Novel predictors of Alzheimer's disease in Down syndrome identified using machine learning

Abstract

Background

Most adults with Down syndrome will develop Alzheimer's disease (AD) due to the triplication of the amyloid precursor protein in the 21^st chromosome. Predictors of condition onset are less known.

Objective

We used Medicaid and Medicare data and machine learning to identify which co-occurring conditions predict incident AD in United States adults with Down syndrome.

Methods

We examined adults with Down syndrome enrolled in Medicaid and/or Medicare between 2011 and 2019. We identified AD and other conditions using ICD 9 and 10 codes. We used a case-control design with risk set sampling to have that controls to mimic the distribution of times of incident AD. We trained gradient boosted trees to identify strongest predictors.

Results

The cohort had a mean age at entry of 44.6 years, 46.2% were male, and 73.7% were white non-Hispanic. 16,398 had incident AD diagnoses over the study period. The machine learning model had an area under the curve of 0.86 and high positive predictive value. Strongest predictors of increased probability of AD were age, dual Medicaid/Medicare enrollment; incident epilepsy or incident ulcer three years before index date; any hypothyroidism, schizophrenia, or hyperlipidemia. We found synergistic interaction between epilepsy and enrollment by age.

Conclusions

Predictors aligned with known predictors in the general population and characteristics that signal AD symptom onset. New onset epilepsy may be a relevant clinical sign. Identifying these predictors highlights areas for further etiologic inquiry and intervention.

Keywords

Alzheimer's disease deafness Down syndrome epilepsy machine learning

Introduction

Down syndrome is an intellectual and developmental disability caused by the full or partial triplication of chromosome 21. Phenotypically, this often leads to short stature, intellectual disability, and cardiac defects.¹ Because the amyloid precursor protein (a major risk factor for Alzheimer's disease; AD) is also located on chromosome 21, the triplication that defines Down syndrome also leads to a triplication of the APP gene, resulting in excess amyloid which leads to a near certain development of the plaques and tangles associated with AD.² In turn, AD is the leading cause of death among adults with Down syndrome.^3,4

The biologic implication of chromosome 21 and AD is well understood epidemiologically, but other health conditions surrounding AD onset in Down syndrome are not as well understood. In the last forty years there has been a major extension of the lifespan for people with Down syndrome, with a median life expectancy of 4 years in 1950 to 58 years in 2010.⁵ Given this fact, there has been little population-level study of health in the older adult Down syndrome population. In early and middle adulthood, people with Down syndrome develop more chronic health conditions with earlier onset compared to peers without Down syndrome.⁶ Many of these conditions, including diabetes, hearing loss, depression, and obesity, are also risk factors for AD in the general population.⁷ Cross-sectional data show increased prevalence of hypothyroidism, depression, and epilepsy in adults with Down syndrome and AD^8,9; yet it is not known whether these conditions hasten or cause initial AD onset, result from biological and social changes spurred from AD, or are unrelated to AD. Identifying which health conditions precede AD diagnosis can be important for the prevention or early and timely assessments of AD¹⁰ and may create opportunities to research treatment of those conditions as a way to delay AD onset (for example, obstructive sleep apnea treatment delaying onset of cognitive dysfunction in the general population¹¹) or initiate early AD assessment.

There are existing longitudinal cohorts that aim to understand the prodromal development of Down syndrome AD¹²; yet there are inefficiencies to this approach. First, AD on set occurs between 45 and 60 years, meaning a long and expensive observation period. Participation in research studies for people with Down syndrome is often tedious and burdensome with obstacles like limits they can be paid and transportation challenges. Further, when cohorts are developed, they tend to not reflect the socio-demographic distribution of the Down syndrome population.¹³ Adults with Down syndrome in the United States (US) are reliant on public health insurers for their medical care; therefore, we can use Medicaid and Medicare data as a near full sample of US adults with Down syndrome.¹⁴ Therefore, we can leverage longitudinal passively collected Medicaid and Medicare claims data to overcome logistical and demographic limitations of previous work.

A machine learning approach is useful for hypothesis generating,¹⁵ as we can use the data to identify which of the hundreds of potential preceding diagnoses (which reflect identified conditions and access to appropriate healthcare) are most strongly predictive of dementia incidence.¹⁶ In the general population, machine learning has been used in claims data to predict dementia onset,¹⁷ and find novel predictors of onset including incontinence, amnesia, metabolic disorders, and bipolar disorders.¹⁸ This type of work has not yet been done in Down syndrome.

Adults with Down syndrome in the US are reliant on public health insurers for their medical care; therefore, we can use Medicaid and Medicare data as a near full sample of US adults with Down syndrome.¹⁴ Given the large data and the need to understand predictors of AD onset in adults with Down syndrome, our objective was to use nine years of Medicaid and Medicare data and a machine learning approach to identify whether the presence, timing, and combination of co-occurring conditions and demographics predict timing of incident AD diagnosis in adults with Down syndrome in the US.

Methods

Data source

We used data from the Down Syndrome Toward Optimal Trajectories and Health Equity using Medicaid Analytic eXtract (DS-TO-THE-MAX) project. DS-TO-THE-MAX is a longitudinal administrative claims cohort following all adults ≥18 years of age with Down syndrome enrolled in Medicaid and/or Medicare at any point between 2011 and 2019. Data files included inpatient-, outpatient-, and long-term care files for Medicaid and Medicare. Down syndrome was identified by examining claims from any source for ICD-9 code 758.0 or ICD-10 codes Q90.0, Q90.1, Q90.2, and Q90.9. More details on cohort derivation are presented in Rubenstein et al. 2023.^14,19 This project was considered not human subjects research by the Boston University Institutional Review Board.

Sample

We included all adults with ≥1 inpatient or ≥2 outpatient Down syndrome claims in Medicaid and/or Medicare, and at least one year of enrollment time in between 2011–2019, in line with previous work assessing Down syndrome and other intellectual and developmental disabilities.^19,20 We restricted to person years where an individual was ≥30 years old, as AD in Down syndrome rarely presents before age 35.²¹ This study was deemed not human subjects research by the Boston University Medical Campus Institutional Review Board and a waiver of consent was not required.

Outcome identification

We used established algorithms from the Center for Medicare and Medicaid Services (CMS) to identify AD.²¹ These algorithms were developed to assess individuals who have at least three continuous years of claims data and consider an individual as having AD if they have at least one inpatient or two outpatient/carrier claims for AD (see code list, Supplemental Table 1). The AD algorithm has high accuracy and specificity in the general population²² and is likely more reliable in the Down syndrome population due to clinical screening practice which may lead to less false negatives (i.e., never assessed for AD). We examined both the specific (AD) and broader (AD and related disorders of senile dementia) and found that there was not a difference between the two in the Down syndrome sample (i.e., all dementia was AD). Building off previous work using Medicaid and Medicare data for adults with Down syndrome, we relaxed the requirement of three years of enrollment to one year of enrollment (see eligibility criteria above). We do this because for people with Down syndrome, Medicaid turnover is low (e.g., people are consistently enrolled), service use is high,¹⁹ and life-span after Alzheimer diagnosis is short.²³ Therefore, relaxing the reference window means we have greater opportunity to identify an AD claim. This approach does not impart bias when estimating prevalence of AD in the population with Down syndrome.²³

Study design

For our analyses we used a case control design with risk set sampling (i.e., incidence density sampling).²⁴ This approach uses all eligible study entrants and examines outcomes prospectively. The unit of analysis for controls is person-time not persons, so the time before a case becomes a case is eligible to be a control. This more mimics a cohort study where non-cases may eventually become cases.²⁵ In cases the index data was the date of incident AD diagnosis and for non-cases the index data was a time point randomly selected from the distribution of times to incident AD diagnoses among cases, ensuring equal person-time observed for both groups. From the index date, we looked backwards in time to document preceding health conditions, and this approach ensures equal amount of person-time for observation between cases and controls. An individual who becomes a case could have their pre-case person time selected as a control, which enables us to compare those with Down syndrome and AD to the full Down syndrome population.

Co-occurring conditions

For input in our machine learning models, we used established algorithms for common chronic and disabling conditions.^26,27 There were 30 chronic conditions and 40 other disabling condition algorithms developed by CMS and are comprehensive in covering common health outcomes. These algorithms are like the AD algorithm described above and use ICD-9, ICD-10, procedure codes, and claim location (e.g., inpatient, outpatient) to identify conditions (See Supplemental Table 1). We used clinical expertise and literature review to identify certain additional conditions that are prevalent in Down syndrome⁶ but not included in these algorithms, e.g., sleep apnea, bone breaks. We determined conditions that present at birth or in childhood (e.g., autism, spina bifida) not to be considered incident in adulthood and created one binary variable for any occurrence. For variables that could be incident (e.g., epilepsy, depression), capturing timing was important so we created indicators for incidence within three months, six months, one year, two years, and three years prior to the index date. These indicators capture new diagnoses, not necessarily that date of condition onset and is an approach consistent with previous studies.²⁸ We also grouped conditions into categories (cancer, mental health, cardiovascular disease, bone break) and created binary indicators for ever having a condition in the category. In total, we examined 106 conditions.

Demographic variables

Demographic variables were from the demographic enrollment files in Medicaid and the Master Beneficiary Summary File in Medicare. We aligned data from Medicaid and Medicare, so all data was consistent regardless of what program one was enrolled in. When race/ethnicity was missing in Medicaid (∼25%) we used multiple imputation to impute race/ethnicity. Briefly, we used individual data (e.g., age, sex, state) and community level data at the zip code level (e.g., percent of the population that is Black non-Hispanic) to model probability that an individual with missing race/ethnicity data is of a certain race/ethnicity (See Rubenstein et al. (2024) for more details).¹⁹ We created binary indicators for each racial/ethnic group, sex, region, health insurer (Medicaid, Medicare, or Dual enrolled), and age category at index date in ten-year increments.

Data processing and statistical analysis

To understand the main effects associations, we first calculated descriptive statistics and univariate odds ratios via logistic regression for all 488 predictors (any presence of a condition, timing of condition onset + demographics) and AD status. We used a data analysis pipeline that included data organization and preprocessing before model training. We had access to all Medicaid and Medicare claims and assumed that a lack of claim signified a lack of diagnosis or assessment (e.g., there were not non-Medicaid or Medicare billed healthcare costs that were not attempted to be paid by Medicaid or Medicare). We randomly split our data into training (80% n = 44,112) and test (20%, n = 11,029) data sets.

After eliminating rare predictors (occurring in less than 5 observations, 22 predictors), we trained gradient boosted trees (nrounds = 500 trees) with the R xgboost²⁹ and caret³⁰ packages for 467 predictors. Xgboost was used because it is a flexible tree-based learner able to model both main effects and interactions among features when making predictions. We used a 10-fold cross-validation grid search to optimize xgboost hyperparameters (e.g., tree depth, eta, gamma) and ensure model robustness. The cross-validation split the training data into ten parts (folds), with the model being developed in one-fold then tested in the other nine, with the process repeating so each fold is used for model development and tested against the other nine.

We used receiver operating curves and area under curve to identify the optimal model in the 80% training data. We then applied the optimal model to the test sample to evaluate performance. In addition to AUC, we divide the individual-level predicted probabilities in the test sample into quintiles and examined AD occurrence across quintiles (i.e., sensitivity, positive predictive value).

As sensitivity analysis, we also ran an elastic net model and a logistic regression model with a training/test split sample to assess whether gradient boosted trees was the optimal approach. Elastic net is a regularized linear regression combining penalties of lasso and ridge regression and is well suited for highly correlated predictors and modelling main events. We used the optimal tuning parameters identified in the gradient boosted tree model and calculated model AUC in the test set. For the logistic regression, we included the top five predictors identified from the gradient boosted trees in a logistic regression model and calculated AUC in the test sample.

We used the R package shapviz³¹ to calculate and visualize SHapley Additive exPlanations (SHAP)³² values from the xgboost model. SHAP is a model-agnostic approach to explaining the results of an ML model, in our case by identifying predictors with the most influence on protection or risk of AD. SHAP uses the trained model to estimate how each variable was associated with the outcome in everyone, then we extract the mean absolute importance for each of those comparisons and displayed them graphically. A higher mean absolute SHAP value indicates that the variable was more important for prediction. Beeswarm plots (and univariate ORs) were used to determine the directionality of predictive associations. We looked at SHAP interaction values among the top 15 predictors by plotting the relationship between each feature and the model output. We selected pairs of predictors that visually showed potential interactions to present. For those predictors, we then examined effect measure modification by comparing stratified odds ratios and tested statistical interaction by running a logistic regression model with AD as the dependent variable and the predictors and their interaction term as the independent variables.

Results

Demographics

Our analytic sample consisted of 55,606 adults with Down syndrome enrolled in Medicaid and/or Medicare who were ≥30 years old at any time from 2011–2019 (study entry flow diagram presented in Supplemental Figure 1). Of those, 16,398 (28.3%) had an incident AD claim. Mean age at study entry of those who developed AD was 50.0 years: 53.3% were female, 81.2% were non-Hispanic white, and 80.7% were enrolled in both Medicaid and Medicare (Table 1). Of those who did not develop AD, mean age at study entry was 42.0 years, 54.0% were female, 69.9% were non-Hispanic white, and 60.2% were enrolled in Medicaid and Medicare.

Table 1.

Demographics of adults with Down syndrome 2011–2019.

	Down Syndrome	With Alzheimer's disease	Without Alzheimer's disease
	N = 55,606	N = 16,398	N = 39,208
	N (%)	N (%)	N (%)
Age at study entry
<35	11,813 (20.1)	562 (3.4)	10,513(26.8)
35–44	19,706 (33.5)	3491 (21.3)	14,793 (37.7)
45–54	18,849 (32.1)	8410 (51.3)	10,031 (25.6)
55–64	6780 (11.5)	3285 (20.0)	2998 (7.6)
65–89	1606 (2.7)	650 (4.0)	873 (2.2)
Mean (SD)	44.3 (10.0)	50.0 (8.1)	42.0 (9.6)
Sex
Male	25,665 (46.2)	7664 (46.7)	18,001 (45.9)
Female	29,941 (53.8)	8734 (53.3)	21,207 (54.1)
Race /Ethnicity
White NH	40,963 (73.7)	13,309(81.2)	27,654 (69.9)
Black NH	6045 (10.9)	1486 (9.1)	4559 (11.6)
Asian /Pacific Islander	1025 (1.8)	192 (1.2)	833 (2.1)
Native American	448 (0.8)	85 (0.5)	363 (0.9)
Mixed Race	599 (1.1)	140 (0.9)	459 (1.2)
Hispanic	6073 (10.9)	1088 (6.6)	4985(12.7)
Missing	453	98	355
Region
Northeast	12,507 (21.3)	4064 (24.8)	7742 (19.7)
Midwest	14,410 (24.5)	4588 (28.0)	9321 (23.7)
South	20,032 (34.1)	5185 (31.8)	13,833 (35.3)
West	11,303 (19.2)	2526 (15.4)	7882 (20.1)
US Territories	477 (0.8)	33 (0.2)	408 (1.0)
Missing	23
Enrollment
Medicaid only	13,593 (23.1)	1656 (10.1)	9668 (24.6)
Dual enrolled	37,007 (73.0)	13,229 (80.7)	23,602 (60.2)
Medicare only	8154 (13.9)	1513 (9.2)	5938 (15.1)

NH: Non-Hispanic; SD: standard deviation.

Univariate analyses

The most common predictors of AD in univariate analyses were white race (73.7%), male sex (53.8%), any claim for hypothyroidism (51.0%), and any hyperlipidemia (50.8%) (Supplemental Table 2). The largest AD odds ratios concerned those with incident traumatic brain injury three months before index date (Odds ratio [OR] 12.4, 95% Confidence interval [CI]: 7.1, 23.4), incident epilepsy three months before index date (OR: 9.9, 95% CI: 8.6, 11.4), and incident traumatic brain injury six months before index date(OR: 9.6, 95% CI: 6.4, 14.9). Of the 467 predictors evaluated, all but 18 (3.9%) had odds ratio estimates > 1.0. Protective predictors were incident obesity three years before index date claim (OR: 0.9, 95% CI: 0.85, 0.95) and race (Black non-Hispanic (OR: 0.8, 95% CI: 0.71, 0.81), Mixed Race (OR: 0.7, 95% CI: 0.60, 0.88), Native American (OR: 0.6, 95% CI: 0.44, 0.70), Asian/Pacific Islander (OR: 0.5, 95% CI: 0.46, 0.64), and Hispanic (OR: 0.5, 95% CI: 0.46, 0.52).

Machine learning model statistics

After completing the hyperparameter grid search and training the model (Supplemental Table 3), the optimal xgboost model achieved an AUC of 0.86 (95% CI: 0.85, 0.87) in the test sample (Supplemental Figure 2). The elastic net model had an AUC of 0.85 (95% CI: 0.84, 0.86) with a and the logistic regression model had an AUC of 0.78 (95% CI: 0.77, 0.79) (Supplement Table 3). Given the slightly higher AUC and flexibility to model interactions, we present results only for the xgboost model.

In the xgboost model, Individuals in the top two quintiles of predicted risk accounted for 76% of all cases of incident AD, i.e., sensitivity in the two “high risk” strata was 76.0% (see Figure 1 for distribution of outcomes across all quintiles). Among those in the top quintile, 73.6% experienced AD (positive predictive value).

Figure 1.

Quintiles of prediction probability by probability of developing Alzheimer's disease, among Medicaid and/or Medicare enrolled adults with Down syndrome. On top of each bar is the maximum and minimum predicted value of each quintile (i.e., the 20% highest predictions had highest predicted values between 0.56 and 1.0). Proportion is within each quintile.

SHAP values

Based on the SHAP values (Figure 2, Supplemental Table 4), our strongest predictors were being older than 45 years, being older than 35 years, dual enrollment in Medicaid and Medicare, incident epilepsy three years before index date; any claim for epilepsy, deafness, hypothyroid, ulcer, schizophrenia, hyperlipidemia, any mental health diagnosis three years before index date. For SHAP feature values (Figure 3, <0 indicates protection, >0 indicates increased risk) presence of the predictor increased the predicted risk of AD.

Figure 2.

Absolute SHAP values for 15 most influential variables predicting Alzheimer's dementia in Medicaid and/or Medicare enrolled adults with Down syndrome. SHapley Additive exPlanations (SHAP) shows the contribution or the importance of each feature on the prediction of the mode.

Figure 3.

Beeswarm plot for SHAP values or 15 most influential variables predicting Alzheimer's disease among adults with Down syndrome enrolled in Medicaid and/or Medicare. SHAP value plots for each feature are shown, ranked by importance (mean absolute SHAP value). Each dot represents 1 individual, with their position representing their SHAP value. Color is used to indicate the specified feature characteristic. Lighter dots indicate those who have the listed characteristic, while purple dots indicate those who do not. For example, for the any claims for deafness, individuals with a yellow dot have claims for deafness and appear predominantly to the right of Y-axis, indicating higher predicted risk to develop AD, while those with a darker dot do not have claims for deafness.

Interactions

Our assessment of SHAP value interaction for the top 15 predictors are presented in Supplemental Figure 4. We highlight two meaningful interactions, as most were not meaningful (as represented by tight clustering around SHAP value 0): any claim for epilepsy by being older than 45 years and incident claim for epilepsy three years from index date by being older than 45 years (Figure 4). For any claim of epilepsy and being older than 45 years, the odds ratio for AD and epilepsy was 3.03 (95% CI: 2.5, 3.6) in those older than 45 and 1.85 (95% CI: 1.6, 2.1) in those 45 and younger, indicating effect measure modification. The interaction p value was <0.001 indicating statistical interaction. Incident epilepsy three years from index date and being older than 45 showed effect measure modification: the odds ratio between AD and incident epilepsy three years before index date was 5.61 (95% CI: 4.4, 7.4) in those older than 45 and 4.42 (95% CI: 3.7, 5.2) in those 45 and younger. The interaction term was not significant in the interaction regression model (p = 0.1).

Figure 4.

Interaction plots between highest valued predictors for Alzheimer's disease among adults with Down syndrome enrolled in Medicaid and/or Medicare.

Discussion

AD is burdensome and lethal condition that profoundly harms the Down syndrome population. The link between the amyloid precursor protein and chromosome 21 is a well understood cause of AD in Down syndrome. There remains substantial variability in populations with genetic risk for AD, such as Down syndrome, and examining other risk factors might explain some of the heterogeneity.³³ We used a full Medicare and Medicaid data set of adults with Down syndrome and flexible tree-based machine learning methods to identify which diagnoses and demographics predict incident AD diagnosis, advancing our knowledge on conditions that precede onset and could be an avenue for intervention.

Demographic predictors

Age older than 45 and older than 35 years at index data had the largest SHAP values while age > 55 was the 15th largest. Age is intrinsically tied to AD onset, and this onset occurs in the Down syndrome population around 54 years,^4,23 which supports being older than 45 years being the strongest predictor and less of an effect for >55 years. Medicaid and Medicare dual enrollment also had a high SHAP value, especially in older adults as evidence by synergistic interaction, which may reflect age and increased healthcare access.¹⁴ In the general population, women are at greater risk then men for AD and onset differs by racial/ethnic group.³⁴ We found that sex and race/ethnicity were strong predictors based on SHAP value. In our data, there was no difference in age at onset of AD comparing men and women or white non-Hispanic and Black non-Hispanic adults with Down syndrome.²³ The lack of difference by demographic group may be due to the homogenous etiology of AD in Down syndrome compared to that of the general population, but also underscores the need to understand risk of dementia based on preceding health conditions. The triplication of the chromosome 21 is the cause of AD in Down syndrome and we hypothesize that social factors and inequities that drive disparity in the general population³⁴ may not have as much of an effect among those with Down syndrome.

Epilepsy

In our machine learning model, incident epilepsy diagnosis three years before index date was the strongest predictor of incident AD diagnosis. Further, our interaction analyses found that incident epilepsy and older age had a synergistic effect. Epilepsy is common among people with Down syndrome at all ages,³⁵ but there is a phenomenon of Late Onset Myoclonic Epilepsy in Down Syndrome (LOMEDS) reported in clinical studies where incident epilepsy occurs after AD onset.³⁶ The strong link between epilepsy and AD occurs because amyloid-β (Aβ) plaques lead to synaptic degeneration, circuit remodeling and abnormal synchronization of neuronal networks.³⁷ This may induce a pro-epileptic effect.³⁸ Studies of LOMEDS find that AD precedes epilepsy,³⁶ whereas we saw epilepsy claims prior to AD claims. It is possible that AD diagnoses are delayed relative to LOMEDS diagnosis or there is a different late onset epilepsy³⁹ that precedes AD.

Chronic conditions

Other predictors identified through our model are also risk factors for AD in the general population. Deafness, hypothyroidism, pulmonary vascular disease, ulcers, and mental health issues, are all factors shown to increase risk for AD.³⁴ AD related pathology, such as amyloid plaque deposits⁴⁰ or inflammation,⁴¹ may cause conditions such as ulcers and pulmonary vascular disease, in those with and without Down syndrome before AD onset. Alternatively, cognitive reserve and isolation^40,42 may also exacerbate AD onset in people with Down syndrome and we are seeing mental health conditions and deafness preceding AD onset. Future studies need to address the timing of these conditions and its effect on the risk in a Down syndrome population that has high of amyloid levels, which might dampen or exacerbate some risk factors; e.g., the effect of APOE4 is lower in Down syndrome AD.⁴³ Misdiagnosis of prodromal AD symptoms as mental health disorder might justify the finding of schizophrenia and other psychotic conditions as strong predictors, when they might be they might be behavioral and psychological symptoms of dementia⁴⁴

Implications

Our findings have implications for research and practice. Epilepsy is well understood to be associated with Down syndrome and AD, but further examination on timing of onset is warranted. Cognitive reserve is an important area of research in prevention of AD onset and is a concept worth studying in Down syndrome. Clinically, interventions for conditions such as deafness and sleep apnea are beneficial in and of themselves,^45,46 but given the strength of the conditions as predictors, there may be added benefit for dementia prevention.⁴⁷ The promising AUC, sensitivity, and positive predictive value achieved by the model suggests it may be viable and useful to develop dementia risk scores in claims data for individuals with Down syndrome. If validated, a risk score could be used to identify which Down syndrome individuals are at greatest need for targeted intervention or dementia-focused prevention.

Limitations

Claims data are indicators of service received and billing and may not capture all underlying conditions. We used Medicare validated algorithms to identify conditions but there may still be misclassification. Further, our index date was timing of diagnosis, not true date of onset so there is potential for reverse causation. We aimed to create windows of observation that would account for potential time bias. We did not have test results, continuous measurements, or other phenotypic data that could inform our analyses. Our risk-set sampling was designed to account for differing person time and loss to follow up but there may still be some bias due to survival (e.g., less healthy people died sooner and had less time to amass claims for conditions). We conducted exploratory interaction analyses for all the top 15 predictors. While we did not conduct any statistical tests, we did make many comparisons which may lead to a spurious finding.

Strengths

We were able to examine predictors for AD in a full Medicare and Medicaid cohort of adults with Down syndrome. We were able to use existing validated algorithms to identify conditions in nine years of data. Our machine learning models identified the most robust model, and we had high AUC and sensitivity.

Conclusion

In Medicaid and Medicare enrolled adults with Down syndrome, age, dual Medicare and Medicaid enrollment, incident epilepsy three years before index date, and ever having deafness, were the strongest predictors of AD onset. Identifying these predictors highlights areas for further etiologic inquiry and intervention.

Supplemental Material

sj-docx-1-alz-10.1177_13872877251385423 - Supplemental material for Novel predictors of Alzheimer's disease in Down syndrome identified using machine learning

Supplemental material, sj-docx-1-alz-10.1177_13872877251385423 for Novel predictors of Alzheimer's disease in Down syndrome identified using machine learning by Eric Rubenstein, Salina Tewolde, Amy Michals, Juan Fortea, Marcia Pescador Jimenez, Brian G Skotko, Yorghos Tripodis, Jennifer Weuve and Anthony J Rosellini in Journal of Alzheimer's Disease

Supplemental Material

sj-xlsx-2-alz-10.1177_13872877251385423 - Supplemental material for Novel predictors of Alzheimer's disease in Down syndrome identified using machine learning

Supplemental material, sj-xlsx-2-alz-10.1177_13872877251385423 for Novel predictors of Alzheimer's disease in Down syndrome identified using machine learning by Eric Rubenstein, Salina Tewolde, Amy Michals, Juan Fortea, Marcia Pescador Jimenez, Brian G Skotko, Yorghos Tripodis, Jennifer Weuve and Anthony J Rosellini in Journal of Alzheimer's Disease

Footnotes

Acknowledgements

The authors have no acknowledgments to report.

ORCID iD

Eric Rubenstein

Ethical considerations

Not applicable.

Consent to participate

Not applicable.

Consent for publication

Not applicable.

Author contribution(s)

Eric Rubenstein: Conceptualization; Data curation; Formal analysis; Funding acquisition; Investigation; Methodology; Project administration; Resources; Supervision; Writing – original draft; Writing – review & editing.

Salina Tewolde: Formal analysis; Validation; Visualization; Writing – review & editing.

Amy Michals: Data curation; Writing – review & editing.

Juan Fortea: Methodology; Writing – review & editing.

Marcia Pescador Jimenez: Writing – review & editing.

Brian Skotko: Conceptualization; Writing – review & editing.

Yorghos Tripodis: Methodology; Writing – review & editing.

Jennifer Weuve: Methodology; Writing – review & editing.

Anthony Rosellini: Conceptualization; Formal analysis; Methodology; Supervision; Validation; Visualization; Writing – review & editing.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was funded by the National Institute on Aging grant R01AG073179, the Instituto de Salud Carlos III (Ministerio de Asuntos Económicos y Transformación Digital, Gobierno de España) through the projects INT21/00073, PI20/01473 and PI23/01786 to J.F.), the Centro de Investigación Biomédica en Red sobre Enfermedades Neurodegenerativas CIBERNED Program 1, partly jointly funded by Fondo Europeo de Desarrollo Regional, Unión Europea, Una Manera de Hacer Europa. This work was also supported by the National Institutes of Health grants (R01 AG056850; R21 AG056974, R01 AG061566, R01 AG081394 and R61AG066543 and 1RF1AG080769–01 to to J.F.), the Departament de Salut de la Generalitat de Catalunya, Fundación Tatiana Pérez de Guzmán el Bueno (IIBSP-DOW-2020-151 to J.F and SG.) and Horizon 2020–Research and Innovation Framework Programme from the European Union (H2020-SC1-BHC-2018-2020 to J.F.), Brightfocus, and Life Molecular Imaging (LMI) to JF.

Declaration of conflicting interests

The authors declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: JF reported serving on the advisory boards, adjudication committees, or speaker honoraria from AC Immune, Adamed, Alzheon, Biogen, Eisai, Esteve, Fujirebio, Ionis, Laboratorios Carnot, Life Molecular Imaging, Lilly, Lundbeck, Novo Nordisk, Perha, Roche, Zambón, Spanish Neurological Society, T21 Research Society, Lumind foundation, Jérôme-Lejeune Foundation, Alzheimer's Association, National Institutes of Health USA, and Instituto de Salud Carlos III. JF reports holding a patent for markers of synaptopathy in neurodegenerative disease (licensed to ADx, EPI8382175.0). No other competing interests were reported.

Dr Skotko occasionally consults on the topic of Down syndrome through Gerson Lehrman Group. He receives remuneration from Down syndrome non-profit organizations for speaking engagements and associated travel expenses. In the past two years, Dr Skotko received annual royalties from Woodbine House, Inc., for the publication of his book, Fasten Your Seatbelt: A Crash Course on Down Syndrome for Brothers and Sistaers. Within the past two years, he has received research funding from AC Immune, and LuMind IDSC Down Syndrome Foundation to conduct clinical trials for people with Down syndrome. Dr Skotko is occasionally asked to serve as an expert witness for legal cases where Down syndrome is discussed. Dr Skotko serves in a non-paid capacity on the Honorary Board of Directors for the Massachusetts Down Syndrome Congress and the Professional Advisory Committee for the National Center for Prenatal and Postnatal Down Syndrome Resources. Dr Skotko has a sister with Down syndrome.

Data availability statement

Data are not available to be shared due to the Centers for Medicare and Medicaid services data use policy.

Supplemental material

Supplemental material for this article is available online.

References

Antonarakis

Skotko

Rafii

, et al. Down syndrome. Nat Rev Dis Primers 2020; 6: 9.

Zigman

Lott

. Alzheimer's disease in Down syndrome: neurobiology and risk. Ment Retard Dev Disabil Res Rev 2007; 13: 237–246.

Landes

Stevens

Turk

. Cause of death in adults with Down syndrome in the United States. Disabil Health J 2020; 13: 100947.

Iulita

Garzon Chavez

Klitgaard Christensen

, et al. Association of Alzheimer disease with life expectancy in people with Down syndrome. JAMA Netw Open 2022; 5: e2212910.

de Graaf

Buckley

Skotko

. Estimation of the number of people with Down syndrome in the United States. Genet Med 2017; 19: 439–447.

Bull

. Down syndrome. N Engl J Med 2020; 382: 2344–2352.

Dintica

Yaffe

. Epidemiology and risk factors for dementia. Psychiatr Clin North Am 2022; 45: 677–689.

Prasher

. Age-specific prevalence, thyroid dysfunction and depressive symptomatology in adults with down syndrome and dementia. Int J Geriatr Psychiatry 1995; 10: 25–31.

Menendez

. Down syndrome, Alzheimer's disease and seizures. Brain Dev 2005; 27: 246–252.

10.

Vaughan

McGee

Guerin

, et al. The challenges of diagnosis and treatment of dementia in Down's syndrome. Ir J Psychol Med 2016; 33: 151–158.

11.

Ercolano

Bencivenga

Palaia

, et al. Intricate relationship between obstructive sleep apnea and dementia in older adults. Geroscience 2024; 46: 99–111.

12.

Handen

Lott

Christian

, et al. The Alzheimer's biomarker consortium-Down syndrome: rationale and methodology. Alzheimers Dement (Amst) 2020; 12: e12065.

13.

Rothman

Greenland

Lash

. Modern Epidemiology, 3rd Edition. Philadelphia, PA: Lippincott, Williams & Wilkins, 2008.

14.

Rubenstein

Tewolde

Levine

, et al. Medicare, Medicaid, and dual enrollment for adults with intellectual and developmental disabilities. Health Serv Res 2024; 59: e14287.

15.

Ludwig

Mullainathan

. Machine learning as a tool for hypothesis generation. Q J Econ 2024; 139: 751–827.

16.

Wang

Liu

Spiteri

, et al. Understanding machine learning applications in dementia research and clinical practice: a review for biomedical scientists and clinicians. Alzheimers Res Ther 2024; 16: 175.

17.

Nori

Hane

Martin

, et al. Identifying incident dementia by applying machine learning to a very large administrative claims dataset. PLoS One 2019; 14: e0203246.

18.

Uspenskaya-Cadoz

Alamuri

Wang

, et al. Machine learning algorithm helps identify non-diagnosed prodromal Alzheimer's disease patients in the general population. J Prev Alzheimers Dis 2019; 6: 185–191.

19.

Rubenstein

Michals

Wang

, et al. Medicaid enrollment and service use among adults with Down syndrome. JAMA Health Forum 2023; 4: e232320.

20.

McDermott

Royer

Cope

, et al. Using Medicaid data to characterize persons with intellectual and developmental disabilities in five U.S. states. Am J Intellect Dev Disabil 2018; 123: 371–381.

21.

Fortea

Zaman

Hartley

, et al. Alzheimer's disease associated with Down syndrome: a genetic form of dementia. Lancet Neurol 2021; 20: 930–942.

22.

Bhattacharyya

Barnes

Chen

, et al. Evaluating linked ICD-10 Medicare claims data as a method of dementia case ascertainment in research settings. Alzheimers Dement 2025; 21: e70200.

23.

Rubenstein

Tewolde

Michals

, et al. Alzheimer dementia among individuals with Down syndrome. JAMA Network Open 2024; 7: e2435018.

24.

Langholz

Goldstein

. Risk set sampling in epidemiologic cohort studies. Stat Sci 1996; 11: 35–53.

25.

O'Brien

Lawrence

Keil

. The case for case-cohort: an applied epidemiologist's guide to reframing case-cohort studies to improve usability and flexibility. Epidemiology 2022; 33: 354–361.

26.

Center for Medicare and Medicaid Services . CMS Chronic Conditions Data Warehouse (CCW), CCW Condition Algorithm. Washington, DC: Department of Health and Human Services, 2021.

27.

Center for Medicare and Medicaid Services . CMS Chronic Conditions Data Warehouse (CCW), CCW Chronic Condition Reference List. Washington, DC: Department of Health and Human Services, 2021.

28.

Gradus

Rosellini

Horvath-Puho

, et al. Prediction of sex-specific suicide risk using machine learning and single-payer health care registry data from Denmark. JAMA Psychiatry 2020; 77: 25–34.

29.

Chen

Benesty

, et al. xgboost: extreme gradient boosting. R package version 1.7.8.1. https://CRAN.R-project.org/package=xgboost (2024). 2024.

30.

Kuhn

. Building predictive models in R using the caret package. J Stat Softw 2008; 28: 1–26.

31.

Mayer

. shapviz: SHAP visualizations. R package version 0.9.6. https://CRAN.R-project.org/package=shapviz2024 (2024).

32.

Lundberg

Nair

Vavilala

, et al. Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nat Biomed Eng 2018; 2: 749–760.

33.

Fortea

Vaque-Alcazar

Pegueroles

, et al. Reply to: challenges to identifying risk versus protective factors in Alzheimer's disease. Nat Med 2024; 30: 3096–3097.

34.

Zhang

Tian

Wang

, et al. The epidemiology of Alzheimer's disease modifiable risk factors and prevention. J Prev Alzheimers Dis 2021; 8: 313–321.

35.

Alexander

Petri

Ding

, et al. Morbidity and medication in a large population of individuals with Down syndrome compared to the general population. Dev Med Child Neurol 2016; 58: 246–254.

36.

Corniello

Dono

Evangelista

, et al. Diagnosis and treatment of late-onset myoclonic epilepsy in Down syndrome (LOMEDS): a systematic review with individual patients’ data analysis. Seizure 2023; 109: 62–67.

37.

Altuna

Gimenez

Fortea

. Epilepsy in Down syndrome: a highly prevalent comorbidity. J Clin Med 2021; 10: 2776.

38.

Lott

Doran

Nguyen

, et al. Down syndrome and dementia: seizures and cognitive decline. J Alzheimers Dis 2012; 29: 177–185.

39.

Punia

Bhansali

Tsai

. Late-onset epilepsy clinic: from clinical diagnostics to biomarkers. Seizure 2025; 128: 68–73.

40.

Griffiths

Lad

Kumar

, et al.

How can hearing loss cause dementia?

Neuron 2020; 108: 401–412.

41.

Kinney

Bemiller

Murtishaw

, et al. Inflammation as a central mechanism in Alzheimer's disease. Alzheimers Dement (N Y) 2018; 4: 575–590.

42.

Nelson

Jester

Petkus

, et al. Cognitive reserve, Alzheimer's neuropathology, and risk of dementia: a systematic review and meta-analysis. Neuropsychol Rev 2021; 31: 233–250.

43.

Iulita

Bejanin

Vilaplana

, et al. Association of biological sex with clinical outcomes and biomarkers of Alzheimer's disease in adults with Down syndrome. Brain Commun 2023; 5: fcad074.

44.

Ismail

Creese

Aarsland

, et al. Psychosis in Alzheimer disease — mechanisms, genetics and therapeutic opportunities. Nat Rev Neurol 2022; 18: 131–144.

45.

Simpson

Oyekan

Ehsan

, et al. Obstructive sleep apnea in patients with Down syndrome: current perspectives. Nat Sci Sleep 2018; 10: 287–293.

46.

Bull

. Improvement of outcomes for children with Down syndrome. J Pediatr 2018; 193: 9–10.

47.

Boyle

Koops

Ances

, et al. Resistance and resilience to Alzheimer's disease in Down syndrome. Alzheimers Dement 2025; 21: e70151.

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

0.38 MB

0.09 MB