Social determinants of health and brain connectivity predict physical activity behavior change after new cardiovascular diagnosis

Abstract

Physical activity is essential for preventing cognitive decline, stroke and dementia in older adults. A new cardiovascular diagnosis offers a critical window for positive lifestyle changes. However, sustaining physical activity behavior change remains challenging and the underlying mechanisms are poorly understood. To identify the neural, behavioral, and contextual determinants of long-term physical activity change after a new cardiovascular diagnosis, we applied support vector machine learning to predict 4-year trajectories of both self-reported and accelerometer-derived moderate-to-vigorous physical activity in 295 cognitively unimpaired older adults from the UK Biobank, testing three models that incorporated baseline: (i) demographic, cognitive, and contextual factors, (ii) baseline resting-state functional connectivity alone, and (iii) combined multimodal features across all predictors. The combined multimodal model had the highest predictive power (r = 0.28, P = 0.001). Key predictors included greenspace access, social support, executive function and between-network functional connectivity within the default mode, and frontoparietal control networks. These findings underscore the importance of behavioral factors and social determinants of health and uncover neural mechanisms that may support lifestyle modifications. In addition to furthering our understanding of the mechanisms underlying successful physical activity behavior change, these findings help to guide the design of interventions and health policy with the ultimate goal of preventing cardiovascular disease burden and late-life cognitive decline.

Significance Statement.

Cardiovascular diseases significantly elevate the risk of dementia and stroke. Physical activity is highly effective in lowering dementia risk among individuals with cardiovascular disease, yet physical inactivity remains widespread. This study represents the largest and most comprehensive assessment of brain, behavioral, and contextual factors predicting successful long-term physical activity behavior change following a cardiovascular diagnosis in aging. Our findings suggest that a multimodal brain-behavior fingerprint may serve as a marker of an individual's propensity for physical activity adherence and help to identify those who are most likely to change their behavior. We also emphasize the importance of structural factors, including greenspace and social support, in promoting physical activity—evidence that are critical to guide policy decision-making and urban planning.

Introduction

Cardiovascular diseases substantially elevate the risk of dementia and stroke due to shared pathophysiological mechanisms across the heart–brain axis (1). Indeed, mixed dementia, comprised of combined vascular and Alzheimer pathological changes, is the most prevalent etiology of dementia in older age (2). With the global dementia burden projected to rise to 132 million by 2050 (3), there is an urgent need for targeted strategies to mitigate the vascular contributions to late-life cognitive decline. Physical activity is highly effective in lowering dementia risk and all-cause mortality among individuals with cardiovascular disease (4). Thus, targeting physical activity engagement as a strategy for dementia prevention following a cardiovascular diagnosis is essential (5, 6).

Despite the well-established benefits of physical activity, physical inactivity remains prevalent, with approximately 27.5% of the global population not meeting recommended activity levels (7). The prevalence is high among older adults and has increased since the COVID-19 pandemic, especially among older adults with chronic conditions (8). Moreover, motivating and maintaining long-term behavior change is difficult. Observational studies show that only 4.3% of individuals adopt lifestyle modifications within 6 months of a cardiovascular event (9). Prior research on behavioral therapy also suggest that most individuals struggle to maintain these changes, with long-term adherence rates often declining significantly over time. For example, premature dropout from health behavior change interventions was reported in 30–60% of participants (10). Understanding why and when individuals initiate physical activity engagement is crucial for designing effective interventions.

To move towards a precision medicine approach to behavior change, it is important to go beyond group-level statistical approaches to identify individual differences and contextual factors at the level of the individual (11, 12). Prior behavioral research applying group-level statistics has highlighted factors such as self-efficacy (13), self-regulation (14), and biological sex, where males generally show higher adherence rates than females (15) in influencing physical activity engagement. Psychological factors, including depression, fatigue, and executive function have also previously been shown to influence adherence (16). Furthermore, social and structural determinants of health (SSDoH), which refers to the environmental conditions in which individuals are born, live, learn, work, play, and age have a cumulative impact on physical, mental, and brain health (17). Factors such as access to greenspace and neighborhood walkability (18), social support (19), socioeconomic status (20) are strongly associated with physical activity levels. Critically however, whether these factors also support physical activity behavior change remains unknown. These determinants may not only shape physical activity behavior but also act as upstream contributors to disparities in health outcomes, including incidence of dementias (21). Further, neuroimaging provides insights into individual differences in brain organization and highlights neurodiversity, that is, how brain functions vary across individuals based on multilevel nonmodifiable factors (e.g. genetics and biological sex) and differential life exposures to SSDoH (11).

Resting-state functional MRI offers a powerful lens into the intrinsic architecture of the brain, providing mechanistic insight into the cognitive and neural processes that support complex behaviors such as sustained physical activity engagement. Prior work has demonstrated that individual variability in network connectivity, particularly within and between default mode, salience, and frontoparietal control networks, is associated with self-regulation, future planning, and reward sensitivity, all of which are critical components of behavior change (22, 23). Moreover, our own and others' work has shown that these large-scale networks are sensitive to the effects of aging and vascular risk, but also exhibit plasticity in response to lifestyle interventions, suggesting they may be both predictive and modifiable (24, 25).

Beyond prediction, resting-state functional MRI holds clinical relevance by identifying neurobiological targets for tailoring interventions. First, such systems could be directly targeted through neuromodulation (e.g. transcranial magnetic stimulation or transcranial direct current stimulation) (26). Second, RSFC profiles can serve as neural markers for personalizing interventions. Prior studies from our lab and others have leveraged resting-state functional MRI to forecast individual responses to behavioral and pharmacological treatments across a range of conditions, including smoking cessation (25), adherence to cognitive training (27), and treatment response in social anxiety disorder (28), underscoring its potential translational utility as a neural marker with relevance to personalized intervention design. Lastly, these neural systems can also be indirectly engaged via interventions that engage specific cognitive processes (e.g. goal-setting, mindfulness training) (29). For example, individuals with differential connectivity in frontoparietal regions may benefit from strategies that offload executive demands, such as implementation intentions or environmental restructuring.

Thus, incorporating resting-state functional MRI derived markers advances both mechanistic understanding and intervention design, two essential pillars for developing precision behavior change approaches in aging populations at risk of cognitive and cardiovascular decline.

Predictive modeling based on resting-state functional connectivity (RSFC) leverages the most relevant features of functional connectivity to predict behavioral outcomes. By mapping the brain's intricate connections and integrating them with data on individual behaviors, it offers a window into the neural basis of highly complex phenomena. Indeed, prior research supports the utility of functional brain connectivity for behavioral prediction (30, 31), and shows that it can outperform the predictive power of structural features for lifestyle adherence (24). This is in line with the Stern theoretical framework of cognitive reserve (the ability to maintain function in the face of age- and disease-related brain changes) that suggests functional measures might best capture the “neural implementation” of cognitive reserve (32).

To better understand the drivers of physical activity behavior, change among older adults who stand most to benefit, the current study adopts a precision medicine framework combined with a whole-brain machine learning approach. Specifically, we examine the roles of sociodemographic factors (e.g. age, sex, and socioeconomic status), behavioral characteristics (e.g. retirement status, general health, pain, and depression), cognitive function (e.g. attention and executive function), social factors (e.g. networks and support), environmental context (e.g. access to green spaces), and baseline RSFC on future physical activity behavior change after physically inactive older individuals receive a new cardiovascular diagnosis. This comprehensive approach is designed to uncover tangible targets for future interventions, including public policy changes, tailored to individual needs for those at a heightened risk of cognitive decline. By employing a rigorous data-driven machine learning approach, the current study aims to uncover the neurobehavioral mechanisms driving successful physical activity at the individual level (11).

Methods

Participants and study design

295 (mean age = 63.13 years ± 7.5, 188 women) cognitively unimpaired and physically inactive older adults from the UK Biobank, a large-scale population-based longitudinal cohort were included in this study. Inclusion criteria were: (i) cognitively unimpaired at enrollment; (ii) reported a new cardiovascular diagnosis (i.e. hypertension, type II diabetes, dyslipidemia, cardiac angina, or myocardial infarction) between baseline (T1; 2014) and follow-up over 4 years later (T2; 2019) (mean duration 4.2 years, SD 1.1); (iii) did not meet the World Health Organization recommendation of 150 min/week of moderate-to-vigorous physical activity (MVPA) at baseline (3); and 4) age ≥ 60. Unimpaired cognition was defined as follows: performance scores on each cognitive test were converted into a percentile rank, and the raw score corresponding to the 5^th percentile (or 95th, on tests where higher scores represented worse performance) was identified as the cut-off for impairment (33). The brain imaging visit (Instance 2 of the UK Biobank) was considered the baseline timepoint, and the first repeat imaging visit (Instance 3 of the UK Biobank) was considered the follow-up timepoint. MRI data were obtained at baseline, before participants had received a new cardiovascular diagnosis, and MVPA self-reported and accelerometry-derived data and cognitive indices were obtained for the two timepoints: at baseline and in follow-up after 4 years (ranging from 8 months to 4.8 years). After excluding four participants for poor quality brain imaging data, the final sample was 295. Participant demographic characteristics are summarized in Table 1. An illustration of the study timeline is shown in Fig. 1.

Table 1.

Participant baseline demographic information for the UK Biobank sample.

Demographic factor (n = 295)	Mean ± SD
Age at baseline (years)	63.13 years ± 7.5
Female sex (count and %)	188 women (63.72%)
Household income (£, most frequent range)	18,000 to 30,999 (61%)
Education (years)	15.4 ± 3.2
Townsend deprivation index score	−1.3 ± 3.2
Accelerometry-derived MVPA at baseline (min/week)	101.62 ± 4.5
Accelerometry-derived MVPA at follow-up (min/week)	109.10 ± 6.8
Frequency of friends and family visits at baseline	4.2 ± 1.03
Able to confide at baseline	3.8 ± 1.67
Greenspace proximity at baseline (%)	34.1 ± 20.59
Coastal proximity at baseline (%)	0.89 ± 2.88
Depression score at baseline	2.1 ± 1.02
Anxiety score at baseline	1.8 ± 0.88
Number retired at baseline (count and %)	111 retired (37.64%)

SD, standard deviation; MVPA, moderate to vigorous physical activity.

Fig. 1.

Study timeline: older participants received a new cardiovascular diagnosis (i.e. hypertension, type II diabetes, dyslipidemia, cardiac angina, or myocardial infarction) between the baseline and follow-up periods (mean 4.2 years), with no cardiovascular diagnoses reported prior to baseline. Assessments included self-reported and accelerometer-derived MVPA, cognitive function, neurobehavioral factors (such as depression, anxiety, general or pain), and SSDoH (including social support, retirement status, and greenspace access). Resting-state functional connectivity (RSFC) MRI was assessed at baseline.

Instruments and assessments

Participants completed a comprehensive battery of psychosocial, behavioral, cognitive, and environmental assessments at both baseline and follow-up (Fig. 1). These assessments are briefly outlined below and additional detail is provided in Supplementary material.

Townsend deprivation index

The Townsend Deprivation Score is an area-based score of social deprivation aggregated from percentage of unemployment rate, noncar ownership rate, nonhome ownership rate and household overcrowding (proportion of households with more people than rooms). This indicator was determined immediately prior to the participant joining the Biobank and was based on data from the preceding national census (34). The Townsend Deprivation Index is a composite, standardized score with higher positive values indicating greater socioeconomic deprivation and lower (negative) values indicating less deprivation. Each participant was assigned a score corresponding to their postal code area.

Physical activity questionnaires

Successful future physical activity engagement, the primary behavioral outcome of interest, was defined as the difference between the overall MVPA in minutes per week measured at follow-up compared to the overall MVPA in minutes per week measured at baseline. This change in MVPA was assessed using the Lifetime Total Physical Activity Questionnaire (35), which captures self-reported MVPA by recording the frequency and duration of each physical activity type performed weekly. The total time spent on moderate and vigorous activities was then calculated to derive the overall MVPA in minutes per week. This total score served as an indicator of each individual's physical activity engagement. The scale was administered at both baseline and follow-up timepoints to assess changes over time.

Leisure-time physical activity was also measured through items capturing activities such as walking for pleasure, light, and heavy do-it-yourself tasks (e.g. pruning, watering the lawn, carpentry, digging, and weeding), and recreational activities (e.g. swimming, cycling, and bowling). The total time spent on these activities was then calculated to derive the overall leisure-time physical activity in minutes per week.

Occupational physical activity was assessed with questions adapted from the UK Biobank, including “Does your work involve heavy manual or physical work?” and “Does your work involve walking or standing for most of the time?” These questions helped capture physical activity levels related to participants' work environments. Scores ranged from 1 (never/rarely) to 4 (always) and were treated as a continuous measure.

Accelerometry-based measurement of physical activity

A subset of participants who provided valid email addresses were randomly selected and invited to wear a wrist-mounted triaxial accelerometer (AX3; Axivity Ltd) within the UKBB, and continuously for 7 days to objectively assess physical activity levels. Physical activity volume was quantified as the mean vector magnitude (expressed in milligravity units), a measure that has been externally validated against the gold-standard doubly labeled water method (36). MVPA was estimated using validated machine learning models trained on data from wearable cameras and detailed time-use diaries.

Initial data processing and quality control were conducted centrally by the UK Biobank research team (37). Participants with fewer than three valid days (≥72 h) of accelerometer wear time, or those lacking data coverage in each 1-h segment of the 24-h cycle, were excluded. Following the removal of 10 individuals with insufficient data, our final analytic sample included 285 participants. Raw accelerometer files (.cwa format) were processed using an open-source pipeline available at https://github.com/activityMonitoring/biobankAccelerometerAnalysis. Data were resampled into 5-s epochs to preserve temporal resolution and retain the average vector magnitude for each epoch. An empirical cumulative distribution function was generated across all epochs to characterize the full spectrum of physical activity intensity for each participant (36).

Nonwear periods were defined as any continuous 60-min interval during which all three axes exhibited a SD below 13.0 mg (38). These segments were imputed using a time-of-day-matched averaging method, wherein values were replaced by the average vector magnitude and intensity distribution observed at the same time across different days for the same participant. This imputation strategy mitigates diurnal bias that could result from systematically lower wear adherence during specific periods, such as sleep, thereby avoiding inflation of activity estimates due to selective wear-time averaging. Finally, MVPA was derived by averaging all valid and imputed values across the measurement period, providing a robust estimate of habitual physical activity for each participant.

Cognitive assessments

A computerized cognitive battery was administered using a touchscreen tablet. The tests were specifically developed for the UK Biobank and have been validated (33), while sharing features with established cognitive assessments. The battery included the following tasks: Reaction time, Numeric memory, Prospective Memory, Fluid intelligence, Matrix pattern completion, Tower rearranging, and Trail making. A detailed description of these tasks can be found in Supplementary material Appendix A. Additionally, a composite cognition score was calculated at both baseline and follow-up by standardizing individual task scores and aggregating them into a single summary index.

Social support

The measures available in the UK Biobank for social support come from the items “How often do you visit friends or family or have them visit you?” and “How often are you able to confide in someone close to you?” Participants rated each item on a Likert scale from 0 (never or almost never) to 6 (almost daily). For the frequency of visits, the categories “never or almost never” and “no friends or family outside the household” were combined into a single category, “never.” This adjustment was made because these responses were similar, and there were only a few participants with no friends or family outside the household (n = 16). Scores ranged from 0 to 6 and were treated as a continuous measure. Loneliness was also assessed using the item “Do you often feel lonely?.” Responses were recorded as yes (1) or no (0).

Greenspace and coastal proximity assessment

Environmental indicators included in this study were the proportion of green space and water within 300 m of residential addresses, using the 2005 Generalised Land Use Database for England and Centre for Ecology and Hydrology 2007 Land Cover Map data for Great Britain (39). The buffer size of 300 m was decided based on relevant health evidence and public policy on both density and accessibility (40). Coastal proximity was estimated using Euclidean distance raster.

Psychosocial and mental health factors

Psychosocial factors were assessed through self-reported experiences, including depression, anxiety, general pain, and lifestyle factors such as retirement status. Depression was evaluated using two items: “Feeling down, depressed, or hopeless” and “Little interest or pleasure in doing things.” Participants rated their experiences on a four-point scale, ranging from 0 (not at all) to 4 (nearly every day). Anxiety was assessed similarly, with two items: “Feeling nervous, anxious, or on edge” and “Not being able to stop or control worrying.” General pain was measured using a single item: “Have you had pains all over your body for more than 3 months?” Responses were recorded as yes (1) or no (0). Participants also rated their overall health perception on a scale from 1 (excellent) to 4 (poor) and were treated as a continuous measure. Additionally, participants indicated their retirement status with a simple yes (1) or no (0) response. These assessments were conducted at both baseline and follow-up timepoints.

MRI data acquisition

Details of image acquisition and processing are available in the UK Biobank Protocol (http://biobank.ctsu.ox.ac.uk/crystal/refer.cgi?id=2367), and Brain Imaging Documentation (http://biobank.ctsu.ox.ac.uk/crystal/refer.cgi?id=1977). Briefly, all brain MRI data were acquired on a Siemens Skyra 3T scanner with a standard Siemens 32-channel RF receiver head coil, using the following parameters: repetition time (TR) = 2000ms; inversion time (TI) = 800 ms; R = 2; field of view (FOV) = 208 × 256 × 256 mm; voxel size = 1 × 1 × 1 mm. For resting-state fMRI scans, two consecutive functional T2*-weighted runs were collected with eyes closed using a blood oxygen level dependent (BOLD) sensitive, single-shot echo planar imaging (EPI) sequence with the following parameters: TR = 735 ms; time to echo (TE) = 39 ms; flip angle = 52°; FOV 88 × 88 × 64 matrix; resolution = 2.4 × 2.4 × 2.4 mm; 490 volumes; and acquisition time = 6 min/run.

Resting-state functional MRI data preprocessing

Preprocessing of raw functional images from the UK Biobank was done using the fMRIprep pipeline (version 20.2.4) (41). For each of the BOLD runs per participant, the following preprocessing was performed: First, the T1w reference was skull-stripped using a Nipype implementation of the antsBrainExtraction.sh (ANTs) tool. A B0-nonuniformity map (or fieldmap) was estimated based on a phase-difference map calculated with a dual-echo gradient-recall echo sequence, which was then co-registered to the target EPI reference run and converted to a displacements field map. A distortion-corrected BOLD EPI reference image was constructed and registered to the T1-weighted reference using a boundary-based approach (using bbregister, Freesurfer). Rigid-body head-motion parameters with respect to the BOLD EPI reference were estimated (using mcflirt, FSL 5.0.9) (42) before spatiotemporal filtering was performed. BOLD runs belonging to the single band acquisition sessions were slice-time corrected (using 3dTshift, AFNI 20160207). The BOLD time series were resampled into their original, native space by applying a single, composite transform to correct for scan-to-scan head motion and susceptibility distortions. Functional scans were spatially smoothed using a 6 mm full width at half maximum Gaussian smoothing kernel.

Additional preprocessing steps were undertaken to remove physiological, subject motion, and outlier-related artifacts, which were implemented using the nilearn package. Non-neuronal sources of noise from white matter and cerebrospinal fluid (CSF) were estimated and removed using the anatomical CompCor method (43) to allow for valid identification of correlated and anticorrelated networks (44). Temporal band-pass filtering (0.008–0.09 Hz) was then applied. Additionally, scan-to-scan mean head motion (framewise displacement) was used as a covariate of noninterest in all second-level analyses (mean head motion = 0.2 mm, SD = 0.1 mm). Head motion is a known important potential confound as it produces systematic and spurious patterns in connectivity and is accentuated in Alzheimer's disease and cognitively typical aging populations (45). Critically, we did not identify a relationship between the mean head motion parameter and the primary behavioral variable of interest, physical activity change (all P > 0.05). The framewise displacement timeseries was determined by calculating the maximum shift in the position of six control points situated at the center of a bounding box around the brain, computed independently for each scan. Four participants were removed from the UK Biobank sample final analysis for having >30 scan volumes flagged, leading to the final sample size of 295 participants. This cut-off was determined based on preserving at least 5 min of scanning time (46).

Resting-state functional connectivity feature extraction

RSFC features were derived from the preprocessed BOLD time series using the Schaefer 400-node cortical parcellation atlas with 17-network assignments, which aligns with the Yeo large-scale functional networks (47). We computed pairwise Pearson correlation coefficients between the average BOLD time series of each node pair, resulting in 79,800 unique edgewise connectivity values per participant. These correlation matrices were Fisher z-transformed to improve normality. From these, a subset of 400 connectivity features was selected based on prior literature indicating their relevance to cognition and physical activity, and to reduce model complexity. These 400 features primarily consisted of within- and between-network connectivity metrics aggregated across predefined network groupings (e.g. default mode, frontoparietal, and dorsal attention).

Data analysis

To establish the relevance of investigating determinants of physical activity, we first examined the relationship between physical activity and cognitive outcomes. Specifically, we tested whether baseline cognitive function predicted future physical activity behavior and whether increases in physical activity at follow-up were linked to cognitive gains. We also conducted moderation analyses to test whether changes in MVPA, both self-reported and objectively measured, interacted with time and influenced cognitive outcomes. These analyses provide critical justification for the study: if physical activity contributes to better cognitive health, it becomes essential to understand the factors that support sustained engagement in physical activity. Building on this foundation, we then evaluated whether social and structural health determinants, behavioral, and neural measures at baseline predicted successful future engagement in physical activity.

To predict future successful physical activity (MVPA) as indexed by both self-reported and accelerometer-derived changes in behavior as a continuous measure following a new cardiovascular diagnosis, we used the support vector machine (SVM) regression algorithm from the scikit-learn (v0.21.3) library, utilizing the pydra-ml (v0.3.1) toolbox. Three separate models were trained: (i) combined demographic, cognitive and contextual features only (ii) neuroimaging features only, and (iii) multimodal model combining all demographic, cognitive, contextual, and neuroimaging features. RSFC was assessed at baseline only, prior to any cardiovascular diagnosis, thereby reducing potential confounding effects related to blood flow alterations (48). Contextual features encompass many factors influencing responses to interventions and overall clinical outcomes, including but not limited to the personal characteristics, and SSDoH (15, 17). This multilevel, complexly interacting framework is essential for understanding physical activity behavior change in individuals with cardiovascular disease and optimizing the effectiveness of preventive strategies and interventions.

The SVM regression works by placing constraints to ensure only a small number of observations (support vectors) are used. SVM regression works with the goal of constructing a regression line that fits the data within some chosen level of error. We used the default parameters, which include the radial basis function kernel to capture nonlinearities in the data. To assess the robustness of our findings, we repeated our analysis using additional machine learning algorithms of increasing complexity, defined by the computational resources required for model simulation. Specifically, we examined linear regression, random forest, and multilayer perceptron algorithms, using default parameters unless otherwise specified. Further details on these algorithms are available in the Supplementary material Appendix B.

We investigated model performance using four features selection strategies for all the prediction models we tested (Linear Regression, SVM regression, Random Forest regression, and multilayer perceptron): (i) using all features, (ii) removing redundant neuroimaging features, (iii) selecting only the top 20 features based on SHAP (SHapley Additive exPlanations) analysis, and (iv) excluding the top 20 features to assess their necessity for predictive performance. To generate independent test and train data splits, we used a bootstrapped group shuffle split sampling scheme. For each iteration of bootstrapping, a random selection of 20% of the participants, balanced between the two groups, was designated as the held-out test set. The remaining 80% of participants were used for training. This process was repeated 50 times, fitting and testing the four classifiers for each test/train split. We used the default of 50 bootstrapping splits from pydra-ml toolbox. We provide several interpretable measures of model performance based on the observed versus predicted values; Pearson's r correlation, the squared correlation, R², root mean squared error (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. The P-value for each model is derived by comparing the correlation coefficient between the observed and predicted values to a null distribution derived from 1,000 nonparametric permutations.

To interpret model predictions, we employed Kernel SHAP to quantify the contribution of baseline RSFC features in predicting successful engagement in physical activity (49). We computed the average absolute SHAP values across all predictions, weighted by the model's median performance, and calculated mean SHAP values across splits for each model. This entire pipeline, encompassing machine learning models, bootstrapping, and SHAP analysis, was implemented using pydra-ml toolbox.

A full list of the demographic, cognitive and contextual input variables used in the prediction models is available in Table S1. Demographic variables including age, sex, years of education, household income, and socioeconomic status (as measured through Townsend deprivation index) were included as covariates of noninterest (50). Household income data were obtained from UK Biobank field 738, which categorizes average total household income before tax into five groups: <£18,000, £18,000–30,999, £31,000–51,999, £52,000–100,000, and >£100,000. Models were also run with demographic variables treated as predictors of interest rather than as covariates of noninterest.

Reducing collinearity using independence factor to enhance model interpretability

Collinearity among features can significantly affect model generation and interpretation, particularly in resting-state functional MRI analyses. To address this, we employed the Independence Factor method (51), which iteratively removes features with strong dependence above a set threshold, ensuring a consistent set of features across models. Using distance correlation, which accommodates nonmonotonic relationships, we systematically increased the threshold to eliminate redundant features while preserving model performance within a narrow margin. Importantly, reducing distance correlation enhances statistical independence among features, thereby improving model interpretability. We applied thresholds ranging from 1.0 (keeping all features) to 0.2 (removing features with distance correlation above 0.2). Our goal was to identify a feature set that maintained model performance within three percentage points of using all features, resulting in a more parsimonious and interpretable model without compromising accuracy, essential for clinical applicability.

Performance using most important and least important features

To address the question of why certain features are important, we evaluated model performance under two scenarios: one using only the top 20 features and another excluding these features. This method mitigates the common pitfall in brain-behavior prediction analyses, where the significance of the top features may not reflect their true impact on model performance. By comparing performance metrics in both scenarios, we can gain a more nuanced understanding of the highlighted features' contributions and derive mechanistic insights into the neural correlates of successful behavior change.

Results

Behavioral results

Following a new cardiovascular diagnosis, participants demonstrated a significant average increase in physical activity engagement of 7.52 min/week ± 1.18, reflecting a 9.16% increase in accelerometer-derived MVPA (r = 0.38. P < 0.01) across the 4-year follow-up period. A positive trend was observed between higher baseline self-reported MVPA and change in MVPA at follow-up among inactive older adults (r = 0.51, P = 0.12). A supplementary Pearson correlation confirmed modest but significant convergence between self-reported and accelerometer-derived MVPA (r = 0.21, P = 0.046).

Medication use was measured at follow-up, with 141 individuals taking cholesterol-lowering medication, 162 taking blood pressure medication, and 85 using both. There were no significant associations were identified between medication use (cholesterol-lowering or blood pressure) and either baseline MVPA (self-reported or accelerometry-derived) or change in MVPA following a new cardiovascular diagnosis. We observed a nonstatistically significant trends towards cognitive decline at the group level from baseline to follow-up (all P > 0.05).

The distribution of cardiovascular conditions was as follows: 183 individuals with hypertension, 20 with diabetes, 161 with high cholesterol, and 10 with cardiac angina or myocardial infarction. Inventories used to measure physical activity, psychosocial, cognitive, and environmental factors in the current study are briefly described below, with further detail in Table S1.

Association between physical activity and cognitive function

While the decline in composite cognitive performance from baseline to follow-up did not reach statistical significance (P = 0.635), we observed a nonsignificant trend toward group-level cognitive decline. This pattern was apparent in the overall composite score but was not driven by changes within any individual cognitive domain.

We tested whether changes in self-reported MVPA moderated the relationship between time (baseline = 0, follow-up = 1) and cognition. A model including the interaction term (time × MVPA change) predicted working memory scores, measured via the Digit Span task: R² = 0.038, F(3, 291) = 3.87, P = 0.01. The interaction term improved model fit, explaining an additional 2.3% of the variance, ΔR² = 0.023, F(1, 291) = 5.47, P = 0.02, qFDR corrected (Fig. S2). These results indicate that individuals who reported greater increases in self-reported MVPA at follow-up demonstrated relatively improved working memory over time, with results corrected for multiple comparisons using the Benjamini–Hochberg false discovery rate (qFDR). No other cognitive domains showed significant moderation effects using self-reported MVPA.

To strengthen our findings, we also tested objective measurement of MVPA, and examined the above relationships with accelerometry-derived MVPA. For composite cognitive performance, the model with main effects of time and MVPA change explained a significant portion of variance, R² = 0.024, F(2, 290) = 3.56, P = 0.03, qFDR corrected (Fig. S3). The addition of the interaction term (Time × MVPA change) accounted for a further 1.9% of variance, ΔR² = 0.01, F(1, 289) = 4.16, P = 0.043, yielding a total of R² = 0.043, F(3, 289) = 4.32, P = 0.005, qFDR corrected. This interaction indicates that increases in objectively measured MVPA were associated with more favorable changes in cognitive performance over time.

To identify which domains were driving this effect, we conducted follow-up moderation models for individual cognitive outcomes. For the Tower Rearranging task (planning and executive function), the Time × MVPA Change interaction was significant (B = 0.25, standard error [SE] = 0.11, β = 0.16, t = 2.30, P = 0.022, qFDR corrected), suggesting that physical activity improvements were associated with gains in executive function. For Fluid Intelligence, the interaction term was also significant (B = 0.18, SE = 0.09, β = 0.13, t = 2.08, P = 0.038, qFDR corrected), indicating that increases in MVPA corresponded with enhanced reasoning and problem-solving skills.

Prediction modeling results

Prediction of future change in both subjective and objectively measured physical activity (MVPA as a continuous variable) among 295 cognitively unimpaired older adults, was conducted separately across three support vector machine (SVM) learning models with inputs that included baseline neuroimaging, behavioral, or combined features as predictors: (model 1) demographic, cognitive, and contextual features, (model 2) RSFC MRI inputs, and (model 3) a multimodal model integrating all behavioral and neural features.

Prediction of objectively measured physical activity change

We evaluated whether changes in objective accelerometer-derived MVPA could be predicted using behavioral, cognitive, contextual, and neuroimaging features. The behavioral–contextual model alone (model 1) did not significantly predict MVPA change (r = 0.17, P = 0.056). In contrast, both the neuroimaging model (model 2: r = 0.31, P = 0.002, q_FDR-corrected) and the multimodal model combining all features (model 3: r = 0.38, P = 0.0008, qFDR-corrected, q_FDR-corrected) significantly predicted objective MVPA change, with results corrected for multiple comparisons using the Benjamini–Hochberg false discovery rate (qFDR) (Table 2). SVM models consistently outperformed other machine learning algorithms, including linear regression, random forest, and multilayer perceptron (performance metrics for the other algorithms described in Table S2).

Table 2.

Performance metrics from support vector machine (SVM) regression models predicting objectively measured MVPA.

Model	R	R 2	RMSE	MAE	P-value
Behavioral and SSDoH	0.17	0.02	0.13	0.11	0.056
Neuroimaging	0.31	0.07	0.11	0.09	0.002
Multimodal	0.38	0.08	0.11	0.09	0.0008

R and R² represent the Pearson 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). MAE represents the absolute mean difference between the predicted and observed values. SSDoH, social and structural determinants of health.

A predictive model that generalizes to different settings has greater clinical utility than a model that only works under specific conditions. The SVM model demonstrated robust performance across all scenarios (Table S2). Given the high dimensionality of resting-state fMRI data, Independence Factor Analysis was applied to neuroimaging features, resulting in an optimal subset of 262 features for subsequent analyses. After removing highly dependent features, based on distance correlation, from the original 400 neuroimaging features, the final model included 262 neuroimaging features and 15 demographic, cognitive, and contextual features, for a total of 277 features.

Mean SHAP values illustrating feature importance across the three models are summarized in Table S3. Significant behavioral predictors from the highest-performing model, the multimodal model, included neighborhood greenspace exposure, social support, higher baseline scores in fluid intelligence and executive function (Tower Rearranging task), and baseline MVPA (P = 0.0033, q_FDR-corrected).

Key RSFC predictors of future MVPA were primarily left-lateralized and spanned multiple large-scale networks: the default mode network (DMN; 7 nodes), frontoparietal control network (FPN; 8 nodes), dorsal attention network (DAN; 6 nodes), and salience/ventral attention network (VAN; 6 nodes). Increased positive connectivity between the FPN and DMN, particularly between the right superior frontal gyrus and both the ventromedial prefrontal cortex and precuneus, was associated with greater MVPA gains. Similarly, enhanced connectivity between the FPN and DAN (specifically between the right superior frontal gyrus and bilateral parietal cortex), and stronger within-network FPN connectivity, predicted greater physical activity. Notably, stronger anticorrelations between the VAN and both DMN and FPN were also predictive of increased accelerometer-derived MVPA (Fig. 2).

Fig. 2.

Baseline brain functional connectivity features predict future increases in objectively measured physical activity following a new cardiovascular diagnosis in aging. a) Neuroanatomical depiction of significant features and their corresponding importance values across the objective MVPA behavior change multimodal model: node size (spheres) depicts the frequency of that brain region among predictive features, while edge thickness (line connecting two nodes) represents the weight or importance of a predictive RSFC feature. Purple signifies positive RSFC whereas grey signifies negative RSFC associated with enhanced physical activity at follow-up compared to baseline. b) The summary of frequency and distribution of predictive nodes grouped by location within canonical neural networks (i.e. Yeo 7 networks).

Prediction of subjective physical activity change

We applied the same machine learning pipeline to predict changes in subjectively measured MVPA. Independence Factor Analysis was used to reduce high-dimensional RSFC data to an optimal subset of 250 features. Combined with 19 behavioral and demographic features, the final feature set included 269 predictors. Both the RSFC-only model (r = 0.25, P = 0.004, q_FDR-corrected) and the multimodal model (r = 0.28, P = 0.001, q_FDR-corrected) significantly predicted MVPA change, outperforming alternative models including linear regression, random forest, and multilayer perceptron classifiers. In the highest-performing multimodal model, neighborhood greenspace percentage, social support (i.e. frequency of visits from friends and family), retirement status, and occupational physical activity showed a significantly positive association with MVPA change, indicating that higher greenspace exposure, more frequent friend and family visits, not being retired, and greater occupational physical activity predicted greater improvements in MVPA (P < 0.05, FDR-corrected). For cognitive features, improved higher baseline executive function (the Tower Rearranging task) emerged as a significant predictor of future increase in MVPA while no other behavioral, cognitive, or contextual features showed significant prediction effects.

Figure S1 highlights the most significant baseline RSFC MRI features from the highest-performing multimodal prediction model for subjective MVPA. These features were primarily located within the left hemisphere and spanned multiple large-scale networks, with critical nodes in the DMN (e.g. temporal lobe and medial prefrontal cortex), FPN (e.g. lateral prefrontal cortex), and VAN (e.g. frontal operculum). Enhanced RSFC within the DMN was associated with increased physical activity at follow-up. Moreover, increased MVPA was associated with greater positive RSFC between FPN and the DMN. Mean SHAP values illustrating feature importance across the three models for subjectively measured MVPA are summarized in Table S4.

Across both objectively and subjectively assessed MVPA models, several predictors consistently emerged as significant: higher neighborhood greenspace exposure, stronger social support, executive function, and greater baseline RSFC between the DMN and FPN.

Discussion

Our findings demonstrate that increases in both self-reported and objectively measured MVPA following a new cardiovascular diagnosis are associated with improvements in cognitive performance in older adults. We observed an overall trend towards decline in composite cognitive scores from baseline to follow-up, consistent with well-documented patterns of age-related cognitive decline (52). Participants who exhibited greater increases in objectively measured MVPA experienced attenuated decline in their composite cognitive score and, notably, showed gains in executive function and fluid intelligence.

Recognizing the role of physical activity in cognitive preservation, we next sought to identify factors that support sustained behavioral change. Our multilevel analyses revealed that environmental and social determinants, particularly access to greenspace and perceived social support were robust predictors of long-term MVPA maintenance. These results suggest that motivational factors alone may be insufficient; instead, structural and contextual factors play a critical role in enabling and maintaining active lifestyles. Importantly, a multimodal predictive model integrating behavioral, contextual, and neuroimaging data offered the highest predictive accuracy for long-term MVPA engagement. RSFC analyses revealed that future sustained MVPA increases were associated with enhanced between-network connectivity between the DMN and the FPN at baseline, implicated in self-regulation, future planning, and goal-directed behavior (23). Specifically, increased positive coupling between the right superior frontal gyrus (within FPN) and both the ventromedial prefrontal cortex and precuneus (within DMN) was linked to greater future MVPA gains. Together, these findings underscore the importance of integrative, systems-level approaches to behavior change. Interventions aiming to promote physical activity in older adults following a cardiovascular event should address both individual-level neurocognitive mechanisms and broader social–environmental factors.

The critical windows theory suggests that successful behavior change may be facilitated by an external threat from a major life event or circumstance (e.g. receiving a diagnosis of a new chronic illness such as a cardiovascular disease, pregnancy, or menopause), which might catalyze the reassessment of goals and increase motivation for change presenting a “teachable moment” in life (53). For example, individuals with chronic conditions, including diabetes and other cardiovascular diseases are often more likely to maintain or increase their leisure-time physical activity levels (54). Consistent with this, our study observed increased physical activity behavior among older adults who reported a new cardiovascular diagnosis. Thus, life transitions may serve as critical windows for intervention, offering opportunities to promote long-term physical activity engagement.

Our findings build on the growing body of literature demonstrating the influence of SSDoH on age-related health outcomes; for example, the influence of upstream factors on downstream protective behaviors such as physical activity engagement. Consistent with prior research, proximity to greenspace and social support were linked to increased physical activity behavior change (18). Similarly, high social support from friends and family was significantly associated with enhanced MVPA (19). However, we found that quantitative aspects of social support (e.g. frequency of visits from family and friends) were stronger predictors of behavior change than qualitative aspects (e.g. ability to confide in others or perceived loneliness). There is likely a complex, bidirectional relationship between social contact frequency and emotional support in influencing physical activity (55).

Even modest increases in MVPA can yield substantial health benefits for individuals with cardiovascular risk factors (56). However, comorbid conditions may necessitate personalized activity targets due to variability in clinically meaningful responses. By identifying individual differences in key factors influencing long-term behavior change, spanning behavioral, cognitive, neural, social, and structural determinants, our findings contribute to the growing evidence base that can be leveraged to develop scalable and effective personalized physical activity interventions. Despite mixed prior findings suggesting that antihypertensives and cholesterol-lowering medications such as beta-blockers and statins can impair exercise capacity due to muscle fatigue or reduced endurance (57), we did not identify a relationship between medication use and behavior change, suggesting these medications may not limit long-term MVPA engagement.

We identified neural markers that predicted successful physical activity behavior change among older adults following a cardiovascular risk diagnosis. Future increases in physical activity were associated with enhanced positive functional connectivity between the DMN and the FPN network. These findings align with prior research showing that DMN, especially the prefrontal cortex, support compensatory mechanisms in aging (30). Prior age-related neuroimaging research has shown that DMN is associated with complex decision-making processes critical for adaptive behavior in aging (58). Moreover, our finding of increased default mode to frontoparietal network coupling with enhanced physical activity behavior change supports the default-executive coupling hypothesis of aging (23): This model suggests that goal-directed cognition in older adults increasingly relies on accumulated knowledge (semanticization of cognition) to offset declining cognitive control resources for successful behavior (58). Default-executive coupling has been associated with positive behavioral outcomes including creative problem-solving and autobiographical memory (59).

Predominantly left-lateralized connections within heteromodal cortices predicted future successful MVPA maintenance. This finding underscores the role of large-scale brain networks in facilitating PA behavior change. A key region identified was the left prefrontal cortex, a hub implicated in executive control (60). Theoretical models proposed within the framework of Structured Event Complexes offer a foundational account of lateralized prefrontal function (22). According to this framework, the left prefrontal cortex is critical for processing the structural and semantic components of plans, including the temporal and logical organization of goal-related actions. Supporting this view, lesion studies have shown that damage to the left prefrontal cortex disproportionately impairs the ability to generate, structure, and execute multistep plans, even when general cognitive abilities are preserved. Wood et al. (61) further emphasized that planning processes relying on the hierarchical analysis of actions are more severely affected by left than right prefrontal lesions. This is further corroborated by the neuropsychological and lesional literature supporting the ROtman–Baycrest Battery to Investigate Attention (ROBBIA) model (62–64). The ROBBIA model outlines the cognitive processes underlying goal-directed behavior including task-setting, monitoring and energization, and where task-setting is dependent on left-lateralized prefrontal systems (65). While beyond the scope of the current study, prior research on hemispheric contributions to goal-directed behavior provide a testable framework for component cognitive processes that may be important to health behavior change.

Our findings point to a possible large-scale network connectivity fingerprint as a marker of individuals who may be the most receptive to changing their lifestyle behavior. This has significant clinical implications, particularly for the development of targeted interventions. While such biomarkers may help identify individuals who are more amenable to lifestyle modification, they also raise critical questions about how best to intervene in those who appear least responsive, arguably the population most in need of support. Understanding the neural correlates of behavioral change may inform personalized approaches that go beyond one-size-fits-all strategies, guiding the development of tailored and potentially more effective treatments. In this context, advances in brain stimulation therapies underscore the translational potential of functional connectivity-based biomarkers. Emerging evidence suggests that therapeutic outcomes are closely tied to the intrinsic functional connectivity of the targeted brain region (66). By leveraging resting-state functional connectivity, it may be possible to refine stimulation targets, and ultimately personalize interventions based on an individual's unique neural architecture. Such an approach aligns with precision medicine efforts and highlights the broader therapeutic promise of functional imaging not only for prognosis and risk stratification but also for informing intervention strategies in populations with lower propensity for behavior change.

Finally, our observation that combining multimodal brain and behavioral features leads to an increase in model performance suggests that these features provide independent and relevant information for predicting changes in physical activity. Previous studies (67) have also demonstrated that multimodal prediction models outperform unimodal ones. This improvement in prediction performance may arise because individual features capture distinct aspects of complex behaviors related to physical activity behavior change, insights that unimodal features alone may fail to capture.

The differential significance of unimodal versus multimodal models may reflect complex interactions between SSDoH and biological mechanisms including brain connectivity. Prior research supports the complex interaction between brain and environment: Chronic exposure to social and environmental stressors induces cumulative biological “wear-and-tear,” or allostatic load, also conceptualized as brain reserve (32, 68), which encompasses dysregulation across multiple physiological systems—including the hypothalamic-pituitary-adrenal axis, autonomic nervous system, immune, cardiovascular, and metabolic systems (69). Crucially, these systems converge in their expression across brain networks, positioning brain-based biomarkers as key integrators of SSDoH effects. By combining neural, behavioral, and contextual variables, multimodal models may better capture the full spectrum of determinants influencing behavior change in aging populations. This supports a systems-level framework, providing a more holistic understanding of modifiable health behaviors and informing precision interventions targeted at individuals most at risk.

Despite these contributions, an important limitation of this work is that the correlational nature of functional connectivity analyses prevents us from determining causality in brain-behavior relationships (70). Second, physical activity was assessed at baseline and at follow-up, but without fine-grained temporal resolution, we cannot determine the timing or stability of behavior change across the 4-year period. Third, some nonimaging features such as depression and anxiety, were measured using brief two-item questionnaires, which may not fully capture the complexity of these constructs. Finally, the absence of an external validation dataset limits the generalizability of our predictive models beyond the current sample. Nonetheless, our study has several notable strengths. It represents the largest and most comprehensive assessment of the brain, behavioral, and contextual factors predicting successful longer-term physical behavior change after cardiovascular diagnosis in aging. This study highlights the importance of going beyond individual-level factors and considering structural factors such as greenspace and social support to promote physical activity behavior change, evidence that is critical to guide policy decision-making and urban planning. Additionally, the consistency of predictive features across both self-reported and objectively measured physical activity models strengthens the robustness and generalizability of our findings, particularly given the well-documented discrepancies between these measurement approaches (71).

Future research must adopt a life course perspective to identify factors in younger or midlife adults and build a comprehensive understanding of physical activity behavior change across the lifespan. Moreover, external validation of the predictive models on independent datasets will be essential to assess their generalizability and reproducibility across diverse populations.

Conclusion

This study demonstrated that individual differences in brain, cognition, behavior, and contextual factors, including SSDoH, drive a complex human behavior: Future engagement in physical activity among older adults that are newly diagnosed with a cardiovascular illness. Leveraging mechanistic predictors of future physical activity and adopting a precision medicine framework will potentially lead to targeted interventions that result in sustained behavioral change and dementia prevention.

Supplementary Material

Supplementary material is available at PNAS Nexus online.

Funding

This research was undertaken thanks in part to funding from the Canada First Research Excellence Fund, awarded through the Healthy Brains, Healthy Lives initiative at McGill University. This research was undertaken thanks in part to funding from a National Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant (DGECR-2022-00299), an NSERC Early Career Researcher Supplement (RGPIN-2022-04496), a Fonds de Recherche Santé Québec (FRSQ) Salary Award, the Canada Brain Research Fund (CBRF), an innovative arrangement between the Government of Canada (through Health Canada) and Brain Canada Foundation, in collaboration with the Canadian Institutes of Health Research (CIHR), a Brain Canada Future Leaders Award, an Alzheimer Society Research Program (ASRP) New Investigator Grant, the Canadian Institutes of Health Research, Fondation J.A. DeSève, the Canada First Research Excellence Fund, awarded through the Healthy Brains, Healthy Lives initiative at McGill University, and the National Institutes of Health (P30 AG048785) to M.R.G., and the Consortium pour l’Identification précoce de la Maladie d’Alzheimer-Québec (CIMA-Q) and a doctoral studentship from the Fonds de Recherche Santé Québec awarded to N.T. S.S.G. was partially supported by National Institutes of Health projects P41EB019936 and RF1MH121885.

Author Contributions

N.T. (Conceptualization, Data Curation, Formal Analysis, Funding Acquisition, Investigation, Methodology, Visualization, Writing Original Draft, Writing Review Editing), S.S.G. (Methodology, Software, Writing Review Editing), Y.I.-M. (Investigation, Methodology, Writing Review Editing), and M.R.G. (Conceptualization, Funding Acquisition, Investigation, Resources, Supervision, Writing Review Editing)

Data Availability

The individual-level UK Biobank data can be obtained from https://www.ukbiobank.ac.uk/. The code required to run the analyses is available through Github (https://github.com/nagatv11/cvd_MVPA.git).

Ethics Statement

This study utilized data from the UK-Biobank study, which obtained ethics approval from the Northwest Multi-Centre Research Ethics Committee (MREC, approval number: 11/NW/0382), and obtained written informed consent from all participants prior to the study in accordance with the Declaration of Helsinki. This research has been conducted using the UK Biobank Resource under application no. 45551.