Sage Journals: Discover world-class research

Abstract

Introduction:

Estimating the effects of comorbidities on risk of all-cause dementia (ACD) could potentially better inform prevention strategies and identify novel risk factors compared to more common post-hoc analyses from predictive modeling.

Methods:

In a retrospective cohort study of patients with mild cognitive impairment (MCI) from US Veterans Affairs Medical Centers between 2009 and 2021, we used machine learning techniques from the treatment effect estimation literature to estimate individualized effects of 25 comorbidities (e.g., hypertension) on ACD risk within 10 years of MCI diagnosis. Age and healthcare utilization were adjusted for using exact matching.

Results:

After matching, of 19,797 MCI patients, 6,767 (34.18%) experienced ACD onset. Dyslipidemia (percentage point increase of ACD risk range across different treatment effect estimation techniques = 0.009–0.044), hypertension (range = 0.007–0.043), and diabetes (range = 0.007–0.191) consistently had non-zero average effects.

Discussion:

Our findings support known associations between dyslipidemia, hypertension, and diabetes that increase the risk of ACD in MCI patients, demonstrating the potential for these approaches to identify novel risk factors.

Keywords

Alzheimer’s/dementia quantitative methodology statistical analysis veterans comorbidity

Introduction

All-cause dementia (ACD) is a leading cause of death among individuals 65 years and older, and understanding what contributes to ACD onset in patients with mild cognitive impairment (MCI) could inform treatment and prevention (Alzheimer’s Association, 2024). Past work has shown that intervening on modifiable lifestyle factors, such as diet and exercise, may slow cognitive decline (Rosenberg et al., 2018). Identifying novel risk factors related to ACD using machine learning (ML) can further stimulate hypothesis generation on the disease mechanism, which can aid in designing interventions to reduce ACD risk. However, current work using ML to identify these factors does so by performing post-hoc analyses on models trained to predict ACD onset (Irwin et al., 2024; Jo et al., 2019; Tang et al., 2024; Tjandra et al., 2020). For example, past studies have used datasets such as the Alzheimer’s Disease Neuroimaging Initiative (Devarakonda et al., 2019; Jo et al., 2019) and institutional electronic health records (EHRs; Irwin et al., 2024; Tang et al., 2024; Tjandra et al., 2020) to predict onset of Alzheimer’s disease and then identified risk factors as the features whose contribution to predictive performance was significant. In contrast, we aim to directly estimate the individual effects of common comorbidities on the risk of ACD rather than measuring post-hoc feature importance. If the effect is strong, then treating the comorbidity could reduce ACD risk.

Ideally, we aim to identify comorbidities that cause ACD onset. However, verifying whether the relationship between a comorbidity (e.g., hypertension) and ACD onset is causal requires a randomized controlled trial (RCT), and for many potential risk factors, an RCT is infeasible. In light of this, we investigate the applicability of ML approaches to an observational EHR cohort to estimate the effects of comorbidities identified by the literature as risk factors for ACD onset (e.g., hypertension and hearing loss). The approaches we use generally aim to quantify the effect of a feature (e.g., hypertension) on an outcome (e.g., ACD onset) from observational data. Under a set of standard assumptions, outlined below, the estimated effects can be interpreted as the amount by which the feature changes the outcome or risk of the outcome. Such knowledge can provide a focused set of hypotheses for future work in the clinical space to test. In past work, these approaches have been used in assessing treatment effects (Xu, Ignatiadis et al., 2023). In the context of ACD onset, we aim to quantify how much known risk factors change the risk of ACD onset over a 10-year horizon. Here, we measure the change in risk of ACD onset (i.e., the estimated effect) as the percentage point increase of having the risk factor compared to not having the risk factor. In theory, measuring the estimated effect allows us to come closer to a causal investigation than post-hoc analyses from predictive modeling, which aim to identify factors that are correlated with ACD onset.

Cardiovascular diseases, including hypertension and cerebrovascular disease, are among the most commonly studied risk factors for ACD onset (Alzheimer’s Association, 2024). Since adequate heart health is required to deliver oxygen to the brain, researchers hypothesize that comorbidities adversely affecting the heart also adversely affect the brain (Kuźma et al., 2018; Mergenthaler et al., 2013). Similarly, factors like smoking can have a negative effect on heart health (Wells, 1994), thus affecting ACD progression in similar ways. Though we cannot test these hypotheses directly, we can check whether our findings are consistent with the clinical literature across different approaches from the treatment effect estimation literature.

In this paper, we identified a cohort of patients with MCI and used ML to estimate the effects of recognized risk factors on the risk of conversion from MCI to ACD at a 10-year horizon as a proof of concept (Beydoun et al., 2008; Choi et al., 2018; Dunietz et al., 2021; Hulse et al., 2005; Newman et al., 2005; Stefanidis et al., 2018; Tamura & Yaffe, 2011; Thomson et al., 2017). Based on our results, we suggest potential mechanisms for how these factors could contribute to ACD conversion. Going forward, these approaches can potentially be used by researchers in dementia to guide clinical research by investigating the effects of novel risk factors on dementia risk in observational data, leading to novel hypotheses that can be used to inform future work and interventions.

Methods

Study Cohort

We included patients with MCI, as defined by the VINCI (VA Informatics and Computing Infrastructure) CIPHER (Centralized Interactive Phenomics Resource) criteria (Honerlaw et al., 2023), from the Veterans Affairs’ (VA) Cerner EHR instance (Cerner Corporation, North Kansas City, MO; VA Informatics and Computing Infrastructure [VINCI], 2008) who had an encounter with any of the 172 VA facilities in the United States between January 1, 2009 and December 31, 2021. Patient timelines were aligned at the first diagnosis of MCI (i.e., MCI onset). We excluded patients with an MCI or ACD diagnosis before 50 years of age, patients who converted to ACD less than 6 months after their MCI diagnosis, and patients with less than 1 year of historical data prior to MCI diagnosis. ACD diagnoses were also defined as described by the VINCI CIPHER criteria (Honerlaw et al., 2023; more detail in Appendix A1). For both conditions, the VINCI CIPHER criteria identified diagnoses based whether the patient was billed for specific diagnostic codes in the EHR. To control for the effects of age at MCI diagnosis and healthcare utilization on the risk of ACD conversion, we matched patients across each time to conversion (e.g., conversion after 1 year) and time of censoring (e.g., lost to follow-up after 1 year) by age of MCI diagnosis and number of BMI measurements within the 5 years leading up to MCI diagnosis (more detail in Appendix A2). Here, the number of BMI measurements acted as a surrogate for healthcare utilization since we assumed that patients generally have their BMI measured during routine clinical encounters. We controlled for these so that our predictions would not be dominated by these factors (e.g., the model mainly uses age at MCI onset to predict ACD risk). This study was carried out between August 2023 and August 2024 and was approved by the Institutional Review Board of the Phoenix Veterans Affairs Health Care Center (protocol Migrino 1593816).

Risk Factors Considered

We considered comorbidities that were diagnosed before MCI onset. Since cardiovascular comorbidities have been identified by the literature as risk factors for ACD, we considered comorbidities like hypertension and cerebrovascular disease (Alzheimer’s Association, 2024). Similarly, since smoking can adversely effect heart health (Wells, 1994), we also considered it as a risk factor. Outside of cardiovascular comorbidities, TBI (traumatic brain injury) is often studied in the context of ACD onset (Logue et al., 2023; Vincent et al., 2014), where TBI has been shown to be associated with increased risk (Alzheimer’s Association, 2024). Additionally, comorbidities affecting mental health (e.g., anxiety and depression) and psychological trauma (e.g., post-traumatic stress disorder [PTSD]) (Logue et al., 2023; Prieto et al., 2023; Yaffe et al., 2009) have been suggested to be associated with increased dementia risk (Byers & Yaffe, 2011; Desmarais et al., 2020; Gardner et al., 2014; Kwak et al., 2017), and thus, were included in our study.

We limited our focus to mid- to late-life modifiable risk factors that can be identified in the EHR to highlight comorbidities that could guide future research for designing risk-reducing interventions. Thus, we did not consider factors like genetics, demographics, education, and socioeconomic status (SES) since they are either (1) not intervenable or (2) not observable in the EHR. However, since they are potential confounders to our study, we considered them as features during model training to account for their effects on the risk of ACD onset. Only demographics were observable in the EHR, so we relied on downstream variables to capture the effects of features like SES as described below.

We assumed that these risk factors are related to ACD onset as shown in Figure 1, where the risk factors we considered are highlighted in blue. Figure 1 (described more in Appendix A3) was constructed based on our literature search of ACD risk factors and aims to explicitly state which risk factors we are considering and what we assume the potential confounders are. Since we considered comorbidities identified with diagnostic billing codes, they are likely confounded by variables like genetics and SES. These variables, in turn, affect related vital sign measurements and laboratory test results. Thus, we included them as confounders. For unobservable variables like education, genetics, and SES, we indirectly accounted for them by relying on downstream variables to capture their effects, such as healthcare utilization, ZIP codes, vital sign measurements, and laboratory test results. If we assume that these relationships hold, then we can apply the approaches outlined below.

Figure 1.

Assumed feature relations. We assume that the conditions we examine are related to each other and ACD onset as shown in the figure in patients with MCI. Arrows show the assumption that the source feature (or feature group) has a causal contribution to the destination feature (or feature group). Blue boxes: comorbidities assumed to contribute to ACD onset. Green boxes with dashed borders: observed confounders. Yellow boxes with dotted borders: potential unobserved confounders whose downstream effects were assumed to be captured by other variables. Dark gray boxes: feature groups. Arrows pointing at a dark gray box point to all features in the box. Arrows coming from a dark gray box are such that there is an arrow coming from each feature in the box.

Data Preprocessing

We extracted 114 covariates relating to the comorbidities mentioned above as well as potential confounders such as demographics, medications, vital signs, laboratory tests, and healthcare utilization from up to 5 years before MCI onset (more detail in Table 1 and Appendix A4).

Table 1.

Cohort characteristics and feature breakdown. We show the characteristics at alignment (i.e., MCI onset). Some patients did not have a race or ethnicity recorded. We report race and ethnicity categories as they are recorded.

Characteristic: N (%) or Median [IQR]		N = 19,797
Demographics (16 features)	Female	774 (3.91%)
	Male	19,023 (96.09%)
	White	15,307 (77.32%)
	Black	2,843 (14.36%)
	Declined to answer	612 (3.09%)
	Two or more races	365 (1.84%)
	Race unknown	186 (0.94%)
	Hawaiian/Pacific Islander	171 (0.86%)
	Asian	138 (0.70%)
	American Indian/Alaskan	126 (0.64%)
	White not of Hispanic Origin	27 (0.14%)
	Non-Hispanic	17,976 (90.80%)
	Hispanic	1,139 (5.75%)
	Decline to answer	306 (1.55%)
	Ethnicity unknown	250 (1.26%)
	High risk ZIP code	2,224 (11.23%)
Cardiovascular comorbidities (6 features)	Hypertension	11,427 (57.72%)
	Cerebrovascular disease	2,833 (14.31%)
	Peripheral vascular disorders	1,983 (10.02%)
	Heart failure	918 (4.64%)
	Myocardial infarction	761 (3.84%)
	Pulmonary circulation disorders	349 (1.76%)
Substance abuse comorbidities (3 features)	Smoking	3,725 (18.82%)
	Alcohol abuse	2,099 (10.60%)
	Drug abuse	1,314 (6.64%)
Mental health comorbidities (4 features)	Depression	6,158 (31.11%)
	Anxiety/related disorders	5,896 (29.78%)
	Post-traumatic stress disorder	3,302 (16.68%)
	Delirium	113 (0.57%)
Other comorbidities (14 features)	Dyslipidemia	11,421 (57.69%)
	Diabetes	5,477 (27.67%)
	Hearing loss	6,514 (32.90%)
	Obesity	4,113 (20.78%)
	Hearing aids	2,699 (13.63%)
	Sleep apnea	1,819 (9.19%)
	Renal failure	1,450 (7.32%)
	Myopia	1,321 (6.67%)
	Gout	1,100 (5.56%)
	Weight loss	864 (4.36%)
	Traumatic brain injury	770 (3.89%)
	Liver disease	762 (3.85%)
	Urine stones	656 (3.31%)
	Coagulopathy	481 (2.43%)
Age and healthcare utilization (15 features)	Age of MCI onset	70 [65–78]
	Number of outpatient encounters	4 [2–6]
	Number of inpatient encounters	0 [0–0]
Most recent vital signs & laboratory tests (51 features)	Systolic blood pressure	130 mmHg [120–140]
	Diastolic blood pressure	75 mmHg [68–81]
	Body mass index	27.97 kg/m² [24.91-31.55]
	Creatinine	1.02 mg/dL [0.90-1.22]
	Hemoglobin A1c	5.9% [5.5-6.7]
	Low density lipoprotein cholesterol	88.8 mg/dL [74.0-113.0]
	High density lipoprotien cholesterol	44.0 mg/dL [37.0-54.0]
	Aspartate transaminase	22 U/L [18–27]
	Alanine transaminase	23 U/L [17–31]
Medications (5 features)	Cardiovascular medications	15,313 (77.35%)
	Depression medications	7,079 (35.76%)
	Diabetes medications	5,139 (25.96%)
	Anxiety medications	3,639 (18.38%)
	Dyslipidemia medications	865 (4.37%)

N = number; IQR = interquartile range.

Comorbidity Effect Estimation

Given the comorbidities and patient covariates, we used ML techniques from the treatment effect estimation literature to estimate the effect of each comorbidity on the risk of ACD onset. We considered the probability of ACD conversion within 10 years of MCI onset as our outcome, and we trained ML models to predict time to ACD onset at the time of MCI onset given patient covariates (Wang et al., 2019). To estimate effects, we estimated the difference in probability of ACD onset within 10 years of MCI onset in the presence and absence of each comorbidity within each patient and then averaged over all patients (more detail in Appendix A5; Rubin, 2005). In summary, the average effect can be interpreted as the percentage point change in the probability of ACD onset within 10 years of MCI onset resulting from the comorbidity.

Model Training

For each comorbidity, we estimated the effects using common approaches from the treatment effect estimation literature, such as the X, R, and DR metalearners (see Appendix A6 for more detail; Funk et al., 2011; Xu, Ignatiadis et al., 2023). To use these approaches, we make the following three assumptions, as is standard in the treatment effect estimation literature (VanderWeele, 2009; Xu, Ignatiadis et al., 2023). The first is overlap: for a comorbidity of interest, the probability of any patient in the dataset having that comorbidity is non-zero. The second is unconfoundedness: the expected outcome given the presence or absence of the comorbidity (e.g., the expected probability of ACD onset in the presence of hypertension) is independent of whether the comorbidity is present, conditioned on patient covariates (i.e., all confounders are included in the covariate set). The third is stable treatment units and consistency: a patient’s observed outcome does not depend on another patient’s outcome, and for all patients, the observed outcome is the outcome directly resulting from the given set of features and whether the comorbidity is present. In making these assumptions, we can use treatment effect estimation approaches to train ML models using observational data and still recover the average effects. This is because the assumptions ensure that what the models learn from patients without the comorbidity, generalize to patients with the comorbidity and vice versa. We test for overlap as described in the preliminary analysis. Unconfoundedness holds if the assumed relationships outlined in Figure 1 hold and there are no additional confounders not shown in Figure 1. We cannot explicitly test for consistency or stable treatment units, but it remains a reasonable assumption given our current understanding of the disease process.

Model Evaluation

We first conducted a preliminary analysis to investigate the overlap assumption and whether the approaches perform as expected in a controlled environment. Then, we conducted the main analysis, where we compared the identified predictors of ACD onset from applying post-hoc analyses to a predictive model to the identified predictors from the treatment effect estimation approaches described earlier.

The preliminary analysis is described in Appendix A7. Our main analysis consisted of two parts. In the first, we trained a standard predictive model, reporting its discriminative performance using the time-dependent AUROC (area under the receiver operating characteristic curve; Lambert & Chevret, 2016). Potential predictors were identified with permutation importance (Breiman, 2001) on the standard model using the held-out test set. In the second part, we used the treatment effect estimation approaches to measure the effects of the comorbidities on ACD onset. To evaluate the approaches, we began by measuring the discriminative performance of all intermediate models trained for each approach and each comorbidity using the time-dependent AUROC. Comorbidities whose 95% CI overlapped with 0.5 (i.e., random performance) for at least one model on at least one approach were excluded from further analysis. For each comorbidity that remained, we measured the average effect of each comorbidity on ACD onset using each approach by estimating the effect for all individuals in the test set and then taking the average. Here, we do not know what the ground truth effect of each comorbidity is so we can only evaluate whether the results among approaches are consistent (i.e., all of the approaches indicate that the comorbidity increases the risk of ACD onset). Inconsistencies among approaches, such as the effects having different signs (i.e., some approaches indicate that the risk factor increases ACD risk while others indicate a decrease in ACD risk), would indicate that the predicted average effects are more likely to be false discoveries resulting from methodological differences among approaches (Xu, Bechler et al., 2023). For example, the X and R learners may be more sensitive to the accuracy of the estimated propensity scores, an intermediate model output describing the probabilty of having the comorbidity that is used for downstream effect estimation. As such, for each comorbidity, we plotted the average effect for each approach and highlighted which comorbidities had consistent predictions, with error bars representing 95% CIs from 1,000 bootstrapped samples. Note that the features identified by consistent average effects and permutation importance are not guaranteed to be the same (see Appendix A8 for more detail).

Results

Cohort Characteristics

After applying our inclusion/exclusion criteria and matching, our cohort contained 19,797 MCI patients (Figure 2). About 6,767 (34.18%) experienced ACD onset within 10 years, 1,320 (6.67%) did not experience ACD onset within 10 years, and the remaining were right censored (i.e., lost to follow-up less than 10 years after MCI onset without meeting the ACD criteria). The median age of MCI onset was 70 years [IQR (interquartile range) 65–78], 774 (3.91%) were female, 15,307 (77.32%) were White, and the median number of outpatient encounters prior to MCI onset was 4 [IQR 2–6]. The most common comorbidity was hypertension, covering 57.72% of the cohort (11,427 patients). More details are in Table 1.

Figure 2.

Inclusion/exclusion criteria. We begin with all patients with an encounter at any VHA facility between January 1 2009 and December 2021. Numbers in each box correspond to the number of patients included or excluded.

Preliminary Analysis

We investigated the overlap assumption in Appendix A9 by plotting the distributions of propensity scores for each comorbidity. Although the means of the distributions differed among patients who did and did not have the comorbidity (i.e., positives and negatives), the range of scores between positive and negative patients had a considerable amount of overlap for all comorbidities, and the majority of scores were within the range [0.1, 0.9]. We investigated the approaches in a controlled environment through a global null analysis in Appendix A10. The environment was controlled such that a synthetic comorbidity was introduced to the dataset whose average effect was 0. In the results, the estimated effects were close to zero, showing that the approaches performed as expected.

Main Analysis

Part 1

Model performance, as measured by the time-dependent AUROC, was 0.61 (95% CI [0.59, 0.64]). The results from running permutation importance on a standard model using the held-out test set are shown in Appendix A11, where the only comorbidity identified as having a significant effect on performance was anxiety and related disorders, which resulted in a drop of 0.003 (95% CI [0.0003, 0.007]).

Part 2

The discriminative performance of the intermediate models trained for each comorbidity and approach is shown in Appendix A12. For each comorbidity, the performances of the models trained on the negative patients were all significantly better than random (time-dependent AUROC range across comorbidities = 0.61–0.68). For the models trained on the positive patients for each comorbidity, some were not significantly better than random (time-dependent AUROC range = 0.41–0.61). In Figure 3, we show the estimated average effects for the comorbidities whose performance on all models trained for each approach were better than random. Dyslipidemia, hypertension, and diabetes were consistently identified as risk factors by all approaches.

Figure 3.

Average estimated effects. We show the average of the estimated effects of each condition on ACD onset. Error bars represent bootstrapped 95% confidence intervals.

Discussion

In our study, we identified hypertension, dyslipidemia, and diabetes as risk factors for ACD using approaches from the treatment effect estimation literature. While EHR data have been used by previous work to identify ACD risk factors, many focus on post-hoc analyses from predictive modeling instead of directly estimating the individual effect of each comorbidity (Irwin et al., 2024; Jo et al., 2019; Tang et al., 2024; Tjandra et al., 2020). For example, these studies identify potential risk factors by measuring feature importance post-hoc using approaches like permutation importance. While these approaches may indicate which features the model uses to make its predictions (i.e., which features may have a higher correlation with the outcome), they do not necessarily indicate which features could inform prevention. From our analysis using permutation importance, we found that the features identified differed from those identified by consistent average effects across approaches. There are many reasons why this could occur. For example, the effects may not be large enough to significantly affect discriminative performance. This may be why hypertension, dyslipidemia, and diabetes were identified by the treatment effect estimation approaches and not by permutation importance. It is also possible that the effects among individuals within a comorbidity cancel each other out at the population level but can significantly change discriminative performance.

Identifying features that inform prevention requires identifying causal relationships. Verifying causal relationships requires an RCT. However, RCTs cannot be used to investigate the effect of comorbidities on the onset of ACD. We address this gap, in part, through retrospective analyses on observational data using ML techniques. With these techniques, we directly estimate how the presence of a comorbidity could change a patient’s probability of ACD onset within some prediction horizon. While these analyses cannot replace RCTs, we have shown that they can identify risk factors that are consistent with the literature, and thus, have the potential to guide clinical research by suggesting avenues for future research and intervention. By observing similar trends across multiple approaches, we can strengthen our claim on whether the directions of our estimates (i.e., risk or protective) hold.

Approaches from the treatment effect estimation literature require a stricter set of assumptions, but these assumptions allow us to have greater confidence in the accuracy of the predictions of the outcomes in the presence and absence of the comorbidity. This is because these assumptions mean that the model can accurately learn the relationship between the comorbidity and outcome (due to consistency) while accounting for confounding (due to unconfoundedness), and that the learned relationship among patients with the comorbidity will generalize to patients without the comorbidity and vice versa (due to overlap). As a result, we can estimate effects by taking the difference between the predictions of the outcome with and without the comorbidity.

Our finding that dyslipidemia, hypertension, and diabetes are risk factors of ACD onset aligns with the literature (Biessels et al., 2006; Walker et al., 2017; Wee et al., 2023). Dyslipidemia increases the chance of cholesterol buildup in the arteries (Kopin & Lowenstein, 2017), which could limit blood flow to the brain. Hypertension increases the chance of heart disease and stroke (Wajngarten & Silva, 2019), both of which can affect the heart’s ability to supply oxygen to the brain. Insulin resistance from type 2 diabetes has been shown to lower insulin levels in the brain, which may contribute to cognitive decline (Gasparini et al., 2001). Further studies into the mechanism through which these comorbidities contribute to ACD progression could shed light on results from current work (Rosenberg et al., 2018) showing that medication or intervening on modifiable risk factors, such as diet and exercise, may slow the rate of cognitive decline. Notably, it has been proposed that drugs for diabetes and hypertension, two risk factors identified in our analyses, could potentially lower the risk of Alzheimer’s disease, the most common form of dementia (Michailidis et al., 2022; Yasar et al., 2013).

In contrast to hypertension, dyslipidemia, and diabetes, some factors from the literature were not consistently identified by the metalearners. For example, smoking and cerebrovascular disease are associated with ACD onset in similar ways to hypertension, but the approaches did not consistently indicate that they increased the probability of ACD onset over 10 years. This may be because some of the comorbidities could have been affected by unobserved confounding. For example, comorbidities like smoking have been shown to be associated with SES (Hitchman et al., 2014). While we included ZIP codes in our feature set, they only serve as a proxy and may not fully capture SES. Despite hypertension, dyslipidemia, and diabetes also being associated with SES (Blok et al., 2022; Espírito Santo et al., 2022), we hypothesize that including direct measures relating to vital signs and laboratory tests (e.g., blood pressure) could more effectively capture the downstream effects of SES.

Our study has several limitations. First, we relied on EHR-based phenotyping tools to identify MCI, ACD, and the comorbidities we included, which may not always be accurate (Tjandra et al., 2020). Second, since the ground truth average effects are unknown, we could not quantitatively assess whether our estimated effects were correct. Third, we were unable to verify whether all of the assumptions required for the approaches held, such as unconfoundedness. Due to the complicated dynamics of ACD progression, it is likely that there are additional confounders that Figure 1 did not include. Despite our inability to check the correctness of the main analysis, our results support well accepted, plausible hypotheses on what contributes to ACD onset. Fourth, our study only considered one comorbidity at a time and not how comorbidities act in combination. Additionally, the findings from our VA cohort should be validated in the general population. Past work has shown that veterans have a higher prevalence of mental health conditions, thus potentially putting them at higher risk for dementia (Veitch et al., 2013). In addition, even among veterans, these conditions may affect male and female patients differently (Yaffe et al., 2019). In our cohort, the prevalence of PTSD and depression were 17% and 31% respectively, and patients were mostly male (96.09%). Despite these differences in our cohort compared to the general population, a previous study (Tjandra et al., 2022) showed that the performance of using machine learning to predict Alzheimer’s disease onset using blood pressure trajectories trained using VA EHR data was similar when applied to EHR data from another institution, even though the demographic compositions were different. However, it is important to note that generalizability to other demographics should be empirically established. Finally, when controlling for age of MCI onset and healthcare utilization, we used matching, which excluded many patients from our final cohort.

Conclusion

We demonstrated the potential of approaches for estimating treatment effects from observational data in the context of directly estimating the effects of comorbidities on the risk of ACD onset. Results from analyses like ours can be used to inform future work in clinical research on identifying novel risk factors in settings where RCTs are infeasible.

Supplemental Material

sj-docx-1-ggm-10.1177_30495334251347053 – Supplemental material for Estimated Effects of Comorbidities on Risk of All-cause Dementia in Patients with Mild Cognitive Impairment

Supplemental material, sj-docx-1-ggm-10.1177_30495334251347053 for Estimated Effects of Comorbidities on Risk of All-cause Dementia in Patients with Mild Cognitive Impairment by Donna Tjandra, Chase Irwin, Raymond Q. Migrino, Bruno Giordani and Jenna Wiens in Sage Open Aging

Footnotes

ORCID iDs

Donna Tjandra

Chase Irwin

Consent Statement

The IRB at the Phoenix Veterans Affairs approved a waiver of informed consent. Patient consent was waived because, in our retrospective analysis of electronic health record data, there were too many patients to feasibly obtain consent given the size of the study team.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research program is supported by the NIH/NIA funded Michigan Alzheimer’s Disease Center (5P30AG053760), the National Science Foundation (IIS 2124127), and the Phoenix Veterans Affairs Office of Research. The views and conclusions in this document are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the NSF, the NIH, the VA, or the United States Government.

Data Availability Statement

The data underlying this article for the Veterans Affairs cannot be shared due to current policies.

Supplemental Material

Supplemental material for this article is available online.

References

Alzheimer’s Association. (2024). 2024 Alzheimer’s disease facts and figures. Alzheimer’s & Dementia, 20(5), 3708–3821.

Beydoun

M. A.

Beydoun

H. A.

Wang

(2008). Obesity and central obesity as risk factors for incident dementia and its subtypes: A systematic review and meta-analysis. Obesity Reviews, 9(3), 204–218.

Biessels

G. J.

Staekenborg

Brunner

Brayne

Scheltens

(2006). Risk of dementia in diabetes mellitus: A systematic review. Lancet Neurology, 5(1), 64–74.

Blok

Haggenburg

Collard

Van Der Linden

E. L.

Galenkamp

Moll van Charante

E. P.

Agyemang

Van Den Born

B. J. H.

(2022). The association between socioeconomic status and prevalence, awareness, treatment and control of hypertension in different ethnic groups: The healthy life in an urban setting study. Journal of Hypertension, 40(5), 897–907.

Breiman

(2001). Random forests. Machine Learning, 45, 5–32.

Byers

A. L.

Yaffe

(2011). Depression and risk of developing dementia. Nature Reviews Neurology, 7(6), 323–331.

Choi

Park

S. M.

(2018). Effect of smoking cessation on the risk of dementia: A longitudinal study. Annals of Clinical and Translational Neurology, 5(10), 1192–1199.

Desmarais

Weidman

Wassef

Bruneau

M. A.

Friedland

Bajsarowicz

Thibodeau

M. P.

Herrmann

Nguyen

Q. D.

(2020). The interplay between post-traumatic stress disorder and dementia: A systematic review. American Journal of Geriatric Psychiatry, 28(1), 48–60.

Devarakonda

S. T.

J. Y.

Fung

Y. R.

Fiterau

(2019, October 28). FLARe: Forecasting by learning anticipated representations [Conference session]. Machine Learning for Healthcare Conference.

10.

Dunietz

G. L.

Chervin

R. D.

Burke

J. F.

Conceicao

A. S.

Braley

T. J.

(2021). Obstructive sleep apnea treatment and dementia risk in older adults. Sleep, 44(9), zsab076.

11.

Espírito Santo

L. R.

Faria

T. O.

Silva

C. S. O.

Xavier

L. A.

Reis

V. C.

Mota

G. A.

Silveira

M. F.

Mill

J. G.

Baldo

M. P

. (2022). Socioeconomic status and education level are associated with dyslipidemia in adults not taking lipid-lowering medication: A population-based study. International Health, 14(4), 346–353.

12.

Funk

M. J.

Westreich

Wiesen

Stürmer

Brookhart

M. A.

Davidian

(2011). Doubly robust estimation of causal effects. American Journal of Epidemiology, 173(7), 761–767.

13.

Gardner

R. C.

Burke

J. F.

Nettiksimmons

Kaup

Barnes

D. E.

Yaffe

(2014). Dementia risk after traumatic brain injury vs nonbrain trauma: The role of age and severity. JAMA Neurology, 71(12), 1490–1497.

14.

Gasparini

Gouras

G. K.

Wang

Gross

R. S.

Beal

M. F.

Greengard

(2001). Stimulation of β-amyloid precursor protein trafficking by insulin reduces intraneuronal β-amyloid and requires mitogen-activated protein kinase signaling. Journal of Neuroscience, 21(8), 2561–2570.

15.

Hitchman

S. C.

Fong

G. T.

Zanna

M. P.

Thrasher

J. F.

Chung-Hall

Siahpush

(2014). Socioeconomic status and smokers’ number of smoking friends: Findings from the International Tobacco Control (ITC) Four Country Survey. Drug and Alcohol Dependence, 143, 158–166.

16.

Honerlaw

Y. L.

Fontin

Gosian

Maripuri

Murray

Sangar

Galloway

Zimolzak

A. J.

Whitbourne

S. B.

Casas

J. P.

Ramoni

R. B.

Gagnon

D. R.

Cai

Liao

K. P.

Gaziano

J. M.

Muralidhar

Cho

(2023). Framework of the Centralized Interactive Phenomics Resource (CIPHER) standard for electronic health data-based phenomics knowledgebase. Journal of the American Medical Informatics Association, 30(5), 958–964.

17.

Hulse

G. K.

Lautenschlager

N. T.

Tait

R. J.

Almeida

O. P.

(2005). Dementia associated with alcohol and other drug use. International Psychogeriatrics, 17(s1), S109–S127.

18.

Irwin

Tjandra

Aggarwal

Lienau

Giordani

Wiens

Migrino

R. Q.

(2024). Predicting 5-year dementia conversion in veterans with mild cognitive impairment. Alzheimer’s & Dementia Diagnosis Assessment & Disease Monitoring, 16(1), e12572.

19.

Nho

Saykin

A. J.

(2019). Deep learning in Alzheimer’s disease: Diagnostic classification and prognostic prediction using neuroimaging data. Frontiers in Aging Neuroscience, 11, 220.

20.

Kopin

Lowenstein

C. J.

(2017). Dyslipidemia. Annals of Internal Medicine, 167(11), ITC81–ITC96.

21.

Kuźma

Lourida

Moore

S. F.

Levine

D. A.

Ukoumunne

O. C.

Llewellyn

D. J.

(2018). Stroke and dementia risk: A systematic review and meta-analysis. Alzheimer’s & Dementia, 14(11), 1416–1426.

22.

Kwak

Y. T.

Yang

Koo

M. S.

(2017). Anxiety in dementia. Dementia and Neurocognitive Disorders, 16(2), 33–39.

23.

Lambert

Chevret

(2016). Summary measure of discrimination in survival models based on cumulative/dynamic time-dependent ROC curves. Statistical Methods in Medical Research, 25(5), 2088–2102.

24.

Logue

M. W.

Miller

M. W.

Sherva

Zhang

Harrington

K. M.

Fonda

J. R.

Merritt

V. C.

Panizzon

M. S.

Hauger

R. L.

Wolf

E. J.

Neale

Gaziano

J. M.

(2023). Alzheimer’s disease and related dementias among aging veterans: Examining gene-by-environment interactions withpost-traumatic stress disorder and traumatic brain injury. Alzheimer’s & Dementia, 19(6), 2549–2559.

25.

Mergenthaler

Lindauer

Dienel

G. A.

Meisel

(2013). Sugar for the brain: The role of glucose in physiological and pathological brain function. Trends in Neurosciences, 36(10), 587–597.

26.

Michailidis

Tata

D. A.

Moraitou

Kavvadas

Karachrysafi

Papamitsou

Vareltzis

Papaliagkas

(2022). Antidiabetic drugs in the treatment of Alzheimer’s disease. International Journal of Molecular Sciences, 23(9), 4641.

27.

Newman

A. B.

Fitzpatrick

A. L.

Lopez

Jackson

Lyketsos

Jagust

Ives

Dekosky

S. T.

Kuller

L. H.

(2005). Dementia and Alzheimer’s disease incidence in relationship to cardiovascular disease in the Cardiovascular Health Study cohort. Journal of the American Geriatrics Society, 53(7), 1101–1107.

28.

Prieto

Nolan

K. E.

Moody

J. N.

Hayes

S. M.

Hayes

J. P.

(2023). Posttraumatic stress symptom severity predicts cognitive decline beyond the effect of Alzheimer’s disease biomarkers in Veterans. Translational Psychiatry, 13(1), 102.

29.

Rosenberg

Ngandu

Rusanen

Antikainen

Bäckman

Havulinna

Hänninen

Laatikainen

Lehtisalo

Levälahti

Lindström

Paajanen

Peltonen

Soininen

Stigsdotter-Neely

Strandberg

Tuomilehto

Solomon

Kivipelto

(2018). Multidomain lifestyle intervention benefits a large elderly population at risk for cognitive decline and dementia regardless of baseline characteristics: The FINGER trial. Alzheimer’s & Dementia, 14(3), 263–270.

30.

Rubin

D. B.

(2005). Causal inference using potential outcomes: Design, modeling, decisions. Journal of the American Statistical Association, 100(469), 322–331.

31.

Stefanidis

K. B.

Askew

C. D.

Greaves

Summers

M. J.

(2018). The effect of non-stroke cardiovascular disease states on risk for cognitive decline and dementia: A systematic and meta-analytic review. Neuropsychology Review, 28(1), 1–15.

32.

Tamura

M. K.

Yaffe

(2011). Dementia and cognitive impairment in ESRD: diagnostic and therapeutic strategies. Kidney International, 79(1), 14–22.

33.

Tang

A. S.

Rankin

K. P.

Cerono

Miramontes

Mills

Roger

Zeng

Nelson

Soman

Woldemariam

Lee

Bove

Glymour

Aghaeepour

Oskotsky

T. T.

Miller

Allen

I. E.

Sanders

S. J.

. . . Sirota

(2024). Leveraging electronic health records and knowledge networks for Alzheimer’s disease prediction and sex-specific biological insights. Nature Aging, 4(3), 379–395.

34.

Thomson

R. S.

Auduong

Miller

A. T.

Gurgel

R. K.

(2017). Hearing loss as a risk factor for dementia: A systematic review. Laryngoscope Investigative Otolaryngology, 2(2), 69–79.

35.

Tjandra

Migrino

R. Q.

Giordani

Wiens

(2020). Cohort discovery and risk stratification for Alzheimer’s disease: An electronic health record-based approach. Alzheimer’s & Dementia Translational Research & Clinical Interventions, 6(1), e12035.

36.

Tjandra

Migrino

R. Q.

Giordani

Wiens

(2022). Use of blood pressure measurements extracted from the electronic health record in predicting Alzheimer’s disease: A retrospective cohort study at two medical centers. Alzheimer’s & Dementia: The Journal of the Alzheimer’s Association, 18(11), 2368–2372.

37.

VA Informatics and Computing Infrastructure (VINCI). (2008). VA HSR RES 13-457, U.S. Department of Veterans Affairs. Retrieved August 24 2023, from https://www.research.va.gov/programs/vinci/

38.

VanderWeele

T. J.

(2009). Concerning the consistency assumption in causal inference. Epidemiology, 20(6), 880–883.

39.

Veitch

D. P.

Friedl

K. E.

Weiner

M. W.

(2013). Military risk factors for cognitive decline, dementia and Alzheimer’s disease. Current Alzheimer Research, 10(9), 907–930.

40.

Vincent

A. S.

Roebuck-Spencer

T. M.

Cernich

(2014). Cognitive changes and dementia risk after traumatic brain injury: Implications for aging military personnel. Alzheimer’s & Dementia, 10, S174–S187.

41.

Wajngarten

Silva

G. S.

(2019). Hypertension and stroke: Update on treatment. European Cardiology Review, 14(2), 111–115.

42.

Walker

K. A.

Power

M. C.

Gottesman

R. F.

(2017). Defining the relationship between hypertension, cognitive decline, and dementia: A review. Current Hypertension Reports, 19, 24–26.

43.

Wang

Reddy

C. K.

(2019). Machine learning for survival analysis: A survey. ACM Computing Surveys, 51(6), 1–36.

44.

Wee

Sukudom

Bhat

Marklund

Peiris

N. J.

Hoyos

C. M.

Patel

Naismith

S. L.

Dwivedi

Misra

(2023). The relationship between midlife dyslipidemia and lifetime incidence of dementia: A systematic review and meta-analysis of cohort studies. Alzheimer‘s & Dementia Diagnosis Assessment & Disease Monitoring, 15(1), e12395.

45.

Wells

A. J.

(1994). Passive smoking as a cause of heart disease. Journal of the American College of Cardiology, 24(2), 546–554.

46.

Bechler

Callahan

Shah

(2023). Principled estimation and evaluation of treatment effect heterogeneity: A case study application to dabigatran for patients with atrial fibrillation. Journal of Biomedical Informatics, 143, 104420.

47.

Ignatiadis

Sverdrup

Fleming

Wager

Shah

(2023). Treatment heterogeneity with survival outcomes. In Zubizarreta

J. R.

Stuart

E. A.

Small

Dylan S.

Rosenbaum

Paul R.

(Eds.), Handbook of matching and weighting adjustments for causal inference (pp. 445–482). Chapman and Hall/CRC.

48.

Yaffe

Lwi

S. J.

Hoang

T. D.

Xia

Barnes

D. E.

Maguen

Peltz

C. B.

(2019). Military-related risk factors in female veterans and risk of dementia. Neurology, 92(3), e205–e211.

49.

Yaffe

Vittinghoff

Lindquist

Barnes

D. E.

Covinsky

K. E.

Neylan

Kluse

Marmar

(2009). Post-traumatic stress disorder and risk of dementia among US veterans. Alzheimer’s & Dementia, 4(5), 104.

50.

Yasar

Xia

Yao

Furberg

C. D.

Xue

Q. L.

Mercado

C. I.

Fitzpatrick

A. L.

Fried

L. P.

Kawas

C. H.

Sink

K. M.

Williamson

J. D.

(2013). Antihypertensive drugs decrease risk of Alzheimer disease: Ginkgo evaluation of memory study. Neurology, 81(10), 1192–1903.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.42 MB