the optimal combination of penalties is “tuned” by running the models over and over and assessing the predictive performance at each level of the penalty. Results of the optimal hyperparameters for each model are found in Table 3. In this analysis we used a nested cross-validation procedure where the dataset is split into a 10-fold outer loop (to evaluate model performance) and a 10-fold inner loop (to tune the hyperparameters using grid search within each inner fold). Nested-cross validation avoids optimization bias that simple cross-validation could potentially suffer from when using the same folds to both tune the hyperparameters and test the prediction performance on. In this case, the folds were kept consistent across each model by setting the same random seed and all variables were centered and scaled within each inner loop and applied to the outer folds. Cross-validation is an important step in prediction modelling whereby the predictive performance of a model is assessed on left-out data (a model is trained in n-1 folds and tested on the left-out fold in an iterative fashion), resulting in an unbiased measure of the prediction performance on unseen data. This step overcomes the issues of simple multiple linear regression whereby a good model (fit to the whole dataset) may not be a good predictive model when applied to new data, resulting in generalizability issues.

Table 3.

Data-driven features

Feature	Psych/cog/health		Imaging		Multimodal
	λ1	λ2	λ1	λ2	λ1	λ2
	n/a	98.5	0.1	100	0.46	56
lh_extrastriate_centralvisual			−0.00048		0.9136
lh_lateralPFC_FPCNa			0.00259		0.00156
lh_lateralPFC_DMNb			0.00572		0.00558
lh_ventralPFC_DMNb			0.00553		0.00572
lh_temporalparietal			−0.00687		−0.00765
rh_extrastriatesuperior_perivisual			−0.00326		−0.00243
rh_S2_SMb			−0.00531		−0.00594
rh_parsopercularis_VANa			−0.00522		−0.00332
rh_frontalmedial_VANa			0.00736		0.00632
rh_medialprefrontal_DMNa			0.00635		0.00636
rh_temporalparietal			−0.00131		0
rh_temporalparietal			−0.00975		−0.0112
lh_parsorbitalis_thickness			0.00512		0.00435
lh_parstriangularis_thickness			0.00169		0.00049
lh_postcentral_thickness			0.00627		0.00779
rh_isthmuscingulate_thickness			−0.01194		−0.01149
rh_postcentral_thickness			0.00453		0.00476
rh_frontalpole_thickness			0.00153		0.00104
lh_cuneus_area			−0.00218		0
lh_parsorbitalis_area			−0.00464		−0.00798
lh_pericalcarine_area			−0.00731		−0.00684
rh_lateraloccipital_area			−0.00563		−0.04543
Self-efficacy for walking (w3)	−0.00073				−0.01009
Barriers self-efficacy (w1)	0.00014				0.00116
Big Five; Conscientiousness	0.00068				0
Avg time in light exercise (7 days)	0.87631				0
Biological sex	−0.04877				−0.02005
Strength self-esteem	−0.00059				−0.00049
Employment status	0.00861				0.00561

Data driven features derived from selection by filtering approach within the cross-validation framework which were used as initial input features in the elastic net models. Imaging feature names are constructed as follows: 'hemishere_region_network/structural measure'. Network is derived from the 17-network cortical parcellation from Yeo et al., 2011 via automatic labeling using the Schaefer 300 atlas. Coefficients from each prediction model represent the standardized and penalized coefficient used to predict Adherence. A value of 0 means that the feature was effectivity removed from the model due to having little or no predictive value. λ1 and λ2 represent the optimal cross-validated hyperparameters for each model. λ1 is “n/a” for the psych/cog/health model as simple cross-validation and Ridge regression (L2 norm penalty only) was used to predict adherence as the Elastic net with nested cross validation failed to find a solution to predict adherence with these features only. Biological sex was coded as 2 for women and 1 for men (i.e., a negative coefficient means that men adhered better than women). Employment status was coded as 1-7 with 1 being full time employed, 2 part time, 3 retired, part time, 4 retired, 5 laid off/unemployed.

First, to select the optimal degree threshold of the functional connectivity metric to use in the final models, functional connectivity only models were trained over each threshold density separately and the model that predicted the left-out outer folds with the smallest root mean squared error (RMSE) was selected to be used in the final models (see Table S4 for full results from this analysis, Supplemental Digital Content, SDC 4, Optimal degree count threshold). To reduce the features used in the final models, a selection by filtering approach was taken within the same cross-validation folds used in the final models to keep test and train data separate and reduce overfitting. Here, features that did not correlate with the outcome at p<0.05 within each fold were removed from the final feature set. Age, biological sex and mean framewise displacement (functional connectivity only) were regressed out of the features. Three separate models were trained: (1) imaging features only, (2) Psychosocial, physical function and activity, health, and cognitive features only and (3) multimodal features containing all features. In the case of model 2, the combination of nested cross-validation and Elastic net failed to predict adherence in the left-out outer folds, potentially due to the small number of features in the model (p=6) and as such, for the non-imaging model only we ran a simple 10-fold cross-validation model using Ridge regression (L1 norm penalty only). We provide several interpretable measures of model performance (prediction on the left-out folds) based on the observed vs predicted values; Pearson’s r correlation, the squared correlation, R², RMSE, which measures the average prediction error as the average difference between the observed and predicted values and the mean absolute error (MAE) as the average absolute difference between the observed and predicted values. RMSE and MAE are related with MAE being less sensitive to outliers and the lower the value the better the model performance. To assess the significance of the prediction performance, 10000 non-parametric permutations were performed on the correlation coefficient between the predicted and observed values, consistent with the fMRI literature [44]. Here, the outcome measure is randomly assigned to different subjects and by using this new label assignment 10000 times we estimated the distribution of the test statistic. The p-value of the permutation tests was then calculated as the proportion of sampled permutations that are greater or equal to the true prediction correlation (see Figure S1, Supplemental Digital Content, SDC 5, Significance testing of the prediction performance). All statistics were performed in RStudio version 3.6.3 using “caret”, “dplyr”, “purr”, “penalized” and “pensim” packages. Figures were generated using “ggplot2”. Code used in this analysis can be found at https://github.com/tpmor-546/Adherence_pred

3. RESULTS

Table 1 presents demographic details of the participants included in this analysis. Our sample consisted of mostly female, white, and educated participants and the mean adherence was 81%.

Table 2 contains prediction metrics for each modality using Elastic net models and Ridge regression for the psychosocial/cognitive/health only model. The psychosocial and demographic model did not significantly predict adherence (r = 0.17, p = 0.056). The imaging (r = 0.24, p = 0.004) and multimodal models (r = 0.28, p = 0.001) did significantly predict adherence (Table 2 and Figure 1).

Table 2.

Model performance metrics

Model	r	R2	RMSE	MAE	P-value
Multimodal	0.28	0.08	0.11	0.09	0.001
Imaging	0.24	0.06	0.11	0.09	0.004
Psych/cog/health	0.17	0.03	0.12	0.10	0.056

Performance metrics derived from nested cross-validation where the optimal hyperparameters for Elastic net regression were tuned in an inner loop and used to predict adherence in the left out outer loop. r and R² represent the Pearson’s correlation and the squared correlation between the predicted and observed values, respectively. RMSE represents the average difference between the observed and predicted values (average prediction error) and MAE represents the absolute mean difference between the predicted and observe values. The p-value for each model is derived by comparing the correlation coefficient between the observed and predictive values to a null distribution derived from 10000 non-parametric permutations.

Figure 1.

Prediction performance (r) of each model (top row). Top row: r represents the correlation between the predicted and observed adherence values via Elastic net regression with nested cross-validation. The multimodal model numerically performed better than the unimodal models alone.

Standardized coefficients for each feature selected in each model are found in Table 3. A summary figure of the multimodal model is found in Figure 2. For the functional connectivity features, higher nodal degree strength in lateral and medial prefrontal regions, spanning both the frontoparietal and default mode networks, predicted higher adherence to the intervention. Conversely, less nodal strength in visual, temporal and somatomotor networks predicted higher adherence to the intervention. Higher cortical thickness in bi-lateral postcentral gyrus, inferior frontal gyrus of the left hemisphere and the left frontal pole predicted higher adherence whereas lower cortical thickness in the right cingulate and lower surface area in bilateral occipital gyrus and left inferior frontal gyrus predicted higher adherence. While the psychosocial and demographic model alone didn’t significantly predict adherence, those features included in the multimodal model that were predictive of adherence (given their non-zero coefficient) included self-efficacy for walking, barriers specific self-efficacy, strength self-esteem, biological sex and employment status (table 3).

Figure 2.

Summary figure depicting the data-driven features of brain structure and function in the multimodal model (best performing model). For the functional connectivity features (top left), each feature represents degree count (number of connections that each node has with every other node in the cortical parcellation). That is, red features (positive nodes) are interpreted as the higher the connectivity between that node and every other node in the cortex, the higher the adherence. Blue features (negative nodes) suggest that less connections between these nodes and the rest of the brain the higher the adherence. Mapping these regions to intrinsic resting state functional networks (17 Yeo networks via the Schaefer 300 atlas) reveals a broad pattern of positive nodes in the DMN and FPCN and negative nodes in primary information processing networks. For cortical brain structure, greater cortical thickness in bi-lateral post-central gyrus and left inferior frontal gyrus, and lower cortical surface area in bi-lateral occipital cortex and left inferior frontal gyrus were predictive of adherence.

4. DISCUSSION

In this study we demonstrated that brain-based measures of functional connectivity, cortical thickness and surface area predicted future adherence to a structured group-based exercise intervention in older adults. Using a machine learning framework that applies penalized regression with cross-validation we also replicated earlier studies indicating that aspects of self-efficacy and biological sex are predictive of adherence to exercise and that together, multimodal features provide the numerically strongest predictive value.

Currently, our model performance remains far from having utility in a clinical setting. While we replicated the finding that traditional psychosocial and demographic measures predict adherence to exercise, albeit in a model that was not significant, the ability of neuroimaging features alone to predict adherence to a structured group-based exercise intervention suggests substantial utility of these measures for future research into precision medicine and adaptive intervention approaches. Additionally, the numerical increase in model performance when combined suggests that multimodal features provide independent relevant information in the prediction of exercise adherence. Multimodal prediction models have been shown to outperform unimodal models in prior work also [42,45]. It is possible that the numerical increase in prediction performance from multimodal features is due to individual features capturing distinct aspects of complex behaviors linked to adherence to exercise, that unimodal features alone may not capture. Additionally, the use of multimodal imaging metrics likely limits the effect of scanning or preprocessing artefacts from any given modality. Notwithstanding, while the correlation between predicted and observed values were higher in the multimodal model, the absolute difference between the predicted and observed values (MAE) between the imaging model and the multimodal model was numerically negligible. Given the psychological, health and demographic model was not significant when tested against a null distribution of the test statistic, it is likely that the imaging features were providing most of the predictive signal.

It is important to highlight simple numerical comparisons between our results (i.e., R²) and those of prior research on correlates of exercise adherence should not be compared directly due to the rigor of prediction modeling compared to simple within-sample inferential statistics. Our prediction methodology is implemented with the explicit goal of improving generalizability to new unseen data [46] whereas prior research using within-sample correlational statistics are prone to biased estimates that may not generalize well to new data.

In a research setting, the application of prediction modelling could aid in the development of adaptive intervention strategies. For example, in those who can successfully engage in a theory-based intervention, positive results have been shown. However, successful engagement in long-term interventions is non-trivial and not all individuals will adhere. If one could predict these individuals from the outset, re-deployment of resources to tailor the intervention in these individuals with the goal of improving adherence and subsequent efficacy could be done. For example, our results could be leveraged in future prospective studies to test whether adaptive interventions based on pre-intervention characteristic and brain signatures (our predictions) result in better adherence and intervention efficacy. Concomitantly, our results provide several testable mechanistic pathways through which individual differences in brain structure and function can affect exercise adherence. For example, if default mode and executive control connectivity are important predictors of exercise adherence then mindfulness meditation training could be used as a primer to enhance network connectivity prior to or during an exercise intervention, as this has been shown to modulate the connectivity of these networks [47]. Future studies could test whether modulation of these measures results in the enhancement of adherence to prospectively test these predictions. Additionally, successful engagement in an exercise intervention may also be related to long-term maintenance of exercise behaviors beyond the intervention [29]. Consequently, applying this type of prediction modeling to recent large population and cohort studies (e.g., UK Biobank or Human Connectome Project) with imaging and self-report physical activity data could help one determine the proportion and characteristics of individuals who would require additional tailoring when designing future exercise interventions.

Regarding our imaging analysis, we took a whole-brain approach to the prediction of exercise. Our results however could be leveraged to validate these results in future hypothesis-driven studies. For example, nodes with high global functional connectivity predictive of adherence in our study were found in the medial and lateral pre-frontal cortex within the DMN, and FPCN networks, two related networks that contribute to internally and externally directed cognition [21] and in the implementation of executive control processes to maintain goals and inhibit distractions, respectively [48]. Similarly, cortical structure in postcentral, and inferior temporal regions have been shown to be associated with executive functions, especially in healthy aging [11]. As such, direct study of the interplay between these networks at rest or during executive function tasks that engage these networks may provide stronger predictive utility.

Numerous randomized controlled trials of exercise in aging have demonstrated that exercise can improve (with mixed effect sizes) cognitive function and positively effect brain structure and functional connectivity [49]. More recently, the hypothesized bi-directional relationship between exercise and the brain and cognition has been tested [13]. In our current study, individual differences in proxy measures of cognitive and brain reserve (functional connectivity nodes, cortical thickness, and surface area, respectively) within primary information processing networks and regions were predictive of exercise adherence. Prior exercise interventions have shown intervention-mediated increases in the functional connectivity of somatosensory networks in older adults [50]. Our results therefore potentially provide supportive evidence in favor of the hypothesis that the relationship between exercise and the brain is bi-directional [51]. Notwithstanding, our prediction models need to be tested in prospective and experimental studies to conclude a causal association. Furthermore, these predictions need to be validated in much larger and more generalizable samples.

Several limitations to this study mean that our results should be interpreted with reasonable caution. Firstly, the complexity of adherence to a 6-month exercise intervention may not be completely captured in a single variable representing the percentage attendance to the intervention settings and so additional explained variance not captured by our measures may be due to other reasons such as scheduling conflicts related to family, work or breaks due to discomfort. Second, it is of note that inferential guarantees regarding variable estimates (coefficients) in penalized regression models cannot be made. i.e., traditional p-values or confidence intervals for each estimate do not exist [52]. Therefore, estimates produced by the elastic net model are biased and should not be interpreted as the population parameter. While advances in post-selection inference methods for the Lasso have been made ([52,53] which allow for more valid confidence intervals and significance testing for Lasso estimates), such methods do not exist for Elastic Net, the use of which was important in our case given the collinearity of the predictors in our study. Third, our sample size was relatively small for machine learning. As such, testing on completely left-out subjects (test-train splits) was not feasible and so we attempted to perform the most rigorous and generalizable approach possible via nested cross-validation. Fourth, we implemented a data-driven approach which used whole-brain imaging measures as features. This approach could be complimented or enhanced in future research using task-evoked MRI or network/region specific measures to reduce the dimensionality of the input data. Fourth, given the difficulty and costs associated with this type of research and the vast number of measures that were available in this unique dataset, replication of this prediction in a completely independent validation dataset is unavailable at this moment. Fifth, our participants were homogenous. Several factors have recently been highlighted as a reason why global physical inactivity rates continue to be low [3] with one being that most of the research on understanding physical activity behaviors has occurred in high-income countries and so more work in diverse populations and low-income countries is needed to fully generalize this type of prediction. Lastly, our battery of neuropsychological tasks included executive function tests that measured constructs such as abstract, inductive and visuospatial reasoning rather than other executive tasks like inhibitory control that may be more related to exercise adherence [12].

CONCLUSIONS

Our results showed that the combination of psychosocial, cognitive, and demographic, and multimodal imaging metrics can predict adherence to an exercise intervention in older adults and provide independent relevant predictive value. Prospective testing of these predictions and their validation will allow researchers and eventually clinicians to leverage them through personalized medicine approaches.

Supplementary Material

Supplemental Data File (.doc, .tif, pdf, etc.)_1

Table S1: Demographics stratified by group assignment

Supplemental Data File (.doc, .tif, pdf, etc.)_2

Table S2: Prediction models in the walking and walking + group only

Supplemental Data File (.doc, .tif, pdf, etc.)_3

Table S3: Complete list of behavioral variables

Supplemental Data File (.doc, .tif, pdf, etc.)_4

Table S4: Optimal degree count threshold

Supplemental Data File (.doc, .tif, pdf, etc.)_5

Figure S1: Significance testing of the prediction performance