Identifying modifiable factors and their joint effect on brain health: an exposome-wide association study

Abstract

Considerable uncertainty remains regarding the associations of multiple factors with brain health. We aimed to conduct an exposome-wide association study on neurodegenerative disease and neuropsychiatry disorders using data of participants from the UK Biobank. Multivariable Cox regression models with the least absolute shrinkage and selection operator technique as well as principal component analyses were used to evaluate the exposures in relation to common disorders of central nervous system (CNS). Restricted cubic splines were conducted to explore potential nonlinear correlations. Then, weighted standardized scores were generated based on the coefficients to calculate the joint effects of risk factors. We also estimated the potential impact of eliminating the unfavorable profiles of risk domains on CNS disorders using population attributable fraction (PAF). Finally, sensitivity analyses were performed to reduce the risk of reverse causality. The current study discovered the significantly associated exposures fell into six primary exposome categories. The joint effects of identified risk factors demonstrated higher risks for common disorders of CNS (HR = 1.278 ~ 3.743, p < 2e-16). The PAF varied by exposome categories, with lifestyle and medical history contributing to majority of disease cases. In total, we estimated that up to 3.7 ~ 64.1% of disease cases could be prevented.

This study yielded modifiable variables of different categories and assessed their joint effects on common disorders of CNS. Targeting the identified exposures might help formulate effective strategies for maintaining brain health.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11357-024-01224-x.

Introduction

As a result of the demographic transition worldwide, neurodegenerative diseases including multiple sclerosis (MS), Alzheimer’s disease (AD), Parkinson’s disease (PD), and all cause dementia are undergoing an increase in their prevalence [1]. Several lines of scientific data suggested that neuropsychiatric disorders (e.g., anxiety, major depressive disorder [MDD], bipolar affective disorder, and schizophrenia) and neurodegenerative disorders shared common underlying cellular and molecular mechanisms, such as oxidative stress and neuroinflammation [2]. Although prevention of these central nervous system (CNS) diseases has been a focus of primary care medicine, it is now recognized that up to 40% of dementia, stroke, and depression cases are attributable to modifiable risk factors [3]. Therefore, identification of the potential modifiable factors for CNS disorders will help us take effective measures to improve brain health on the long term.

More than 50% of complex disease risk results from differences in an individual's environment [4]. Many cardiovascular risk factors and lifestyles (including obesity, hypertension, diabetes, and smoking) have been reported to increase the load of CNS disorders [5–7]. Besides, environmental factors also play a critical role in modulating cardiovascular and brain aging [8]. However, conventional exposures just account for a small proportion of the “exposome”, the total exposure load occurring throughout a person’s lifetime [4]. Hence, investigating one or a handful of exposures at a time may cause type I errors and selective reports. Individuals are simultaneously exposed to multiple risk factors in real-world; therefore, a new model is required to discover exposures associated with CNS diseases while mitigating possibilities of selective reporting.

The exposome-wide association study (EWAS) represents an approach through which multiple environmental factors can be systematically searched for and validate their associations with complex diseases [9]. EWAS evaluates multiple exposures for association, with proper adjustment for multiplicity and collinearity of comparisons. Moreover, even in its partial forms, the exposome provides a useful framework to systematically assess associations, which can avoid the limitations and challenges of single-exposure studies [10]. Using both the least absolute shrinkage and selection operator (LASSO) technique and principal component analysis (PCA), we conducted an EWAS to comprehensively identify risk factors for CNS disorders. Besides, restricted cubic splines (RCS) were conducted to explore potential nonlinear correlations and probe the optimal ranges of continuous variables for maintaining brain health. The joint effects of risk factors were calculated by building weighted standardized scores. We then quantified the population attributable fraction (PAF) for each domain and in total for CNS disorders. Finally, sensitivity analyses by excluding those who was followed up for less than 3 years were performed to avoid reverse causality (Fig. 1).

Fig. 1.

Overview of analytic design. Analytical procedure to identify modifiable risk factors associated with central nervous disorders in the UK Biobank. Abbreviations: PAF, population attributable fraction; LASSO, least absolute shrinkage and selection operator

Methods

Participants

Data were from individuals enrolled in UK Biobank, a large-scale longitudinal cohort study that recruited more than 500,000 participants in 22 centers across the UK [11]. The baseline demographic, physiological, and clinical data were collected between March 2006 and October 2010. Participants were followed up until the date of first diagnosis, death, loss to follow-up, or the last date with available information (December 2020), whichever came first. Participants with dementia, PD, MS, epilepsy, stroke, anxiety, MDD, schizophrenia, bipolar affective disorder, or other chronic neurological problems (nervous system infection, brain abscess, encephalitis, demyelinating disease, cerebral aneurysm, cerebral palsy, and brain hemorrhage) at baseline were excluded from the analysis. All participants provided informed consent at recruitment.

Exposure variables

In the initial screening, we excluded those with missing values > 20% and retained the remaining exposure variables. Then, we manually removed obviously unmodifiable exposure variables, such as those from the “Genomics”, “Cognitive function”, and “Biological samples” categories. We further selected variables with no less than two levels and set all meaningless negative values as missing (NA). For binary variables, a major disease was determined using the International Statistical Classification of Diseases and Related Health Problems 10th Revision (ICD-10) codes and the self-reported diseases field. A total of 34 major diseases such as cardiovascular disease, diabetes, and hypertension were included in the analysis (details in Table S1). For ordinal variables, we reset the order of levels following ordinal progression (details in Table S2) and removed variables with only one level. For some continuous variables (such as hand grip strength), we generated substitutive variables by summing or averaging the values (details in Table S3). The remaining 166 variables fall into the following broad categories: (i) systematic diseases (i.e., cardiovascular disease, diabetes, and hypertension); (ii) lifestyle factors; (iii) anthropometric indicators (physical measures); (iv) blood sample examinations (biochemical markers); and (v) others (such as personality traits and social support indicators).

CNS disorders

We selected ten brain health–related phenotypes as our outcomes, including stroke, all cause dementia, AD, PD, MS, anxiety, MDD, bipolar affective disorder, schizophrenia, and epilepsy. The brain disorders were ascertained and classified according to the corresponding ICD codes, extracted from UKB health outcome datasets first occurrences of health outcomes (category 1712), and algorithmically defined outcomes (category 42). Specifically, the dementia cases were defined as all-cause dementia containing AD, vascular dementia, frontotemporal dementia, dementia with Lewy bodies, and dementia in other neurodegenerative or specified diseases. The stroke cases consisted of ischemic stroke (transient cerebral ischemic attacks and cerebral infarction), hemorrhagic stroke (intracerebral hemorrhage and subarachnoid hemorrhage), and stroke not specified as hemorrhage or infarction. Follow-up visits began from the date of attending the assessment center (field 53) to the earliest date of any brain disorder diagnosis, date of death (field 40,000), or the last available date from the hospital inpatient data (field 41,280–41281) or primary care data (field 42,040), whichever occurred first.

Statistical analyses

Exposure variables with missing values < 20% were imputed based on a maximum likelihood estimation method which was informed by the observed correlation structure within the data. Further, we performed standardization and generated Z score (mean = 0 and SD = l) for continuous variables.

The main analyses were conducted in 4 steps. Firstly, we randomly divided the data into the discovery dataset and the validation dataset. Multivariable Cox regression models with the LASSO technique were used to reduce potential false positive results due to the correlation between exposures. LASSO technique used a shrinking (regularization) process in which the coefficients of the regression variables were penalized, thus shrinking some of them to zero. Valid risk factors were defined as those showing significant associations with CNS diseases in both the discovery and validation dataset. Besides, as supplements for LASSO regression analyses, we performed PCA for quantitative variables and multiple correspondence analysis (MCA) for qualitative variables, respectively. Secondly, to estimate the nonlinear relationships of exposures with CNS disorders and explore the optimal range of continuous variables, we used RCS with knots at the 5th, 27.5th, 50th, 72.5th, and 95th percentiles of the exposures. Thirdly, exposures were reverse coded to reflect the detrimental aspect and participants scored 1 point for each detrimental factor. Weighted standardized scores generated based on the coefficients of each variable to evaluate the joint effects of risk factors on CNS diseases. PAF was following calculated to estimate the potential impact of eliminating the unfavorable profiles of risk domains on CNS disorders. Finally, we performed the sensitivity analyses by (1) excluding factors with high multicollinearity (r² > 0.8) to alleviate collinearity; (2) excluding those who was followed up for less than 3 years to reduce the risk of reverse causality.

All analyses were adjusted by age, sex, deprivation index (additional APOE ε4 status for cognitive disorders). Statistical analyses were performed using the R 4.0.2 language and environment (The R Foundation for Statistical Computing, Vienna, Austria) and IBM SPSS Statistics 25. Statistical significance was set at a two-tailed p value < 0.05 in regression models.

Results

A total of 363,125 participants were included in our study (median [quartile] age, 59 [51–64] years; 53% women). The flow diagram was displayed in Fig. S1, and baseline demographic characteristics of participants were presented in Table S4–5.

LASSO regression analyses

As shown in Fig. 2a, our study revealed that 7 factors significantly reduced risk for all-cause dementia (e.g., able to confide, hand grip strength, body fat percentage, and apolipoprotein A), and 3 factors (household income, peaking expiratory flow, and hand grip strength) significantly reduced the risk for AD. Inversely, another 16 factors (e.g., CHD, diabetes, diet variation, and systolic blood pressure) significantly increased risk for all-cause dementia and AD. As for other movement disorders, 3 factors related to physical measures and lifestyle factors significantly reduced PD risk. Besides, 3 factors (e.g., walking pace and vitamin D) were found to be protective against MS. Exposures classified into systematic diseases (e.g., CHD and diabetes), unhealthy lifestyle factors (e.g., current smoking), unfavorable anthropometric indicators (e.g., high blood pressure) as well as blood biochemical parameters (e.g., HbA1C) significantly increased the risk of stroke, while favorable physical measures, biochemical parameters and social support reduced stroke risk (details in Table S6).

Fig. 2.

Associations between exposome and CNS disorders. Different colored dots represent different brain disorders. Significant results from discovery dataset were showed in volcano plots and significant results from discovery dataset were showed in forest plots. As for volcano plots, horizontal axis indicates hazard ratios while vertical axis indicates the exp-transformed P value of the correlations between exposures and brain health disorders; the horizontal dotted line indicates the critical significance line (P = 0.05). AS for forest plots, horizontal axis indicates hazard ratios while vertical axis represents exposures; the vertical dotted line indicates the 0 scale line (HR = 1). Abbreviations: HR, hazard ratio; AD, Alzheimer’s disease; PD, Parkinson’s disease; MS, Multiple sclerosis; CHD, chronic heart disease; CLD, chronic liver disease; CKD, chronic kidney disease; IBS, irritable bowel syndrome; IBD, inflammatory bowel disease; DBP, diastolic blood pressure; SBP, systolic blood pressure; Apo A, apolipoprotein A; Apo B, apolipoprotein B; HbA1C, glycosylated hemoglobin; PEF, peaking expiratory flow

As for psychiatry disorders (Fig. 2b), systematic diseases, emotional instability, unhealthy lifestyle factors and unfavorable anthropometric indicators showed detrimental effects on anxiety. Similarly, exposures related to those four categories were also significantly associated with higher risk of MDD, while exposures related to social support (e.g., household income and able to confide) and favorable anthropometric indicators (e.g., walking pace and hand grip strength) were associated with a lower risk of MDD. Results also indicated the detrimental roles of systematic disease and unhealthy lifestyles, as well as the protective role of social support in schizophrenia and bipolar affective disease. Several systematic diseases (e.g., alcohol related disease) and unhealthy lifestyles increased the risk of epilepsy, whereas well social support decreased the risk (details in Table S6).

Additional analyses

The coefficients for nearly all the correlations between binary exposures were lower than 0.4. Hence, each binary exposure was considered as an independent principal component (PC). In addition, correlated exposures were mostly within the same category. All the continuous exposure variables in total population were merged into 11 PCs with eigenvalue > 1, explaining 72.3% of the data variation. All the binary exposure variables were considered as PCs due to the lower correlation coefficients. All continuous exposure variables were correlated with one of the PCs without much overlap (details in Fig. S2–3). In multivariable Cox models adjusted for age, sex, and deprivation index, PCs related to systematic diseases, unhealthy lifestyles as well as emotional instability were positively associated with the risks of CNS diseases, while PCs related to good lung function, well social support, and healthy diet protected against the risks (details in Fig. S4).

Restricted cubic splines

Figure 3 showed the nonlinear associations between continuous variables and CNS disorders. Physical measures (e.g., pause rate, BMI, and waist and hip circumstance) and diet (e.g., vegetable and meat intake) had U-shaped associations with brain health. Significant increases in CNS disease risks were observed at higher levels of tobacco exposure, HbA1C, and apolipoprotein B. Inversely, significant decreases in CNS disease risks were observed at higher levels of hand grip strength, lung function, and apolipoprotein A. Besides, the RCS also figured out a series of cut-off points of continuous variables (including cutoffs for physical measures, lifestyle factors, and biochemistry indicators), which indicated the optimal range of these markers to maintain brain health.

Fig. 3.

Nonlinear relationships between continuous variables and CNS disorders. Horizontal axis indicates exposures; vertical axis indicates the exp-transformed estimation of the correlations between exposures and brain health; the dotted line indicates the 0 scale line (HR = 1). Abbreviations: CNS, central nervous system

Joint-effects analyses of identified factors

Compared with the favorable profile, intermediate and unfavorable profiles of risk factors significantly increased the risk of CNS diseases with trend toward significance. As shown in Fig. 4a, the joint effects were the following: anxiety [HR = 1.986, p = 2.00e-16], MDD [HR = 2.391, p = 2.00e-16], bipolar disorder [HR = 2.780, p = 3.48e-14], schizophrenia [HR = 3.743, p = 2.61e-5], all cause dementia [HR = 1.592, p = 2.00e-16], AD [HR = 1.592, p = 2.00e-16], PD [HR = 1.278, p = 2.29e-9], MS [HR = 1.781, p = 9.06e-6], stroke [HR = 1.596, p = 2.00e-16], and epilepsy [HR = 1.573, p = 7.24e-10].

Fig. 4.

The joint effects of identified factors on CNS disorders and weighted population attributable fraction for the six domains. a Joint effects of identified factors on common CNS disorders. b Weighted population attributable fraction for the six domains

PAF estimates

When shifting all unfavorable profiles to intermediate and favorable ones, PAF estimation suggests that 3.7 ~ 64.1% of disease cases could be prevented. Interestingly, almost all the CNS disorders could be affected by and unhealthy lifestyles (PAF: 0.8 ~ 23.1%). Neuropsychiatric disorders were largely affected by mood instability (PAF: 7.2 ~ 16.7%) (Fig. 4b, details in Table S7).

Sensitivity analyses

Our results remained robust after excluding exposures with high multicollinearity (r²> 0.8) from analysis. Excluding those who was followed up for less than 3 years, the results were almost unchanged (details in Fig. S5-6).

Discussion

Using the exposome-based approach, we found that the exposures which had significant associations with common CNS disorders fell into six exposome categories, including systematic diseases, lifestyle factors, social support, anthropometric indicators, personality traits as well as biochemical markers. Overall, 3.7 ~ 64.1% of disease cases could be prevented through adhering to more favorable profiles in these six domains.

The mechanisms underpinning the development of depression or anxiety were unclear, but our results supported some previous hypotheses involving neurophysiological and psychosocial pathways. Several lines of evidence have shown elevated levels of pro-inflammatory markers in patients with neuropsychiatric disorders, suggesting that chronic low-grade inflammation might play an important role in the pathophysiology of mood disorders [12]. Since chronic inflammation emerged as a key feature of cardiovascular disease, metabolic disorders, as well as immune diseases [13], the associations of systematic diseases (e.g., CHD, obesity, asthma, alcohol related disease) with anxiety and depression could be explained by inflammation. Besides, brain functional and structural alterations were observed in patients with anxiety and depression, and the most affected brain structure in people with mood disorders was the hippocampus, an area involved in emotional processing and stress regulation [14]. Therefore, the associations of mood swings with anxiety and depression might be mediated by hippocampal structural alterations. On the contrary, physical activity improves anxiety and depression via various neuro-molecular mechanisms, including elevated expression of neurotrophic factors, increased availability of serotonin and norepinephrine, regulation of hypothalamic–pituitary–adrenal axis activity and reduced systemic inflammation [15, 16]. Moreover, we also found the protective effects of social support (e.g., household income and able to confide) on psychiatric disorders, including schizophrenia. Hence, healthy lifestyle behaviors, emotional stability, and social support contribute to the prevention mental disorders.

Among the included systematic diseases in the current study, diabetes and CHD are the leading causes of dementia/AD. As for AD pathologies, a potential “head-to-heart” link has been proposed, which has been ascribed to a reduction in cerebral perfusion [17]. Cerebral hypoperfusion tended to cause acidosis and oxidative stress, in which the altered metabolism of neurons would lead to the formation of neurofibrillary tangles and amyloid plaques [18]. Similarly, (neuro)inflammation also offered a vital connection for the metabolic impairments and dementia; that is, increased reactive oxygen species and nitrous oxide generation would cause tissue damage and Aβ aggregation [19]. Besides, several metabolic disturbances, such as diabetes, could also indirectly affect dementia risk through causing cardiovascular dysfunction [20]. Our PCA revealed the protective effects of body shape- and body composition–related anthropometric variables on dementia/AD. However, some previous studies showed higher BMI in midlife increased dementia risk, while higher late-life BMI might decrease the risk [21]. Hence, we should interpret the result with caution because of the age-dependent effects of anthropometric indicators. Moreover, the significant associations of social support and lifestyles with dementia in the current study indicate future prevention strategies.

PD begins with the atrophy and progressive loss of dopamine-containing neurons, followed by α-synuclein aggregation in the brain [22]. The mechanism underlying the association of diabetes with the increased incidence of PD has not yet been elucidated. However, impaired insulin signaling, dysregulation of glucose metabolism, as well as the pathogenic protein folding, and aggregation and accumulation lead to reduced neuronal survival and the occurrence of PD [23]. We also found the protective effects of physical activity (such as walking pace) on PD, which could be explained by exercise-induced increases in brain-derived neurotrophic factor [24]. Likewise, the current results also declared the protective effects of physical activity on MS risk. Future government initiatives are required to improve physical activity across the lifespan.

Our results suggested that atherosclerosis increased the risk of epilepsy, of which the molecular mechanism was still unclear. It is well established that chronic epilepsy causes cognitive impairment. Some of the chronic cognitive dysfunction of epilepsy might be atherosclerotic in etiology [25]. Thus, we speculated that atherosclerosis might indirectly lead to epilepsy by causing cerebrovascular disease, such as cerebral infarction. Future preservation of cognition might be accomplished via means which are very different from just the traditional strategy of epilepsy prevention [26]. Although there was a lack of molecular evidence, we also speculated that nap during a day might increase epilepsy risk by inducing poor sleep quality, since both poor sleep quality and excessive daytime sleepiness were reported as contributors to epilepsy [27]. We also revealed the adverse effect of smoking on epilepsy risk; therefore, traditional antiepileptic drugs supplemented with healthy lifestyles might yield significant benefits for epileptics.

In the current study, we disclosed the adverse effects of systematic disease, unfavorable anthropometric indicators and biochemical markers, and unhealthy lifestyles on brain health. All these risk factors could exacerbate fundamental cellular and molecular aging processes, promoting accelerated aging phenotypes [28]. The mechanisms include increased oxidative stress, cellular mitochondrial and energetic dysfunction, impaired cellular stress resilience, heightened state of inflammation, and disruption of intercellular communication (including endocrine changes) [8, 28]. Interestingly, we also revealed the protective roles of favorable anthropometric indicators (such as lung function, hand grip strength), social support, healthy lifestyles and vitamin D in brain health. Resistance training is known to increase brain-derived neurotrophic factor levels, a marker of neuronal growth and plasticity [29]. Therefore, adherence to healthy lifestyle supplemented with neurotrophic factors might yield significant benefits for brain health.

Our systematic approach aimed to overcome biases (e.g., selective reporting), but several limitations should be considered when interpreting our findings. First, a high level of volunteer participation bias might confine our interpretation of results. In addition, it had a relatively young age at recruitment and a short follow-up duration. Consequently, it had a low incidence of several CNS disorders, such as MS. Second, the current study was restricted by the available variables from the UK Biobank database, so estimation of the PAF of potentially modifiable risk factors is lower than previous estimates [30]. Third, though LASSO regression with bootstrapping combined the strength of LASSO for identifying important features for prediction (by shrinking uninformative variables to zero) with the robustness of bootstrapping, we still could not completely overcome the influence of collinearity among variables. However, PCA reduced the spectrum of the exposures into a smaller number of clusters without multicollinearity, so we consider PCA as an assistant analysis in interpreting LASSO regression results.

In conclusion, we used a comprehensive approach to systematically screen and rigorously assess the associations of a broad range of exposures with common CNS disorders. We summarized the variables associated with brain health in the six primary exposome categories, including several systematic diseases, lifestyle factors, social supports, anthropometric indicators, personality traits as well as biochemical markers. This highlighted the potential of integrating a broad array of exposures to revise effective strategies for maintaining brain health.

Supplementary Information

Below is the link to the electronic supplementary material.

Authors’ contributions

All authors had full access to the data in the study and accept responsibility to submit for publication. JT Yu designed the study. LY Huang, YJ Ge, and Y Fu conducted the main analyses and drafted the manuscript. YL Zhao, YN Ou, Y Zhang, LZ Ma, SD Chen, and ZX Guo contributed to imaging and data analyses. JF Feng, W Cheng, L Tan, and JT Yu critically revised the manuscript, and all authors approved the final version.

Funding

This study was supported by grants from the Science and Technology Innovation 2030 Major Projects (2022ZD0211600), National Natural Science Foundation of China (82071201, 82071997, 82271475), Shanghai Municipal Science and Technology Major Project (2018SHZDZX01), Research Start-up Fund of Huashan Hospital (2022QD002), Excellence 2025 Talent Cultivation Program at Fudan University (3030277001), Shanghai Talent Development Funding for The Project (2019074), Shanghai Rising-Star Program (21QA1408700), and 111 Project (B18015).

Data availability

The main data used in this study were accessed from the publicly available the UK Biobank Resource under application number 19542, which cannot be shared with other investigators. Any other data generated in the analysis process can be requested from the corresponding author.

Declarations

Ethics approval and consent to participate

All participants gave written informed consent prior data collection. The UK Biobank has full ethical approval from the NHS National Research Ethics Service (16/NW/0274).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.