Source-linked AI summary
Causal discovery identifies pathways linking physical activity to dementia risk in the UK BioBank
Wasif Khan, Panayiotis V. Benos, Joshua K. Wong, Ruogu Fang
TL;DR
Physical activity is associated with lower dementia risk, but the pathways linking activity to dementia are incompletely understood. The study combines LLM-guided causal discovery and mediation analysis in older UK Biobank adults, identifying depression as a central pathway among interconnected behavioral, psychological, functional, and cardiometabolic routes.
Problem
Although physical activity is associated with lower dementia risk, the mechanisms linking activity to dementia prevention remain incompletely understood.
Source: §1 · p. 2Method
The study combined LLM-guided feature selection, causal discovery, and mediation analyses in 42,293 older UK Biobank adults to identify and assess pathways linking objectively measured MVPA to dementia risk.
Source: §1 · p. 3Results
Depression accounted for approximately 15% of the total association, with stronger mediation in males than females, amid multiple behavioral, psychological, functional, and cardiometabolic pathways.
Source: §4 · pp. 8–9Takeaways & Limitations
The findings support an integrated prevention framework combining physical activity promotion, mental health care, and cardiometabolic risk management.
Source: §4 · p. 10Because the analyses are observational, bidirectional relationships between depression and MVPA cannot be fully excluded.
Source: §4 · p. 11
Abstract
from arXiv · showhide
Physical Activity (PA) is consistently associated with lower risk of dementia, yet the mechanism linking PA to dementia prevention remain incomopletely understood. Here, we integrate large language model (LLM)-guided causal discovery with mediation analysis in 42,293 older adults aged 60 years or older from the UK Biobank to systematically identify pathways connecting objectively measured moderate-to-vigorous physical activity (MVPA) to dementia risk. Across behavioral, psychological, functional, and clinical domains, causal discovery consistently identified interconnected pathways linking higher MVPA to lower dementia risk through depression, functional capacity, smoking behavior, hypertension, cardiovascular disease, chronic kidney disease, and brain injury. Chain mediation analyses further identified depression as a central pathway, accounting for 15.1% of the overall association between MVPA and dementia risk. Sex-stratified analyses revealed distinct mechanistic patterns, with females exhibiting predominantly metabolic and functional pathways, while males showed behavioral and cardiometabolic cascades involving smoking and cardiovascular disease. Together, these findings suggest that the protective association between PA and dementia is mediated through interconnected behavioral, psychological, and cardiometabolic processes rather than a single pathway. Depression emerged as a prominent and potentially modifiable pathway, highlighting opportunities for integrating dementia prevention strategies that combine PA promotion with mental health and cardiovascular risk management.
1. Introduction
Although physical activity is associated with lower dementia risk, the pathways connecting activity to dementia remain incompletely understood. This study combines LLM-guided feature selection, causal discovery, and mediation analyses to examine objectively measured MVPA in older UK Biobank participants.
1. Introduction
- Prior research consistently associated higher physical activity with lower dementia risk, but whether activity acts directly or through intermediate pathways remained unclear.Proposed intermediates included biological, psychological, and functional processes that may interact.Source: §1 · pp. 2–3
- The study used LLM-guided feature selection alongside causal discovery and mediation analyses to map and quantify pathways, then assessed their robustness with bootstrap and sensitivity analyses.The analysis covered behavioral, psychological, functional, and clinical variables.Source: §1 · p. 3
2. Results
- Among 42,293 participants aged 60 years or older, 647 (1.5%) developed dementia after one year; mean MVPA was 40.2 minutes per day.Participants averaged 67.0 years of age, and 53.3% were female.Source: §2 · p. 5
Causal pathways linking physical activity to dementia risk
- Causal discovery linked MVPA to dementia through depression and walking pace, including a walking pace–smoking–depression cascade; other contributors included alcohol use and major comorbidities.Depression was directly connected to both MVPA and dementia in the inferred network.Source: Causal pathways linking physical… · pp. 5–6
Chain Mediation Analysis
Depression partially mediated the association between MVPA and dementia, and its mediated share increased with longer exclusion windows. The inferred pathways also differed by sex, while depression remained a consistent mediator.
Chain Mediation Analysis
- 15.06% of the total MVPA–dementia association was mediated by depression, with a higher proportion in males (19.31%) than females (12.27%).The indirect effect was β = −0.0028 and the total effect was β = −0.0183.Source: Chain Mediation Analysis · p. 6
- Male estimates rose from 19.3% at 1 year to 60.8% at 5 years, whereas female estimates remained relatively stable at 11.6–13.0%.The modeled direct association progressively attenuated and was no longer statistically significant after 5 years in the overall cohort and after 2 years in males. Both depression pathways remained significant across all exclusion periods.Source: Chain Mediation Analysis · p. 7
3. Sensitivity and Subgroup Analysis
- Females primarily showed metabolic and functional pathways, whereas males showed depression-linked routes involving cardiovascular and clinical conditions, smoking, particulate exposure, and hearing impairment.Depression was identified as a mediator across sexes; longer exclusion windows also revealed cardiometabolic and neurological pathways with low bootstrap stability (0.07–0.10).Source: §3 · pp. 7–8
4. Discussion
The analyses suggest that MVPA’s association with dementia risk may involve interconnected psychological, behavioral, functional, cardiometabolic, and neurological pathways. Depression was a consistent pathway, while mediation estimates and other pathway patterns varied by sex and exclusion window.
4. Discussion
- MVPA–dementia pathways included depression, functional capacity, smoking, hypertension, cardiovascular disease, chronic kidney disease, and brain injury.The discussion also identifies age, alcohol use, stroke, and hearing impairment as direct influences on dementia risk.Source: §4 · p. 8
- The depression-mediated pathway remained stable when incident dementia cases were excluded; longer exclusion windows revealed additional hypertension, heart disease, brain injury, and chronic kidney disease pathways.The discussion interprets these patterns as potentially reflecting psychological and behavioral pathways in the shorter term and cardiometabolic and neurovascular mechanisms over longer periods.Source: §4 · pp. 8–9
- Females showed predominantly metabolic and functional pathways, whereas males showed more complex behavioral and cardiometabolic pathways; depression was consistent across both sexes.The proposed explanations involving body composition, hormonal regulation, energy metabolism, lifestyle patterns, and disease burden are presented as possibilities.Source: §4 · p. 9
- Depression accounted for approximately 15% of the total effect, with stronger mediation in males (~19%) than females (~12%).The estimated proportion mediated rose from 15% at a 1-year exclusion window to 23% at 5 years; the authors caution that this relative measure is influenced by attenuation of the residual direct association.Source: §4 · p. 9
- The authors support prevention strategies combining PA promotion with depression prevention and early mental health screening, alongside cardiometabolic risk management.They also identify functional capacity, including walking pace, as part of the pathways linking MVPA to dementia risk.Source: §4 · p. 10
- The authors caution that observational data cannot establish definitive causality or exclude residual confounding, and that UK Biobank characteristics may limit generalizability.They also note that seven-day accelerometry may not capture lifetime activity patterns, disease severity may be incompletely represented by ICD codes, and bidirectional depression–MVPA relationships cannot be excluded.Source: §4 · pp. 10–11
Study design and participants
The study used accelerometer-derived PA data from UK Biobank participants and restricted analyses to older adults with valid accelerometer data and complete covariate information.
Study design and participants
- The analysis drew on accelerometer-derived PA data from a UK Biobank assessment using a wrist-worn device for seven days between 2013 and 2015.MVPA was quantified with a machine learning–based Random Forest approach; the analytic sample was restricted to adults aged 60 years or older with valid activity data and complete covariates.Source: Study design and participants · p. 11
Inclusion and exclusion criteria
Participants were excluded for consent withdrawal, incomplete or poor-quality accelerometer data, dementia diagnosed within one year after measurement, or missing model covariates.
Inclusion and exclusion criteria
- Participants were excluded for withdrawn consent, missing any hourly activity data, or more than 10% of daily MVPA during 01:00–04:00.Accelerometer quality exclusions included insufficient wear time, incomplete 24-hour coverage, calibration problems, recording interruptions, and excessive data anomalies.Source: Inclusion and exclusion criteria · p. 11
- Poor-quality accelerometer records were excluded, including wear time under 72 hours or failure to cover all 24-hour periods across the seven-day span.Other listed issues included unreliable file size, uncalibrated or reused devices without recalibration, recording interruptions, and more than 768 anomalies.Source: Inclusion and exclusion criteria · pp. 11–12
- Participants diagnosed with dementia within one year after wearing the accelerometer and those missing model covariates were excluded.Dementia included Alzheimer’s disease, vascular dementia, and other unspecified neurodegenerative dementias identified using specified ICD-10 codes; covariate exclusions yielded complete-case analysis.Source: Inclusion and exclusion criteria · p. 12
LLM-based semantic feature selection
The study used a staged LLM-guided process to select and curate UK Biobank variables before causal discovery and mediation analyses.
LLM-based semantic feature selection
- Candidate variables were screened in batches by an LLM relevance classifier, then scored by four GPT-based models; variables with mean scores below 0.60 were excluded.Expert-informed reviewer prompts and consensus ranking further assessed relevance, with top-ranked sets including the top 500, 250, and 100 variables considered for downstream modeling.Source: LLM-based semantic feature selection · p. 12
- The top 100 ranked variables were manually checked for data quality and epidemiological relevance, with highly missing variables excluded.Pairwise correlation analysis was used to remove redundant variables and reduce multicollinearity.Source: LLM-based semantic feature selection · p. 12
- The curated variables were used for causal discovery integrating domain knowledge, the PC algorithm, and bootstrapping to estimate a DAG of MVPA–dementia pathways.Identified pathways were then assessed with chain mediation analysis to quantify direct, indirect, and total effects.Source: Figure 3 · p. 13
- Established dementia risk factors, including hearing impairment and brain injury history, were included when absent and derived from harmonized ICD-9 and ICD-10 codes.These additions followed automated ranking and manual curation.Source: LLM-based semantic feature selection · pp. 12, 14
Causal Discovery and Mediation analysis
The study imposed domain-informed constraints on causal discovery, checked inferred edges with bootstrap resampling, and used chain mediation to estimate pathway-specific effects. Mediation models adjusted for demographic, socioeconomic, and behavioral covariates.
Causal Discovery and Mediation analysis
- Domain constraints treated demographic and socioeconomic factors as exogenous and dementia as a sink node in the causal graph.The PC algorithm tested conditional independence using Fisher’s Z transform with α = 0.05.Source: Causal Discovery and Mediation analysis · p. 14
- Bootstrap resampling re-estimated the causal graph to assess the robustness of inferred edges.The resampled datasets used identical settings.Source: Causal Discovery and Mediation analysis · p. 14
- Chain mediation used standardized path coefficients to estimate indirect effects as products along mediator sequences and total effects as direct plus indirect effects.The proportion mediated was calculated from the indirect and total effects.Source: Causal Discovery and Mediation analysis · p. 14
- Mediation models adjusted for age, sex, education, ethnicity, deprivation, smoking, and alcohol consumption to reduce potential confounding.Age was modeled continuously.Source: Covariates · pp. 14–15
Ethics statement
The study received UK Biobank ethics approval and obtained written informed consent from all participants.
Ethics statement
- The study was approved by the UK Biobank Ethics Advisory Committee, followed the Declaration of Helsinki, and obtained written informed consent from all participants.The analysis used UK Biobank accelerometer sub-study data under Application 48388.Source: Ethics statement · p. 15
Data availability statement
The study used UK Biobank data that are available only through an application and approval process.
Data availability statement
- The supporting UK Biobank data are not publicly available and may be accessed upon application approval.The study used the dataset under approved application ID 48388.Source: Data availability statement · pp. 15–16
p-value
The supplied statistical note specifies the tests and threshold used for baseline comparisons.
p-value
- Continuous-variable p-values were calculated with the Mann–Whitney U test, categorical-variable p-values with the chi-squared test, and significance was set at p < 0.05.Values were reported as mean ± standard deviation for continuous variables and n (%) for categorical and binary variables.Source: p-value · p. 25