Skip to main content
Advertisement
Browse Subject Areas
?

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Identifying a group of factors predicting cognitive impairment among older adults

  • Longgang Zhao,

    Roles Conceptualization, Data curation, Formal analysis, Methodology, Software, Visualization, Writing – original draft

    Affiliation Department of Epidemiology and Biostatistics, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina, United States of America

  • Yuan Wang,

    Roles Supervision, Writing – review & editing

    Affiliation Department of Epidemiology and Biostatistics, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina, United States of America

  • Eric Mishio Bawa,

    Roles Writing – original draft, Writing – review & editing

    Affiliation Department of Epidemiology and Biostatistics, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina, United States of America

  • Zichun Meng,

    Roles Methodology, Software, Writing – review & editing

    Affiliation Department of Epidemiology and Biostatistics, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina, United States of America

  • Jingkai Wei,

    Roles Writing – review & editing

    Affiliation Department of Epidemiology and Biostatistics, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina, United States of America

  • Sarah Newman-Norlund,

    Roles Writing – review & editing

    Affiliation Communication Sciences and Disorders, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina, United States of America

  • Tushar Trivedi,

    Roles Writing – review & editing

    Affiliation Regional Medical Center Primary Care Stroke, Orangeburg, SC, United States of America

  • Hatice Hasturk,

    Roles Writing – review & editing

    Affiliation Center for Clinical and Translational Research, Forsyth Institute, Boston, MA, United States of America

  • Roger D. Newman-Norlund,

    Roles Writing – review & editing

    Affiliation Communication Sciences and Disorders, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina, United States of America

  • Julius Fridriksson,

    Roles Writing – review & editing

    Affiliation Communication Sciences and Disorders, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina, United States of America

  • Anwar T. Merchant

    Roles Conceptualization, Funding acquisition, Methodology, Project administration, Resources, Supervision, Writing – original draft, Writing – review & editing

    merchant@mailbox.sc.edu

    Affiliation Department of Epidemiology and Biostatistics, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina, United States of America

Abstract

Background

Cognitive impairment has multiple risk factors spanning several domains, but few studies have evaluated risk factor clusters. We aimed to identify naturally occurring clusters of risk factors of poor cognition among middle-aged and older adults and evaluate associations between measures of cognition and these risk factor clusters.

Methods

We used data from the National Health and Nutrition Examination Survey (NHANES) III (training dataset, n = 4074) and the NHANES 2011–2014 (validation dataset, n = 2510). Risk factors were selected based on the literature. We used both traditional logistic models and support vector machine methods to construct a composite score of risk factor clusters. We evaluated associations between the risk score and cognitive performance using the logistic model by estimating odds ratios (OR) and 95% confidence intervals (CI).

Results

Using the training dataset, we developed a composite risk score that predicted undiagnosed cognitive decline based on ten selected predictive risk factors including age, waist circumference, healthy eating index, race, education, income, physical activity, diabetes, hypercholesterolemia, and annual visit to dentist. The risk score was significantly associated with poor cognitive performance both in the training dataset (OR Tertile 3 verse tertile 1 = 8.15, 95% CI: 5.36–12.4) and validation dataset (OR Tertile 3 verse tertile 1 = 4.31, 95% CI: 2.62–7.08). The area under the receiver operating characteristics curve for the predictive model was 0.74 and 0.77 for crude model and model adjusted for age, sex, and race.

Conclusion

The model based on selected risk factors may be used to identify high risk individuals with cognitive impairment.

Introduction

Between 29% and 76% of patients with dementia may remain undiagnosed [13] even though the Centers for Medicare and Medicaid Services recommends the assessment of impaired cognition for older individuals at annual wellness visits. As routine screening for impaired cognition is not recommended, evaluation or diagnosis of impaired cognition results from complaints by the individual or caregiver or provider suspicion [4].

Increasing age and the ε4 allele of the apolipoprotein E gene are the strongest predictors of impaired cognition [5], but several other factors are correlated with impaired cognition including blood pressure, diabetes, smoking, alcohol use, diet, physical activity, serum cholesterol [6, 7], and oral health status [8]. With the exception of age and the ε4 allele of the apolipoprotein E gene, these factors are potentially modifiable to prevent or delay cognitive decline. Blood pressure, blood sugar, serum cholesterol smoking, alcohol use, diet, and physical activity, were part of the Brain Care Score which predicted incident dementia, suggesting that improving the Brain Risk Score could reduce dementia risk [9]. While information on these risk factors is often present in medical records, a formal way of including them in the assessment of impaired cognition is challenging. We included information on oral health status in addition to other risk factors, including those mentioned above, because oral health status is correlated with poor cognition but infrequently included in risk scores. Moreover, this information is typically present in electronic health records and was available in the dataset that we were planning to use.

The goal of this study was to create and validate a score predicting the prevalence of impaired cognition in older adults from information on known risk factors using statistical and machine learning methods. A validated algorithm embedded into electronic health records could use pertinent information available in health records to calculate the likelihood of impaired cognition of their patient using all available data and cue the health care provider. Formal screening of individuals identified following this impaired cognition score could potentially increase the positive predictive value of existing screening tools such as the Mini-Mental State Examination (MMSE).

Methods

Data source

We used a subset of cross-sectional data from the Third National Health and Nutrition Examination Survey (NHANES III) which had measures of cognition for participants 40 years and older [10] and its commonly known risk factors as the data source to develop a impaired cognition score predicting impaired cognition. We called this the training dataset. The validation dataset consisted of cross-sectional information from participants of continuous NHANES surveys 2011 through 2014 who were 60 years and older with data on cognitive assessment and covariates [11]. This validation dataset was used to confirm the ability of the impaired cognition score developed in the training dataset to predict impaired cognition, and estimate its sensitivity and specificity.

There were 7,869 participants in the training dataset (NHANES III) and 19,931 participants in the validation dataset (NHANES 2011–2014). In the training dataset, 3795 participants were excluded because of missing values in covariates or cognition variables. Similarly, 16,259 participants in the validation dataset were excluded for no valid measurement for cognition and 962 for missing values in covariates. After exclusions, 4074 and 2510 participants remained in the training and validation datasets respectively (S1 Fig in S1 Appendix). NHANES III participants consented to participate in the study voluntarily. The data used in these analyses was deidentified and classified as exempt from review by the Institutional Review Board of the University of South Carolina.

Exposure assessment

Training dataset.

Predictive factors consisted of sociodemographic factors which included age, sex (male and female), race (white, black, and other), educational level (<12 years, ≥12 years completed education), poverty income ratio (PIR) divided into three groups (≤1.3, 1.3<PIR≤3.5, >3.5) (a value <1 indicates that household income is below poverty level, a value of 1.3 means that household income is 30% above poverty level and so on. The higher this number the wealthier the household), lifestyle factors consisting of smoking status (non-smoker, ever smoker, and current smoker), drinking status (non-drinker and drinker), healthy eating index (HEI) (higher number indicating healthier diet), health related factors including, body mass index (BMI) (normal: ≤24.9 kg/m2; overweight: 25 to ≤29.9 kg/m2, and obese: ≥30 kg/m2), waist circumference (in cm), physical activity (derived based on a structured physical activity questionnaire, and further classified as three groups: sedentary, moderate, vigorous), systolic blood pressure (SBP), diastolic blood pressure (DBP), white blood cell count (WBC), C-reactive protein (CRP), total cholesterol, high density lipoprotein cholesterol (HDL), triglycerides, disease prevalence including diabetes, hypertension, cardiovascular diseases, cancer, and depression, oral health information consisting of periodontal disease (none or mild versus moderate or severe) [12, 13], and annual dentist visits, and social connectedness measures consisting of frequency of talking to friends and family, visiting friends and family, attending church services and club meetings. Details about these variables are available at the CDC website [11, 12].

Validation dataset.

Covariates measured included age in years, sex (male and female), race (white, black, and other), education ((<12 years, ≥12 years completed education), PIR (≤1.3, 1.3<PIR≤3.5, >3.5), smoking (non-smoker, ever smoker, and current smoker), drinking status (non-drinker and drinker), HEI score, BMI (normal: ≤24.9 kg/m2; overweight: 25 to ≤29.9 kg/m2, and obese: ≥30 kg/m2), waist circumference, physical activity (sedentary, moderate, vigorous), SBP, DBP, WBC, total cholesterol, HDL, triglycerides, history of diabetes, hypertension, cardiovascular diseases, cancer, and depression, periodontal diseases, and annual dentist visits. The detailed collection methods for these covariates were similar to the training dataset.

Cognition assessment

Training dataset.

Cognition was measured using a version of the Mini-Mental State Examination (MMSE), which was administered through a home interview and at the Mobile Examination Center (MEC) [1416]. It consisted of six orientation, six recall, and five attention related questions. Each correct response was assigned 1 point and an incorrect response received a score of 0. The outcome used in these analyses was the total score, which was obtained by summing the points assigned to the responses and ranged from 0 to 17 with higher scores indicating better cognitive function. The six orientation items asked about the day of the week, the date, and participant’s complete address including street, city or town, state, and ZIP code (adult questionnaire). To evaluate recall, the interviewer told the participant the names of three items (apple, table, and penny) and asked them to repeat the names. Each participant was given up to six tries to learn the words. If they correctly recalled the items at any of the six tries, the response was considered to be correct. This exercise was repeated after assessing attention to achieve a maximum of 6 points for recall. To evaluate attention, the interviewer asked the participant to serially subtract 3 from 20 and repeat this exercise for up to five times. For example, they were asked to subtract $3 from $20, and $3 from $17, and so on and assigned one point for each correct answer. Based on the prevalence of cognitive impairment among US older adults [17], we defined cognitive impairment as the lowest 10% of the distribution (total score = 10) of the MMSE score.

Validation dataset.

Cognition was measured using three tests consisting of the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD), the Animal Fluency (AF) test, and the Digit Symbol Substitution Test (DSST) [18]. The CERAD model assesses both immediate and delayed memory of new verbal information [19], the AF test assesses the verbal fluency domain of executive function [20], and DSST assesses attention and processing speed. A score, which is the sum of the total number of correct symbols and number pairs within the time frame was assigned. The assessment was done at the MEC. Individuals in the lowest 10% of the score distribution were categorized as cognitively impaired. Details are provided in S1 Appendix. We also used the percentile 20% in the sensitivity analyses.

Statistical analysis

We developed models to predict cognitive impairment in NHANES III data (training dataset) using logistic regression and machine learning (support vector machine). We used machine learning methods to supplement the logistic regression models because they do not depend on the same modeling assumption. Comparing the results from the machine learning would increase our confidence that the final logistic regression models included all the important variables. We further used those models to predict cognitive impairment in the NHANES 2011 through 2014 (validation dataset). We chose the validation dataset because it had most of the variables that were measured in the training dataset.

Training logistic regression model.

The binary outcome for cognitive impairment was related to the predictors (cognition-related risk factors) in forward, backward, and stepwise models, from which the most predictive model was selected. We derived the beta coefficients for each variable from the most predictive model to construct the impaired cognition score.

Training support-vector machine (SVM) learning models.

We next obtained predictive models in the developmental dataset by relating the outcome (cognitive impairment) with the predictor variables using SVM. We used the 5-fold cross validation method with 30 replications to get the best estimation of the model. We ranked the contribution of each feature and compared the top 15 features from the SVM based on the development and validation datasets (Fig 1). Finally, we constructed the impaired cognition score based on the logistic model using the following formula:

thumbnail
Fig 1. Top 15 covariates for ranked contributions of covariates with poor cognition performance from support vector machine methods in both the development and validation datasets; BMI, body mass index; CRP, C-reactive protein; CVD, cardiovascular diseases; DBP, diastolic blood pressure; HDP, HDL cholesterol; HEI, healthy eating index; SBP, systolic blood pressure; TCP, total cholesterol; TGP, triglycerides; WBC, white blood cell; WC, waist circumference.

https://doi.org/10.1371/journal.pone.0301979.g001

Model validation.

To evaluate the performance of the prediction models obtained in the training dataset to external data, we used the final prediction model (logistic model) obtained from NHANES III data to predict impaired cognition in the continuous NHANES 2011 through 2014 data. To do that we calculated a impaired cognition score for each participant using the final logistic regression models obtained earlier and categorized the impaired cognition score into tertiles (low risk, moderate risk, and high risk). We estimated the odds ratios (OR) and the corresponding 95% confidence intervals (CI) using logistic regression to evaluate the association between the impaired cognition score in tertiles and the outcome (impaired cognition) adjusting for age, sex, and race (Table 3). We then calculated the area under the curve (AUC) graphs to assess the prediction performance (Fig 1).

Data management and statistical analyses were conducted using SAS 9.4 (SAS Institute, Cary, NC). The SVM model was performed in R program (Version 3.5.0, R core team, Vienna, Austria) with the “e1071” package. The sample weights provided by NHANES III were used in all procedures. The threshold for statistical significance was 0.05. The manuscript is compliant with the STROBE statement. The University of South Carolina IRB determined the analyses to be not Human Subjects research.

Ethical considerations

The NHANES study was approved by National Center for Health Statistics (NCHS) Ethics Review Board (Protocol #2011–17) and documented consent was obtained from all participants.

Results

Study population characteristics

We documented 411 participants with low cognitive performance. Table 1 shows the characteristics of participants in the training dataset from NHANES III. Participants with low cognitive performance compared with their peers with high cognitive performance were older (p<0.001), more likely to be African American, had higher systolic blood pressure (p = 0.001), lower HEI scores (p<0.001), fewer social connection by phone (p<0.001); these individuals were less likely to have visited a dentist in the past year (p<0.001) and more likely to have periodontal disease (p = 0.01).

thumbnail
Table 1. Characteristics of the training dataset using participants from NHANES III.

https://doi.org/10.1371/journal.pone.0301979.t001

In the validation dataset (NHANES 2011–2014), participants with low cognitive performance compared with their peers with good cognitive performance were older (p<0.001), more likely to be African American (<0.001), have higher systolic blood pressure (p<0.001), lower HEI score (p = 0.001), and visited the dentist less frequently in the last year (p<0.001) (S1 Table in S1 Appendix).

Associations between predictors and cognition performance

Table 2 describes associations between risk factors and cognition performance in the training dataset after adjusting for all covariates. The odds of low cognitive performance were greater with higher age but lower healthy eating index score, fewer social connections by phone, and fewer social connections in attending church activities. Among the participants of the training dataset, the odds of low cognition performance 1.30 times higher for every 5-year increase in age (aOR = 1.30, 95% CI = 1.19–1.41), and 1.58 times higher for African Americans compared with their White peers (aOR = 1.58, 95% CI = 1.15–2.17). The odds of low cognitive performance were 6% lower for a 5-unit increase in the healthy eating index score (aOR = 0.94, 95% CI = 0.89–0.98); 2% lower with more frequent daily social connection by phone (aOR = 0.98, 95% CI = 0.97–0.99); 13% lower with more frequent weekly participation in church activities (aOR = 0.87, 95% CI = 0.76–0.99), and 41% lower among those who visited a dentist in the last year.

thumbnail
Table 2. Associations between risk factors and cognition performance in the training dataset from the NHANES III.

https://doi.org/10.1371/journal.pone.0301979.t002

Construction and prediction of the overall impaired cognition score

The top 15 contributors to poor cognitive performance in both the training dataset and validation datasets are shown in Fig 1. The highest contributor in the validation dataset was education. In the training dataset, CRP was the highest contributor. Waist circumference and physical activity had the same rank in both datasets. Age was part of the top three contributors in either dataset. Visiting the dentist was ranked tenth in the validation dataset and eleventh in the training dataset, while healthy eating index and race were among the top ten contributors in both the training and validation datasets.

Table 3 describes the impaired cognition scores for cognitive performance in the training and the validation dataset using the logistic and SVM models. In both datasets, there were higher odds of poor cognition for those in tertiles 2 and 3 when compared to tertile 1. These findings were similar in both the logistic and SVM models in both datasets. In the training dataset, the odds of poor cognition among participants in tertile 3 was 8.15 times the odds among participants in tertile 1 (aOR = 8.15, 95% CI = 5.36–12.4) and 8.77 times that of the odds among participants in tertile 1 (aOR = 8.77, 95% CI = 5.57–13.8) for the logistic and SVM models respectively. In the validation dataset, the odds of poor cognition among those in tertile 3 was 4.31 times the odds of participants in tertile 1 (aOR = 4.31, 95% CI = 2.62–7.08) and 3.95 times that of the odds among participants in tertile 1 (aOR = 3.95, 95% CI = 2.34–6.66) for the logistic and SVM models respectively. Using 20% as the cutoff of impaired cognitive performance in the validation dataset yielded similar results (S2 Table in S1 Appendix).

thumbnail
Table 3. Impaired cognition score for cognition performance based on the development and prediction datasets.

https://doi.org/10.1371/journal.pone.0301979.t003

The logistic and SVM models had similar discriminatory abilities in predicting poor cognition using the validation dataset. The area under the curve (AUC) for the crude model in was 0.74 and 0.70 for the logistic model and SVM model, respectively. The AUC for the age, sex and race-adjusted model in the logistic and SVM models was 0.77 and 0.76 respectively, and the AUC for the models that adjusted for all covariates in the logistic and SVM methods was 0.78 and 0.79 respectively (Fig 2).

thumbnail
Fig 2.

Performance of the prediction model using both logistic regression and support vector machine methods in the validation datasets; (a) logistic model; (b) SVM model. Adjusted model included age, sex, and race. Covariates model additionally adjusted for education, history of cardiovascular disease, depression, and white blood cells.

https://doi.org/10.1371/journal.pone.0301979.g002

Discussion

We developed a model using statistical and machine learning methods that predicted impaired cognition using ten known risk factors and validated it in a replication dataset among free-living older adults. Information on the risk factors in our model is commonly available in electronic health records but is not systematically used to predict impaired cognition. The model could be embedded in an electronic health record system to predict the likelihood of impaired cognition for individuals using information in their respective health records, which could assist their health care providers to decide on the next steps. The positive predictive value would be increased for all tools used to evaluate impaired cognition by screening individuals identified as high-risk using our model, thus facilitating the diagnosis of impaired cognition among older adults.

Although the causes of cognitive impairment are not yet completely understood, evidence of associations between individual risk factors and cognitive loss is accumulating. Besides increasing age, low education level is associated with higher risk of cognition impairment [21], Some unhealthy lifestyles, including smoking, lack of physical activity, unhealthy eating are potential risk factors [22]. In addition, people with diabetes, obesity, depression, hypertension, and hyperlipidemia have an increased risk of developing cognition decline [23]. However, few studies focused on a combination of these risk factors. Kivipelto and colleagues developed the CAIDE impaired cognition score based on modifiable risk factors that predicted the incidence of dementia over 20 years [24]. Combinations of risk factors have been shown to predict dementia incidence over time in other prospective studies [5, 2527]. In contrast, we combined information on risk factors to predict the prevalence of impaired cognition. Our goal was to facilitate case identification at one moment in time rather than predicting dementia risk over time. Tolea and colleagues using a modified CAIDE instrument in a cross-sectional analysis reported an AUC of 0.63 to predict mild cognitive impairment versus no impairment [28]. In comparison, AUC estimates to assess impaired cognition in our study ranged from 0.70 to 0.79 in various models.

The USPTF does not recommend routine screening for cognitive impairment in older adults (≥65 years) [29] due to the insignificant clinical benefit. Screening tools are effective in identifying cognitive impairment, do not cause any harm, and but benefits from subsequent interventions are modest [30]. Early diagnosis of impaired cognition may nevertheless benefit individuals and their caregivers to manage and monitor the condition and plan for the future, and health care providers to focus on reversing modifiable risk factors and making appropriate treatment plans. Among nationally representative older US adults, 12 potentially modifiable risk factors accounted for 41% of dementia cases overall and 46% and 47% of dementia cases among African Americans and Hispanics respectively [31]. Another analysis of a nationally representative sample of US adults identified 8 modifiable risk factors that contributed to 35.9% and 30.1% of ADRD cases in men and women respectively, 40% of cases among African Americans, and 34% of cases among Hispanics [32]. Additionally, it has been shown that pathologic changes in the brain may precede development of symptoms by several years [33]. Together these studies suggest that managing risk factors early, particularly in high-risk groups, may be beneficial in reducing disease burden. Our model identifies individuals likely to have impaired cognition which could help health care providers to diagnose impaired cognition.

Our study has several strengths. We applied different methods (traditional logistic regression and SVM models) to derive robust and stable models. Though a systematic review found no benefit in machine learning over logistic regression in clinical prediction [34], machine learning may perform better than logistic regression in smaller samples [35], and with high dimension data [36]. The reason for using both logistic regression and SVM models was to increase the chances that important variables would be included in the final prediction model. Though we used logistic regression as the final predictive model, at the outset we did not know which variables would be selected by which method. Using both methods to identify risk factor clusters may therefore have been an advantage. Additionally, the training and validation datasets were from samples representative of free-living older US adults. However, our study also has some limitations. For example, the predictive properties of our model are in the acceptable range but are not excellent [37]. One possible reason for this is that the NHANES datasets were not specifically designed for evaluating cognition. The model could be refined in a dataset with more accurate measures of cognition such as more extensive neurocognitive testing, and neuroimaging such as MRI and EEG [38]. Another way to strengthen the model is by evaluating a wider range of risk factors such as sleep, physical function, speech and hearing abilities, vision, reading history, access to healthcare, social interactions, family engagement, and oral health [3842]. Many of these factors have been related with cognition but are not formally used to predict undiagnosed cognitive impairment. For example, though poor oral health is associated with worse cognition, oral health information is not widely used to predict impaired cognition. Another limitation is that the statistical approach used to identify risk factors assumes that the model assumptions are met. For example, to identify the initial set of possible predictors we used logistic regression, which assumes that the relation between continuous risk factors and the log odds of the outcome are linearly related. A violation of this assumption could result in loss of efficiency. To address this concern, we repeated the variable selection process using machine learning methods which do not have this limitation. The final list of predictors was obtained using information from both logistic regression and machine learning methods. Another limitation was that some domains were measured differently in the development (NHANES III) and validation datasets (continuous NHANES 2011–2014). For example, the cognitive score assessed in the validation dataset used newer methods. Though we used the same overall criteria to evaluate cognitive impairment (lowest 10% of cognition score) in both the development and validation datasets, cognitive assessment in the latter dataset was likely more accurate. This could also have reduced the predictive properties of the model. Despite these limitations, the approach seems promising for clinical translation. For example, based on the predicted probability of cognitive impairment obtained for a particular patient from the model, a health care provider could decide whether to test that patient.

Conclusions

We developed a model consisting of 10 common risk factors that predicted undiagnosed cognitive impairment in a representative sample of older US adults. This model could be used to identify groups with greater prevalence of cognitive impairment in its current form to aid in screening. This model could be enhanced, and its prediction improved by refining it in a richer dataset with more accurate and varied outcome measures and a broader range of predictors.

References

  1. 1. Chodosh J, Petitti DB, Elliott M, Hays RD, Crooks VC, Reuben DB, et al. Physician recognition of cognitive impairment: evaluating the need for improvement. J Am Geriatr Soc. 2004;52(7):1051–9. pmid:15209641.
  2. 2. Olafsdottir M, Skoog I, Marcusson J. Detection of dementia in primary care: the Linkoping study. Dement Geriatr Cogn Disord. 2000;11(4):223–9. pmid:10867449.
  3. 3. Valcour VG, Masaki KH, Curb JD, Blanchette PL. The detection of dementia in the primary care setting. Arch Intern Med. 2000;160(19):2964–8. pmid:11041904.
  4. 4. 410.15 Annual wellness visits providing Personalized Prevention Plan Services: Conditions for and limitations on coverage: National Archives; 2023 [cited 2023 02/11/2023]. Available from: https://www.ecfr.gov/current/title-42/chapter-IV/subchapter-B/part-410/subpart-B/section-410.15.
  5. 5. Shang X, Zhu Z, Zhang X, Huang Y, Zhang X, Liu J, et al. Association of a wide range of chronic diseases and apolipoprotein E4 genotype with subsequent risk of dementia in community-dwelling adults: A retrospective cohort study. EClinicalMedicine. 2022;45:101335. Epub 2022/03/19. pmid:35299656; PubMed Central PMCID: PMC8921546.
  6. 6. Ballard C, Gauthier S, Corbett A, Brayne C, Aarsland D, Jones E. Alzheimer’s disease. Lancet. 2011;377(9770):1019–31. pmid:21371747.
  7. 7. Kuehn BM. In Alzheimer Research, Glucose Metabolism Moves to Center Stage. JAMA. 2020;323(4):297–9. pmid:31913419.
  8. 8. Dominy SS, Lynch C, Ermini F, Benedyk M, Marczyk A, Konradi A, et al. Porphyromonas gingivalis in Alzheimer’s disease brains: Evidence for disease causation and treatment with small-molecule inhibitors. Sci Adv. 2019;5(1):eaau3333. pmid:30746447; PubMed Central PMCID: PMC6357742.
  9. 9. Singh SD, Oreskovic T, Carr S, Papier K, Conroy M, Senff JR, et al. The predictive validity of a Brain Care Score for dementia and stroke: data from the UK Biobank cohort. Front Neurol. 2023;14:1291020. Epub 2023/12/18. pmid:38107629; PubMed Central PMCID: PMC10725202.
  10. 10. NHANES III: Antibodies to Periodontal Pathogens 2008 [updated 2008]. Available from: https://wwwn.cdc.gov/nchs/data/nhanes3/30a/spsdeppx.pdf.
  11. 11. Centers for Disease Control and Prevention (CDC). National Center for Health Statistics (NCHS). National Health and Nutrition Examination Survey Data. Hyattsville, MD: U.S. Department of Health and Human Services, Centers for Disease Control and Prevention; 2011–2014 [cited 2022 October 20, 2022]. Available from: https://wwwn.cdc.gov/nchs/nhanes/continuousnhanes/default.aspx?BeginYear=2011.
  12. 12. National Health and Nutrition Examination Survey III: Oral Examination Component. Rockville, MD: 1992 March 1992. Report No.
  13. 13. Eke PI, Dye BA, Wei L, Slade GD, Thornton-Evans GO, Borgnakke WS, et al. Update on Prevalence of periodontitis in adults in the United States: NHANES 2009 to 2012. J Periodontol. 2015;86(5):611–22. Epub 02/17. pmid:25688694.
  14. 14. Obisesan TO, Gillum RF. Cognitive function, social integration and mortality in a U.S. national cohort study of older adults. BMC Geriatr. 2009;9:33. pmid:19638207; PubMed Central PMCID: PMC2724371.
  15. 15. Pfeiffer E. A short portable mental status questionnaire for the assessment of organic brain deficit in elderly patients. J Am Geriatr Soc. 1975;23(10):433–41. pmid:1159263.
  16. 16. Perkins AJ, Hendrie HC, Callahan CM, Gao S, Unverzagt FW, Xu Y, et al. Association of antioxidants with memory in a multiethnic elderly sample using the Third National Health and Nutrition Examination Survey. Am J Epidemiol. 1999;150(1):37–44. pmid:10400551.
  17. 17. Pantell M, Rehkopf D, Jutte D, Syme SL, Balmes J, Adler N. Social isolation: a predictor of mortality comparable to traditional clinical risk factors. Am J Public Health. 2013;103(11):2056–62. Epub 2013/09/14. pmid:24028260; PubMed Central PMCID: PMC3871270.
  18. 18. National Health and Nutrition Examination Survey, 2011–2012 Data Documentation, Codebook, and Frequencies, Cognitive Functioning (CFQ_G). Hyattsville MD: 2017.
  19. 19. Morris JC, Heyman A, Mohs RC, Hughes JP, van Belle G, Fillenbaum G, et al. The Consortium to Establish a Registry for Alzheimer’s Disease (CERAD). Part I. Clinical and neuropsychological assessment of Alzheimer’s disease. Neurology. 1989;39(9):1159–65. pmid:2771064.
  20. 20. Strauss E, Sherman EMS, Spreen O, Spreen O. A compendium of neuropsychological tests: administration, norms, and commentary. 3rd ed. Oxford; New York: Oxford University Press; 2006. xvii, 1216 p. p.
  21. 21. Andrews SJ, Fulton-Howard B, O’Reilly P, Marcora E, Goate AM, collaborators of the Alzheimer’s Disease Genetics C. Causal Associations Between Modifiable Risk Factors and the Alzheimer’s Phenome. Ann Neurol. 2021;89(1):54–65. Epub 2020/10/01. pmid:32996171; PubMed Central PMCID: PMC8088901.
  22. 22. Zhang XX, Tian Y, Wang ZT, Ma YH, Tan L, Yu JT. The Epidemiology of Alzheimer’s Disease Modifiable Risk Factors and Prevention. J Prev Alzheimers Dis. 2021;8(3):313–21. Epub 2021/06/09. pmid:34101789.
  23. 23. Frisardi V, Solfrizzi V, Seripa D, Capurso C, Santamato A, Sancarlo D, et al. Metabolic-cognitive syndrome: a cross-talk between metabolic syndrome and Alzheimer’s disease. Ageing Res Rev. 2010;9(4):399–417. Epub 2010/05/07. pmid:20444434.
  24. 24. Kivipelto M, Ngandu T, Laatikainen T, Winblad B, Soininen H, Tuomilehto J. Risk score for the prediction of dementia risk in 20 years among middle aged people: a longitudinal, population-based study. The Lancet Neurology. 2006;5(9):735–41. Epub 2006/08/18. pmid:16914401.
  25. 25. Exalto LG, Quesenberry CP, Barnes D, Kivipelto M, Biessels GJ, Whitmer RA. Midlife risk score for the prediction of dementia four decades later. Alzheimer’s & dementia: the journal of the Alzheimer’s Association. 2014;10(5):562–70. Epub 2013/09/17. pmid:24035147.
  26. 26. Ngandu T, Helkala EL, Soininen H, Winblad B, Tuomilehto J, Nissinen A, et al. Alcohol drinking and cognitive functions: findings from the Cardiovascular Risk Factors Aging and Dementia (CAIDE) Study. Dement Geriatr Cogn Disord. 2007;23(3):140–9. Epub 2006/12/16. pmid:17170526.
  27. 27. You J, Zhang YR, Wang HF, Yang M, Feng JF, Yu JT, et al. Development of a novel dementia risk prediction model in the general population: A large, longitudinal, population-based machine-learning study. EClinicalMedicine. 2022;53:101665. Epub 2022/10/04. pmid:36187723; PubMed Central PMCID: PMC9519470.
  28. 28. Tolea MI, Heo J, Chrisphonte S, Galvin JE. A Modified CAIDE Risk Score as a Screening Tool for Cognitive Impairment in Older Adults. J Alzheimers Dis. 2021;82(4):1755–68. pmid:34219721; PubMed Central PMCID: PMC8483620.
  29. 29. Force USPST, Owens DK, Davidson KW, Krist AH, Barry MJ, Cabana M, et al. Screening for Cognitive Impairment in Older Adults: US Preventive Services Task Force Recommendation Statement. JAMA. 2020;323(8):757–63. pmid:32096858.
  30. 30. Patnode CD, Perdue LA, Rossom RC, Rushkin MC, Redmond N, Thomas RG, et al. Screening for Cognitive Impairment in Older Adults: Updated Evidence Report and Systematic Review for the US Preventive Services Task Force. JAMA. 2020;323(8):764–85. pmid:32096857.
  31. 31. Lee M, Whitsel E, Avery C, Hughes TM, Griswold ME, Sedaghat S, et al. Variation in Population Attributable Fraction of Dementia Associated With Potentially Modifiable Risk Factors by Race and Ethnicity in the US. JAMA Netw Open. 2022;5(7):e2219672. pmid:35793088; PubMed Central PMCID: PMC9260480.
  32. 32. Nianogo RA, Rosenwohl-Mack A, Yaffe K, Carrasco A, Hoffmann CM, Barnes DE. Risk Factors Associated With Alzheimer Disease and Related Dementias by Sex and Race and Ethnicity in the US. JAMA Neurol. 2022;79(6):584–91. pmid:35532912; PubMed Central PMCID: PMC9086930.
  33. 33. Guzman-Velez E, Diez I, Schoemaker D, Pardilla-Delgado E, Vila-Castelar C, Fox-Fuller JT, et al. Amyloid-beta and tau pathologies relate to distinctive brain dysconnectomics in preclinical autosomal-dominant Alzheimer’s disease. Proc Natl Acad Sci U S A. 2022;119(15):e2113641119. pmid:35380901; PubMed Central PMCID: PMC9169643.
  34. 34. Christodoulou E, Ma J, Collins GS, Steyerberg EW, Verbakel JY, Van Calster B. A systematic review shows no performance benefit of machine learning over logistic regression for clinical prediction models. J Clin Epidemiol. 2019;110:12–22. Epub 2019/02/15. pmid:30763612.
  35. 35. Balea-Fernandez FJ, Martinez-Vega B, Ortega S, Fabelo H, Leon R, Callico GM, et al. Analysis of Risk Factors in Dementia Through Machine Learning. J Alzheimers Dis. 2021;79(2):845–61. Epub 2020/12/29. pmid:33361594.
  36. 36. Feng JZ, Wang Y, Peng J, Sun MW, Zeng J, Jiang H. Comparison between logistic regression and machine learning algorithms on survival prediction of traumatic brain injuries. J Crit Care. 2019;54:110–6. Epub 2019/08/14. pmid:31408805.
  37. 37. Mandrekar JN. Receiver operating characteristic curve in diagnostic test assessment. J Thorac Oncol. 2010;5(9):1315–6. pmid:20736804.
  38. 38. Newman-Norlund RD, Newman-Norlund SE, Sayers S, Nemati S, Riccardi N, Rorden C, et al. The Aging Brain Cohort (ABC) repository: The University of South Carolina’s multimodal lifespan database for studying the relationship between the brain, cognition, genetics and behavior in healthy aging. Neuroimage Rep. 2021;1(1).
  39. 39. Duchowny KA, Ackley SF, Brenowitz WD, Wang J, Zimmerman SC, Caunca MR, et al. Associations Between Handgrip Strength and Dementia Risk, Cognition, and Neuroimaging Outcomes in the UK Biobank Cohort Study. JAMA Netw Open. 2022;5(6):e2218314. pmid:35737388; PubMed Central PMCID: PMC9227006.
  40. 40. Daly B, Thompsell A, Sharpling J, Rooney YM, Hillman L, Wanyonyi KL, et al. Evidence summary: the relationship between oral health and dementia. Br Dent J. 2018;223(11):846–53. pmid:29192686.
  41. 41. Delwel S, Binnekade TT, Perez R, Hertogh C, Scherder EJA, Lobbezoo F. Oral hygiene and oral health in older people with dementia: a comprehensive review with focus on oral soft tissues. Clin Oral Investig. 2018;22(1):93–108. Epub 2017/11/17. pmid:29143189; PubMed Central PMCID: PMC5748411.
  42. 42. Harding A, Gonder U, Robinson SJ, Crean S, Singhrao SK. Exploring the Association between Alzheimer’s Disease, Oral Health, Microbial Endocrinology and Nutrition. Front Aging Neurosci. 2017;9:398. Epub 2017/12/19. pmid:29249963; PubMed Central PMCID: PMC5717030.