Predicting Age at Alzheimer’s Dementia Onset with the Cognitive Clock

Abstract

INTRODUCTION:

Intervention of Alzheimer’s dementia hinges on early diagnosis and advanced planning. This work utilizes the cognitive clock, a novel indicator of brain health, to develop a dementia prediction model that can be easily applied in broad settings.

METHODS:

Data came from over 3,000 community-dwelling older adults. Cognitive age was estimated by aligning MMSE scores to a clock that represents the typical cognitive aging profile. We identified a mean cognitive age at Alzheimer’s dementia onset and predicted the corresponding chronological age at person-specific level.

RESULTS:

The mean chronological age at baseline was 78 years. 881 (28%) participants developed Alzheimer’s dementia. The mean cognitive age at onset was 91 years. The predicted chronological age at onset had a mean (standard deviation) of 87.6 (6.7) years. The model’s prediction accuracy was supported by multiple testing statistics.

DISCUSSION:

Our model offers an easy-to-use tool for predicting person-specific age at Alzheimer’s dementia onset.

INTRODUCTION

Alzheimer’s dementia is a progressive and irreversible brain disorder. Six and a half million people aged 65 years and older in the United States currently live with Alzheimer’s dementia [1]. Alzheimer’s dementia impairs memory, language, decision making, and the basic functions of daily living. Caregivers of people with Alzheimer’s dementia endure billions of expenses in unpaid care [1], and caregiving exacts a major toll on the health and wellbeing of caregivers themselves [2–4]. Prediction of dementia has been a focus of research for many years, as early diagnosis and advanced planning are pivotal to alleviate the shock of dementia diagnosis, reduce financial and health burdens associated with dementia care, and improve the quality of life and care both for individuals with dementia and their caregivers [5]. Early diagnosis and advanced planning require accurate prediction, at a person-specific level, of age at dementia onset, but this has proven difficult.

Recent advances in cerebrospinal fluid (CSF) and neuroimaging biomarkers for Alzheimer’s disease (AD), specifically β-amyloid and tau, have facilitated the development of models for predicting the onset of Alzheimer’s dementia [6–8]. These advances are invaluable for the early detection of abnormal brain changes that could lead to AD. Yet, reliance on invasive and expensive CSF or neuroimaging analysis limits their use in general clinical settings. In addition, Alzheimer’s dementia is a heterogeneous clinical syndrome that can be attributable to a combination of neuropathologies, not just AD [9]. As such, AD biomarkers alone may not be sufficient for accurate prediction. Alternative approaches that use easily accessible cognitive or behavioral measures are needed for broader applicability.

We recently established a cognitive clock as a novel indicator of brain health [10]. The cognitive clock was developed using data from very well characterized cohort of older persons studied for over 20 years, and reflects the common trajectory of late life cognitive change over time that spans from no impairment to dementia through mild cognitive impairment (MCI). Data for chronological age and longitudinal cognitive assessments, i.e. Mini-Mental State Examination (MMSE), were used to construct this clock. We showed that, after aligning individual’s MMSE scores to the clock, we can derive a metric of person-specific cognitive age, which outperforms chronological age in predicting adverse outcomes including Alzheimer’s dementia. In this study, we expanded our prior work by investigating the prognostic utility of the cognitive clock for predicting the onset of Alzheimer’s dementia at the individual level. Our goal is to provide an algorithm that, given an individual’s chronologic age and MMSE scores, accurately predicts the timing of the onset of Alzheimer’s dementia. Briefly, the cognitive clock was profiled using data from 1,749 community-dwelling older adults who were free of dementia at baseline, followed annually for up to 30 years, and had died. The Alzheimer’s dementia-free survival curve was determined using an accelerated failure time (AFT) model on the cognitive age scale, and each individual’s chronological age at diagnosis was predicted using the mean cognitive age at diagnosis. The model performance was tested in a separate sample of 465 older adults who were alive.

METHODS

Study Participants

Data came from community-dwelling older adults who participated in one of two ongoing clinicopathologic cohort studies of aging and dementia, i.e. the Religious Orders Study or the Rush Memory and Aging Project (ROSMAP). Participants were enrolled without known dementia and agreed to annual uniform detailed evaluations that include MMSE and clinical assessments for Alzheimer’s dementia [11]. The studies were approved by an institutional review board of Rush University Medical Center, and written informed consents were provided by all participants.

At the time of these analyses, 3,454 participants over the age of 65 completed the baseline evaluation and were free of dementia. Of those, 3,128 with at least 2 annual assessments were included in the analyses. To capture the complete history of cognitive aging, the cognitive clock was profiled using data from the participants who had died (N=1,749), and the same sample was used for the survival model that predicts the timing of Alzheimer’s dementia onset. Among the remaining 1,379 participants who were alive, 465 were at risk for Alzheimer dementia (i.e., those who reached cognitive age of 85) at prediction and used to assess prediction accuracy.

Diagnosis of Alzheimer’s Dementia

Uniform structured clinical evaluations were administered to the ROSMAP participants each year [12]. Annual evaluations include an interview on medical history, a comprehensive cognitive testing as well as a complete neurological examination. Cognitive tests were scored and converted to an impairment rating using computer. A neuropsychologist reviewed the cognitive impairment ratings and determined the presence of cognitive impairment. A clinician reviewed all available data and rendered a clinical judgement on the presence of dementia and its likely etiology. The diagnosis of Alzheimer’s dementia follows the 1984 McKhann guideline [13], which requires a history of cognitive decline, and impairment in memory and at least 1 additional cognitive domain.

Statistical Analysis for Prediction

The statistical analyses for predicting age at Alzheimer’s dementia onset were conducted in three steps. First, we estimated the Alzheimer’s dementia-free survival by fitting an AFT model on the cognitive age scale. Person specific cognitive age was estimated using a cognitive clock as previously reported [10], with a slight modification (eMethods, eFigures 1&2). The AFT model takes the form of $\log \left(T_i\right)=\log \left(T_0\right)+{\epsilon }_i$. Here, the reference failure time ${(T}_0)$ follows a generalized gamma distribution. $T_i$ is the cognitive age at onset for participant $i$, and $T_i$ is censored at the last evaluation before death for individuals without a diagnosis. The resulting survival curve estimates the probability of Alzheimer’s dementia onset at any given cognitive age. We identified the mean cognitive age at diagnosis as our endpoint for prediction.

Next, we predicted individual’s chronological age corresponding to the mean cognitive age at diagnosis, which was determined a priori at cognitive age 85. We chose this threshold because a cognitive age at 85 represents an impending Alzheimer’s dementia. Specifically, the survival curve shows that almost all participants at cognitive age 85 were free of dementia and yet they were on the edge of a drastic drop in Alzheimer’s dementia free survival probability. A similar pattern was also observed on the cognitive clock.

Finally, we tested prediction performance using a separate sample of ROSMAP participants who were alive. The overall model performance was assessed using an integrated area under the curve (eMethods). Separately, person-specific predictions were assessed using root mean square error (RMSE) and median absolute deviation (MAD) (eMethods).

All data processing and statistical analyses were performed using SAS/STAT software version 15.2 and R program version 3.6.0.19. The shape invariant model and AFT model were fit using the R packages nlme and flexsurv respectively [14, 15].

RESULTS

Profile of the Cognitive Clock

Following our prior work, we built the cognitive clock using a modified shape invariant model and a larger sample (N=1,749). In this sample, the mean (standard deviation [SD]) chronological age at baseline was 79.9 (6.9) years, and the mean chronological age at death was 89.6 (6.5) years. Approximately 69.2% were female. The participants had a mean education of 16.2 (3.7) years. They were followed for an average of 7.2 (5.4) years. At death, 594 (34.0%) remained cognitively intact, 431 (24.6%) were MCI, and 716 (40.9%) were diagnosed with Alzheimer’s dementia.

The clock profile was almost identical to what we previously reported. There was no or little decline prior to cognitive age 80, which was followed by a period of moderate decline until cognitive age 90, and then a terminal stage of precipitous decline after cognitive age 90. Notably, the clock leverages the trajectories of the entire study sample and captures a broad spectrum of cognitive aging from unimpaired to dementia through MCI (Figure 1). It is evident that for older adults who died without impairment, their cognitive trajectories were truncated at an early phase of the clock. By contrast, the cognitive trajectories for those who died with dementia covered the entire spectrum.

Figure 1: Alignment of person-specific MMSE trajectories on the cognitive clock.

Top panels aligned person-specific MMSE trajectories onto the cognitive clock by the diagnostic groups of NCI, MCI, and Alzheimer’s dementia. Lower panel is the boxplot of the estimated cognitive age at death for the three diagnostic groups.

Cognitive Age and Incident Alzheimer’s Dementia

We aligned each individual’s cognitive trajectory to the clock and estimated cognitive age at diagnosis or proximate to death. We fit an AFT model to obtain an Alzheimer’s dementia free survival curve on the cognitive age scale. The model-derived survival curve overlaps with the Kaplan-Meier curve and features a sharp bend (Figure 2), suggesting that the AFT model fits the data well. Notably, almost all Alzheimer’s dementia diagnoses occurred within a 10-year window between cognitive age of 85 and 95. The mean cognitive age at Alzheimer’s dementia onset was about 91 years.

Figure 2: Alzheimer’s dementia-free survival.

The figure shows Alzheimer’s dementia-free survival curves on cognitive age scale: observed (Kaplan-Meier curve, solid black) vs. model-fitted (dashed red). The shaded area covers the interval between cognitive age 85 and 95. The blue dashed line is the reference line at cognitive age 91.

Next, using cognitive age 91 as an endpoint for prediction, we back solved the corresponding chronological age. Time at prediction was anchored at cognitive age 85. Among the deceased, 1,068 ROSMAP participants (61.1%) reached cognitive age 85 during the follow-up. At cognitive age 85, the mean (SD, Range) chronological age was 84.8 (5.7, 63.1-102.2) and the median (interquartile range) MMSE score was 26 (25-26). The model predicted chronological age at cognitive age 91 had a mean (SD, Range) of 88.8 (5.7, 66.0-107.9).

Assessing Model Prediction Accuracy

We used a separate sample of living participants in ROSMAP (N=1,379) to assess the performance of person-specific predictions of age at Alzheimer’s dementia onset. On average, participants in this sample were younger, mostly female, had a shorter mean length of follow up and fewer incident Alzheimer’s dementia cases (Table 1). Importantly, compared to the deceased, the alignment of cognitive trajectories onto the clock is similar in the living, and the Alzheimer’s dementia-free survival distributions are almost identical (eFigure 3). By the time of these analyses, 156 participants (11.3%) were diagnosed with Alzheimer’s dementia.

Table 1.

Basic characteristic of study participants

	All	Died	Alive
N	3128	1749	1379
Chronological age at baseline	78.1 (7.4)	79.9 (6.9)	75.9 (7.4)
Length of follow-ups (years)	9.1 (5.8)	9.7 (5.6)	8.5 (6.0)
Education (years)	16.4 (3.7)	16.2 (3.7)	16.6 (3.8)
Men	826 (26.4%)	538 (30.8%)	288 (20.9%)
# Incident Alzheimer’s dementia	881 (28.2%)	725 (41.5%)	156 (11.3%)
Time to Alzheimer’s dementia diagnosis	7.4 (5.6)	7.2 (5.4)	8.5 (6.3)
Chronological age at diagnosis	87.6 (6.6)	87.9 (6.5)	86.2 (6.9)

In this testing sample, 465 participants (33.7%) reached cognitive age 85. At cognitive age 85, the mean (SD, Range) chronological age was 82.9 (6.3, 63.1-98.2) and the median (interquartile range) MMSE score was 26 (25-27). The model predicted chronological age at cognitive age 91 had a mean (SD, Range) of 87.6 (6.7, 65.0-105.1). Together, these results predicted that after an older adult reached cognitive age 85, they would have a mean of 4.7 years before the onset of Alzheimer’s dementia. The corresponding SD was 2.9 years, with a range of 0.04 to 27 years.

We first assessed the overall model performance in the testing sample. The integrated AUC was 0.87 for individuals who developed Alzheimer’s dementia, and 0.85 if censored individuals were included. Both statistics suggest a high temporal concordance between the predicted and observed chronological ages at Alzheimer’s dementia onset. Next, we assessed the accuracy of person-specific prediction. The RMSE was 4.6 years, and the MAD was 2.2 years (Figure 3). The latter value demonstrates that a majority of our predictions fall within a window of +/− 2 years of observed age at onset, suggesting a high degree of accuracy.

Figure 3: Comparison of observed and predicted age at Alzheimer’s dementia onset.

The figure shows a scatter plot of the predicted age at Alzheimer’s dementia diagnosis (y-axis) vs the observed age at Alzheimer’s dementia diagnosis (x-axis) for the ROSMAP participants who were alive and had Alzheimer’s dementia during the follow-up. The red solid line in the center is the 45° identity line.

An important feature of our model is its ability to leverage all available MMSE scores for dynamic prediction. We investigated the extent to which prior history of MMSE assessments improves the prediction accuracy. We compared our results with an alternative model that used only a single MMSE score at the time of prediction. Interestingly, we only observe a negligible difference in the prediction accuracy when comparing one versus multiple assessments.

To summarize the prediction algorithm, (1) the available MMSE scores by the time of prediction are aligned to the cognitive clock to calculate cognitive age at prediction, (2) the algorithm stops if the predicted cognitive age is under 85, and (3) for an individual whose cognitive age is equal to or greater than 85, the algorithm predicts age at diagnosis by back solving for the chronological age that corresponds to cognitive age 91. Table 2 shows the predicted ages at Alzheimer’s dementia onset based on a range of chronological ages and MMSE scores. According to the table, e.g., for a 75-year-old individual who had a MMSE score of 26 and was free of dementia, our model predicts that this individual will have 9 years until Alzheimer’s dementia onset. In comparison, an 85-year-old individual who had the same MMSE score will have a shorter time horizon of 5 years. Additional prediction results are included in the supplementary materials for easy reference (eTable 1).

Table 2.

Predicted Age at Alzheimer’s Dementia Diagnosis

		MMSE
Chronological Age at prediction		30	29	28	27	26	25	24	23	22
65	Cognitive Age	65	65	65	65	65	86	88	89	90
	AAO	-	-	-	-	-	78	74	72	69
70	Cognitive Age	70	70	70	70	84	87	88	89	90
	AAO	-	-	-	-	-	80	77	75	73
75	Cognitive Age	75	75	76	82	86	87	89	90	90
	AAO	-	-	-	-	84	82	80	79	77
80	Cognitive Age	78	79	82	85	86	88	89	90	91
	AAO	-	-	-	88	87	85	84	83	81
85	Cognitive Age	80	82	84	85	87	88	89	90	91
	AAO	-	-	-	91	90	89	88	87	86
90	Cognitive Age	82	83	85	86	87	88	89	90	91
	AAO	-	-	95	95	94	93	92	91	91
95	Cognitive Age	83	84	85	86	88	89	89	90	91
	AAO	-	-	99	98	98	97	97	96	95

Cognitive Age: Cognitive Age at Prediction; AAO: Predicted age at onset of Alzheimer’s dementia. Cognitive age at prediction judges whether an individual is at risk for Alzheimer’s dementia. For those reached cognitive age 85 at prediction, AAO provides the predicted chronological age at onset.

DISCUSSION

Effective interventions for Alzheimer’s dementia in older age require early diagnosis and advanced planning. Early diagnosis of Alzheimer’s dementia offers at-risk individuals the much needed opportunity to be involved in decision making for critical matters such as long term care and estate planning. Advanced planning significantly reduces the financial and healthcare burdens imposed on family members, care providers, and society as a whole. In this work, we demonstrate the prognostic utility of the cognitive clock in predicting the timing of the onset of Alzheimer’s dementia. Notably, cognitive age estimated from the clock outperforms chronological age and the MMSE score in overall prediction. At a person specific level, the predicted timing of dementia onset is largely accurate without heavily relying on prior history of cognitive data, meaning that predictions using even a single MMSE score are robust. The clinical implications of these findings are discussed below.

Advanced planning is crucial for older adults at high risk of developing dementia. In particular, planning allows them to communicate their preferences and wishes for the future, and this can involve a wide range of activities, such as healthcare choices, financial decision making, designation of durable power of attorney, living wills, and estate planning. The benefits of advanced planning are well documented [16]. Advanced planning improves patient centered care by not only honoring the patient’s preferences, but also “promoting more thoughtful and compassionate approaches to care” at an advanced disease stage [17, 18]. Prior data also suggest that advanced planning can reduce medical expenditures and healthcare costs [19]. Importantly, advanced planning for dementia is a long and emotional process that would likely be more successful if done at a comfortable pace and involves conversations between patients, family members and physicians. The accurate prediction of age at dementia onset provides a much needed window of time for the affected and their family members to make well informed and thoughtful decisions. The potential public health and financial benefits of having this time for planning are considerable.

An important advantage of this work is that we used only chronological age and brief cognitive assessments (i.e., MMSE scores) for prediction. Our approach complements recent models that rely on AD biomarkers by providing greater feasibility in clinical settings. Novel AD biomarkers, including CSF assays for Aβ and tau, MRI measures of brain atrophy, and PET techniques for brain metabolism and Aβ, detect neuropathologic changes before clinical manifestation and are crucial for early diagnosis and monitoring brain changes as the disease progresses. Notably, however, AD biomarkers are relatively nascent and not readily accessible to healthcare providers and patients. The cognitive age metric leveraged in this work is derived from the MMSE, the most widely used cognitive screening instrument for Alzheimer’s dementia diagnosis, and thus it is suitable for broad application in clinical and research settings. The overall model performance, as measured by the AUCs, suggests good sensitivity and specificity. Notably, the person-specific prediction accuracy is comparable with findings from a recent publication that used amyloid PET imaging to estimate dementia onset [7]. Schindler and colleagues measured brain amyloid level via mean cortical standardized uptake value ratio (SUVR) from PET scans, and predicted time to Alzheimer’s dementia using estimated age at which people reached an amyloid accumulation threshold of SUVR=1.2. For individuals who developed Alzheimer’s dementia, the RMSE was 4.5 years. For comparison, our prediction gives a RMSE of 4.6 years, suggesting that our model using age and MMSE performs just as well as the one that relies on considerably more expensive and invasive AD biomarkers. In addition, we assessed the accuracy of cognitive age for predicting age of Alzheimer’s dementia onset in an independent sample from a biracial population-based cohort study. Both the overall model performance and person-specific prediction accuracy were similar comparing to the results from the ROSMAP testing sample (eResults). This provides strong evidence that the proposed cognitive age metric can be applied to independent datasets to predict age at Alzheimer’s dementia onset.

Notably, most AD biomarkers are specific to key pathophysiologic processes in AD. As such, these biomarkers are intended to capture AD [20], but they do not take into account the complexity of neuropathologic changes that are present in Alzheimer’s dementia. It is increasingly recognized that non-AD neurodegeneration and cerebrovascular conditions have significant impacts on late-life cognitive decline and Alzheimer’s dementia [21], and indeed most Alzheimer’s dementia cases are due to mixed pathologies [22]. AD biomarkers alone may not be sensitive in predicting individuals with dementia due to mixed pathologies or with non-AD predominant conditions. By contrast, models derived from more traditional cognitive or behavioral assessments, while less disease specific, are likely to better detect older adults with mixed pathologies and who are at high risk of developing dementia.

Cognitive impairment is the core and defining clinical feature of Alzheimer’s dementia, and the effects of risk factors on Alzheimer’s dementia work largely through functional impairment as reflected in the MMSE scores. Therefore, by leveraging easily accessible cognitive data for predicting Alzheimer’s dementia, we reduce model complexity and maximize prediction accuracy. We illustrate this by using APOE as an example. APOE ε4 is by far the most potent genetic risk factor for late onset Alzheimer’s dementia [23]. It is estimated that a single copy of ε4 triples the risk of developing Alzheimer’s dementia and the risk is much higher for homozygous carriers. We compared the prediction models with and without including the data for APOE ε4 status. We only observed a modest improvement in the prediction accuracy, such that the RMSE was improved by 0.2 year and MAD by 0.5 year. Similar results were also observed for demographic characteristics of sex and education (eResults). It is worth noting that the cognitive age derived using MMSE scores may not be sensitive enough to detect subclinical change in cognition including MCI. Therefore, our current prediction was tailored to at-risk older adults who reached the cognitive age of 85. Incorporating other factors, particularly novel disease biomarkers, may hold potential for a more accurate prediction and earlier detection of Alzheimer’s dementia. Our findings suggest, however, that the current metric of cognitive age is a good and easily attainable predictor of cognitive outcomes.

We also assessed the extent to which the number of prior cognitive assessments may affect the estimation of cognitive age, and subsequently the prediction accuracy for Alzheimer’s dementia onset. Longitudinal testing scores capture the history of an individual’s cognitive decline, and we hypothesize that more assessments would improve the prediction. Surprisingly, compared to the model using only a single score at prediction, the RMSE and MAD from the model using multiple MMSE scores were similar. This result suggests that prediction by leveraging the cognitive clock does not necessarily require multiple assessments. One possible reason for this is that cognitive age was estimated by mapping MMSE scores onto a robust cognitive profile that provides a blueprint for cognitive aging at the population level. That is, the clock was constructed using longitudinal cognitive data that were densely collected for up to 30 years and based on a large sample of older adults with diverse trajectories and diagnoses at death. As such, the benefits of longitudinal assessments are likely already absorbed in clock construction. Notably, for prediction models, it is important to strike a balance between improved prediction accuracy and the extra testing burden associated with the need for multiple assessments. Thus, if confirmed, the finding that only one MMSE is sufficient for a good prediction of dementia onset is of particular clinical relevance and would significantly reduce both the costs and burden of testing for patients and healthcare providers.

Chronological age is a known risk factor for Alzheimer’s dementia, and yet the prediction accuracy of dementia onset using chronological age alone is limited due to person-specific differences in psychosocial (e.g. education) and disease-related factors (e.g., neuropathologies). Our metric of cognitive age leverages cognitive performance, a key output of brain function, as well as chronological age. As such, chronologic age for a given cognitive age can vary widely depending on an individual’s cognitive function. Consequently, while on the cognitive age scale we restricted our prediction interval to 6 years (i.e., from cognitive age 85 to cognitive age 91), the predicted time horizon of Alzheimer’s dementia onset differs from person to person with a widespread distribution. As such, our model was able to provide person-specific predictions.

This study has many strengths. ROSMAP participants entered the study without dementia and were followed with uniform detailed evaluations each year up until death. The follow-up rates among the survivors exceeded 90%. Over the course of the study, which spanned several decades, some participants remained cognitively unimpaired, some developed mild cognitive impairment, and others were diagnosed with dementia. The design provides longitudinal data that are densely spaced, and cover the entire spectrum of cognitive aging from normality to dementia. These high-quality longitudinal data, coupled with a novel analytic approach, facilitate the robust characterization of a cognitive clock which serves as the backbone for the proposed prediction algorithm. Second, our model uses basic cognitive data that can be readily and inexpensively collected during life. Therefore, the model can be easily applied to Alzheimer’s dementia prediction by other cohorts and in other settings. Indeed, the person-specific prediction accuracy in this work is comparable to that of a recent prediction model that uses amyloid PET imaging.

Limitations are noted. Participants in the study are on average older and may not be representative of the general aging population. Further, ROSMAP are cohorts of predominantly non-Latino Whites, and the extent to which cognitive aging in minority populations follows the same clock remains to be confirmed. However, we note that the generalizability of the clock has previously been validated in an independent biracial population-based cohort. In addition, prediction using cognitive age depends on the accurate profiling of cognitive aging. Although widely used, the MMSE and the cognitive age measure derived from it are not sensitive for detecting MCI. Importantly, however, the method can be adapted to more comprehensive testing batteries for that purpose. The current work focused on the clinical application of cognitive age. We previously reported that older adults whose cognitive age was older than their chronological age had a higher burden of neuropathologies, including AD and other neuropathologic conditions. The extent to which common neuropathologies are implicated in cognitive age is not fully understood and we think warrants a separate investigation. Finally, our findings provide strong evidence that the proposed cognitive clock robustly captures the core pattern of late life cognitive decline and that the derived cognitive age can be applied to various cohorts. Nonetheless, for the cognitive age to be clinically applicable, more extensive validation is required and additional studies are needed to determine how the measure can effectively be used by healthcare professionals.

RESEARCH IN CONTEXT.

Systematic Review

The literature review was conducted using PubMed. Intervention of Alzheimer’s dementia hinges on early diagnosis and advanced planning, which require accurate prediction of age at dementia onset at a person-specific level. Cerebrospinal fluid and neuroimaging biomarkers for Alzheimer’s disease (AD) are invaluable for early detection of abnormal brain changes. Alternative approaches using easily accessible cognitive or behavioral measures are also needed for broader applicability.

Interpretation

By leveraging the cognitive clock, a novel indicator of brain health, we developed a dementia prediction model using only chronological age and MMSE scores. Our prediction accuracy is on par with a recent model that used amyloid positron emission tomography.

Future directions

This study offers an easy-to-use tool for predicting person-specific age at Alzheimer’s dementia onset. It informs on the time horizon for at-risk older adults and facilitates better healthcare and financial planning. Our model will be adapted for predicting mild cognitive impairment.