Predicting Brain Amyloid Status Using the National Institute of Health Toolbox (NIHTB) for Assessment of Neurological and Behavioral Function

Abstract

Background

Amyloid-beta (Aβ) plaque is a neuropathological hallmark of Alzheimer's disease (AD). As anti-amyloid monoclonal antibodies enter the market, predicting brain amyloid status is critical to determine treatment eligibility.

Objective

To predict brain amyloid status utilizing machine learning approaches in the Advancing Reliable Measurement in Alzheimer's Disease and Cognitive Aging (ARMADA) study.

Design

ARMADA is a multisite study that implemented the National Institute of Health Toolbox for Assessment of Neurological and Behavioral Function (NIHTB) in older adults with different cognitive ability levels (normal, mild cognitive impairment, early-stage dementia of the AD type).

Setting

Participants across various sites were involved in the ARMADA study for validating the NIHTB.

Participants

199 ARMADA participants had either PET or CSF information (mean age 76.3 ± 7.7, 51.3% women, 42.3% some or complete college education, 50.3% graduate education, 88.9% White, 33.2% with positive AD biomarkers).

Measurements

We used cognition, emotion, motor, sensation scores from NIHTB, and demographics to predict amyloid status measured by PET or CSF. We applied LASSO and random forest models and used the area under the receiver operating curve (AUROC) to evaluate the ability to identify amyloid positivity.

Results

The random forest model reached AUROC of 0.74 with higher specificity than sensitivity (AUROC 95% CI:0.73 -0.76, Sensitivity 0.50, Specificity 0.88) on the held-out test set; higher than the LASSO model (0.68 (95% CI:0.68 – 0.69)). The 10 features with the highest importance from the random forest model are: picture sequence memory, cognition total composite, cognition fluid composite, list sorting working memory, words-in-noise test (hearing), pattern comparison processing speed, odor identification, 2-minutes-walk endurance, 4-meter walk gait speed, and picture vocabulary. Overall, our model revealed the validity of measurements in cognition, motor, and sensation domains, in associating with AD biomarkers.

Conclusion

Our results support the utilization of the NIH toolbox as an efficient and standardizable AD biomarker measurement that is better at identifying amyloid negative (i.e., high specificity) than positive cases (i.e., low sensitivity).

Abbreviations

Aβ

Amyloid-beta

NFT

neurofibrillary tangles

AD

Alzheimer's disease

MCI

mild cognitive impairment

NIHTB

National Institute of Health Toolbox for Assessment of Neurological and Behavioral Function

ARMADA

Advancing Reliable Measurement in Alzheimer's Disease and Cognitive Aging

PET

positron emission tomography

CSF

cerebrospinal fluid

EPS

Extrapyramidal signs

SMD

standardized mean differences

PCA

principal component analyses

PC

principal components

LASSO

Least Absolute Shrinkage and Selection Operator

AUROC or AUC

Receiver Operating Characteristics Area Under Curve

CI

confidence interval

NU

Northwestern University

UM

University of Michigan

UW-M

University of Wisconsin-Madison

MCF

Mayo Clinic-Jacksonville, Florida

U Pitt

University of Pittsburgh

UCSD

Emory University, University of California-San Diego

MGH

Massachusetts General Hospital

IRB

Institutional Review Boards

Introduction

Alzheimer's disease (AD) is the most prevalent cause of dementia, significantly impacting individuals' cognition, daily functioning, and life expectancy (1). The advent of anti-amyloid therapies (2) and the revised criteria for diagnosis and staging of AD (3) underscore the need for accurate measurement of brain amyloid levels. However, the current gold standard for measuring brain amyloid —utilizing positron emission tomography (PET) (4) or cerebrospinal fluid (CSF) (5)—poses challenges due to their substantial financial burdens and invasiveness. In contrast to traditional biomarker measures, neuropsychological tests, which are non-invasive, cost-effective, and easily accessible offer a promising alternative. Previous neuropsychological tests demonstrated correlations between cognitive functions and brain amyloid burden (6, 7, 8), making them particularly appealing for widespread screening and monitoring of amyloid status.

The National Institute of Health Toolbox for Assessment of Neurological and Behavioral Function (NIHTB) is a comprehensive set of neuropsychological tests developed to assess a broad spectrum of functions (i.e., in cognitive, emotional, motor, and sensory domains) in the general population spanning ages 3 to 85 years (9, 10). Notably, previous applications of the NIHTB cognitive battery have demonstrated its efficacy in distinguishing among different cognitive groups of older adults, such as those experiencing so-called normal aging, those with mild cognitive impairment (MCI) and those with early-stage dementia of the AD type, as well as those who were considered as “SuperAgers” (11, 12). Studies using the NIHTB cognitive battery have shown an association between cognitive scores and tau deposition, rather than amyloid deposition, in non-demented older adults (13), while the NIHTB Face Name Associative Memory Exam (FNAME) score indicated an association with amyloid status in people across the clinical spectrum (14). Given that both tau and amyloid are hallmark pathologies of AD (15), the findings call for a deeper exploration of how brain amyloid status affects a range of cognitive functions measured by the NIHTB, aside from those such as associative memory which are already established. Furthermore, it raises the possibility of studying the association of brain amyloid with other domains, such as the sensory, motor, and emotional domains assessed by the NIHTB.

In the present study, we evaluate performance on discriminating amyloid positive and negative groups using cognitive, sensory, motor, and emotional functions, as part of Advancing Reliable Measurement in Alzheimer's Disease and Cognitive Aging (ARMADA), a multisite study carried out to validate the NIHTB in evaluating the cognitive decline in older adults (16). Baseline characteristics of the ARMADA study cohorts were summarized in a recent paper (17). In the following overview, we outline the relationship between each of the NIHTB domains and AD pathology, with a focus on the amyloid pathology, based on previous literature.

The relationship between cognition and Alzheimer's pathology varies across groups and cognitive stages. In healthy adults, multiple studies connect amyloid deposition with cognitive functions such as episodic memory (18), semantic memory (19), and language (20), although some research suggests a stronger correlation with executive function (21). In preclinical AD, associations between amyloid and cognitive domains that cover episodic memory, semantic memory, and executive function are also inconsistent (21, 22, 23, 24, 25, 26), with gender-specific effects (26). Studies involving individuals diagnosed with MCI have shown links between amyloid and various cognitive functions including episodic memory, language, attention, and visuospatial function (18, 27). In those with dementia due to AD, results vary; some studies associate amyloid with both memory and non-memory functions (18, 28), while others suggest weak or negligible connections with these functions (27, 29, 30). Cross-group comparisons showed that amyloid levels were related to episodic memory distinctions between AD-type dementia and control groups (28), and between amyloid-positive (including non-demented elderly, MCI, and AD dementia) and amyloid-negative groups, particularly in verbal and visual episodic memory (18). We appreciate the suggestion that the complexity of amyloid's association with cognition across different stages may stem from variations in amyloid deposition levels, as indicated by the ‘Jack' curves model (31). Alternatively, the differences we observed among groups at various cognitive stages can be driven by differences in the average cognitive scores when amyloid level variances are accounted for, that is, cognitive performance aligns with the amyloid burden, with this alignment becoming more evident when groups are contrasted across different cognitive stages.

Research findings linking motor function impairments to AD pathology, such as amyloid plaques and neurofibrillary tangles (NFTs), have been consistent, utilizing both postmortem analyses and in vivo studies (32, 33). Gait impairments, specifically, have been correlated with NFTs in the substantia nigra, regardless of dementia status (34). AD-related motor symptoms, including bradykinesia and parkinsonian gait, are also associated with pathologies in the substantia nigra, such as α-synuclein aggregation (34). Pathology is not confined to the substantia nigra but also affects other motor-related brain regions, including the primary motor cortex (35) and striatum (36). In vivo research supports these findings, linking amyloid beta with body mass index (BMI) and gait speed in cognitively normal adults (37), and in preclinical and MCI individuals (38). Despite the clear association between AD pathology and motor dysfunction, the relationship's nuances across different cognitive stages have yet to be clarified, suggesting a need for further research.

The relationship between sensory functions and AD pathology is complex and varies across different senses. Hearing loss in the elderly shows mixed correlations with amyloid but is regularly linked to tau levels, regardless of AD dementia stage (39, 40, 41). Postmortem research corroborates the correlations by identifying amyloid accumulation specifically within the auditory cortices of patients with AD dementia (42). Animal studies also suggest a connection between auditory deficits and amyloid levels (43). Olfactory impairments observed in the early stages of AD (44) align with postmortem discoveries of amyloid and tau accumulations in areas related to olfaction, including the olfactory epithelium (45), olfactory bulb and tract (46, 47), as well as the entorhinal cortex (46); additionally, a recent ARMADA study found lower olfactory identification scores in individuals with positive AD biomarkers (48). Visual impairments, which affect specific visual functions, are particularly prevalent in certain subgroups of AD, such as those diagnosed with posterior cortical atrophy (PCA) (49); amyloid and tau deposits within primary visual structures, however, have been identified in various forms of AD, not limited to PCA alone (49). Additionally, chronic pain is noted for its pathological concurrence with AD (50). These sensory impairments, particularly in hearing and smell, highlight their potential as early markers for AD risk and progression.

Research has demonstrated a significant association between AD pathologies and emotional responses. In amyloid-positive older adults, there was a tendency for increased emotional reactivity with age, alongside a reduction in perceived interpersonal warmth, notably in females (51). Additionally, amyloid-positive individuals exhibited greater emotional empathy concern rated by the informant compared to amyloid-negative healthy controls (52). Animal studies further support these findings; amyloid accumulation in mice's amygdala and anterior insula is correlated with fear and anxiety symptoms (53), while mice with amyloid depositions exhibited depressive-like behaviors (54) and emotional changes (55). Mice with both amyloid and tau pathologies show increased anxiety, despite normal sensorimotor functions (56). Although there is relatively less evidence regarding the association between AD pathology and emotion compared to other functions, such as cognition or sensations, it is intriguing to explore how different emotions, both positive and negative, correlate with AD pathologies and the strength of these connections when considering other functional domains simultaneously.

The primary objective of this study is to assess the effectiveness of the NIHTB, a promising neuropsychological tool, for predicting amyloid status. Additionally, we aim to determine the relative significance of association with amyloid status of four major domains—cognition, motor skills, sensation, and emotion—factors that are not comprehensively addressed in commonly employed tests like Clinical Dementia Rating (57) and Functional Activities Questionnaire (58). By utilizing data on brain amyloid status and NIHTB scores obtained from a subset of participants in the ARMADA study, we seek to employ machine learning techniques to predict amyloid status with NIHTB scores. To the best of our knowledge, this study is the first to comprehensively assess all four functional domains concurrently, shedding light on their relationships with AD pathology.

Methods

Measurements

AD biomarker collection

We utilized data from a subset of the ARMADA cohort whose amyloid status information was available. In the ARMADA study, amyloid status was measured by positron emission tomography (PET) or lumbar puncture, detailed in Rentz et al. (2023) and Weintraub et al., (2022). Sites followed harmonized criteria for determining amyloid PET positivity or CSF results. CSF measures used Luminex or Elisa with amyloid-beta 42/tau ratio cutoffs1 , while PET scans used Pittsburgh Compound B or florbetapir, guided by specific volumetric or uptake value ratios against the cerebellar cortex baseline2 . A 24-month timeframe was established for the stability of negative biomarker results considering the amyloid levels were expected to remain relatively constant in this duration (59), while positive results were exempt from this temporal constraint.

NIHTB scores

The NIHTB evaluates a range of normal functions across cognitive, emotional, motor, and sensory domains (10, 17), with changes in these functions during adulthood potentially signaling cognitive decline. It includes cognitive tests for executive function, attention, memory, language, processing speed, and working memory; motor tests for dexterity, strength, balance, mobility, and endurance; sensory tests for hearing, vision, smell, and pain; and emotional assessments via self-report surveys. In this study, raw scores or uncorrected scores from tests in the 4 NIHTB domains (i.e., cognition, emotion, motor, sensation) were used for the analyses. In the final analyses, there were a total of 41 test scores, including 10 test scores in the cognitive domain, 7 test scores in the motor domain, 5 test scores in the sensory domain, and 19 test scores in the emotional domain. The comprehensive assessment offered by the NIH Toolbox encompasses approximately 2 hours of administration time across its 4 domain modules (16), striking a balance between thoroughness and efficiency in neuropsychological testing.

Exploratory Data Analyses

All analyses and plots were conducted using R (version 4.2.1).

Class imbalance assessment

To understand how the data were distributed between amyloid positive and negative groups, we assessed standardized mean differences (SMD) between groups for each variable using the ‘CreateTableOne' function in the “tableone” package (version 0.13.2). These variables include demographics (i.e., age, sex, ethnicity, race, education, and handedness) and scores in each NIHTB neuropsychological test. We used a SMD threshold of 0.1 to indicate the balance of each measurement, i.e., if the SMD value is between −0.1 and 0.1, it indicates a good balance of the measurement.

To understand whether the differences between amyloid positive and negative groups were statistically significant, we used logistic regression to test the association of each variable and the amyloid groups using the ‘glm' function in the “stats” package (version 4.2.1). Each variable, including demographic factors and scores from the NIHTB neuropsychological tests, served as an independent variable in these analyses. Age, sex, and education level were accounted for as covariates to mitigate their potential confounding effects, with the amyloid status designated as the outcome variable. Reported results were adjusted for multiple comparison (60).

Correlation analyses

To understand the relationship among neuropsychological tests, we first conducted bivariate Pearson correlation for tests within each of the NIHTB domains, respectively. The analyses were conducted using the “cor” function in the “stats” package (version 4.2.1) and results were plotted using the “heatmap” function in the R «ComplexHeatmap» package (version 2.15.1).

Principal Component Analyses

To better understand data variability within each NIHTB domain, we conducted principal component analyses (PCA). Scores from each test (i.e., each variable) were preprocessed first by being centered around 0 to remove bias or offset in the data, then being scaled to ensure each test score has a standard deviation of 1 to make all variables were comparable in terms of their contribution to the model. Next, the ‘prcomp' function in the “stats” package (version 4.2.1) was used to perform PCA on scores from each NIHTB domain, respectively. All principal components with eigen values above 1 were extracted from the corresponding eigenvectors produced by PCA using the ‘get_pca_var' function in the “factoextra” package (version 1.0.7). Then we calculated the loadings for each variable on each principal component to assess the contribution of each variable to the qualified principal components.

Our utilization of correlation analyses and PCA was driven by the need to thoroughly understand the underlying structure and inter-relationships of cognitive measures within the NIHTB neuropsychological tests. While these analyses did not directly inform the feature selection for the machine learning models used in this study, they served a foundational role in characterizing the relationship of cognitive abilities with each other The logistic regression model was chosen. Recognizing the unique contributions of each test, we included all variables in the machine learning model to ensure an all-encompassing evaluation. The insights garnered from these exploratory steps are vital for our future research directions, which include refining the test battery and enhancing the efficiency of cognitive assessments.

Machine learning classification

Binary classifiers

Our binary classification approach to determining participants' amyloid status employs two distinct models: logistic regression and a random forest. The logistic regression model was chosen for its interpretability and established efficacy in clinical research, while random forest classification model was chosen for its robustness and ability to handle complex datasets with diverse feature types. In addition, both models had proficiency in yielding output probabilities and feature importance. Logistic regression gauges feature importance via the magnitude of predictor coefficients, whereas the random forest assesses it by the impurity of which measures lacked homogeneity within the resulting nodes.

The dataset comprised a total of 47 features, encompassing 6 demographic variables and 41 neuropsychological tests that tap cognitive, motor, sensory and emotional domains. To standardize the data, we performed centering around 0 and scaling with a standard deviation of 1, akin to the preprocessing step in PCA analyses.

For each binary classification task, we randomly partitioned the dataset into a training set comprising 75% of the data and a test set with the remaining 25%. The training process incorporated a threefold repetition of 5-fold cross-validation to fine-tune the models, ensuring each fold maintained an equal ratio of amyloid-positive cases, which was approximately 33%. To evaluate the performance of each model, we calculated the average metrics from these cross-validation iterations. This approach provided a robust assessment of the models' capability to accurately identify amyloid positivity while maintaining independence between the training and test sets.

In the logistic regression model, L1 (Least Absolute Shrinkage and Selection Operator or LASSO) regularization was applied to address potential multicollinearity and prevent overfitting. The tuning parameter, lambda, was systematically evaluated over a range from 0.001 to 0.1, with increments of 0.001, to determine the optimal level of regularization. With L1 regularization, only features that exhibit a strong relationship with predictors are selected for making predictions.

In the random forest classification, the optimal number of predictor variables considered for splitting at each tree node was determined using a grid search that spanned from 1 to 10 in increments of 1. The number of trees in the random forest was set to 500, and all available features were used in the model.

The modeling process was conducted using the «caret» package (version 6.0.93) and the feature importance ranking figure was generated using the “ggplot2” package (version 3.4.0).

Classifier performance evaluation

For both classifiers, we reported accuracy of predictions using Receiver Operating Characteristics Area Under Curve (AUROC) along with its 95% confidence intervals that were estimated with 2000 bootstraps. Then we examined sensitivity and specificity on the test set.

In this study, missing data was replaced with the mean of their respective columns, ensuring that our machine learning algorithms—random forest and logistic LASSO regression—could leverage a complete dataset. This approach was chosen to preserve the integrity and breadth of our analysis. We recognize that mean imputation may reduce data variability and potentially impact the model's predictive capacity. Nevertheless, the algorithms we employed are designed to mitigate such effects through their inherent mechanisms, such as regularization in logistic LASSO regression and the ensemble nature of random forests, which generally diminish the influence of minor variances.

Results

Cohort characteristics

Data used in the preparation of this article was obtained from the ARMADA study. In the ARMADA study, participants were older adults with different levels of cognitive ability clinically characterized as normal-for-age, mild cognitive impairment or early stage dementia of the AD type. Participants were enrolled from 9 Alzheimer's Disease Research Centers (ADRCs) and affiliated related studies of cognitive aging studies, including: Northwestern University (NU, lead site), University of Michigan (UM), University of Wisconsin-Madison (UW-M), Mayo Clinic-Jacksonville, Florida (MCF), University of Pittsburgh (U Pitt), Emory University, University of California-San Diego (UCSD), Columbia University, and Massachusetts General Hospital (MGH). Informed consent was obtained at each ADRC in accordance with the approval of their respective Institutional Review Boards (IRBs). Please see Weintraub et al., (2022) for recruitment details and Karpouzian-Rogers et al. (2023) for the baseline characterization of the ARMADA cohort3 .

Our final cohort comprised 199 participants who underwent PET or CSF measurement from the total pool of 462 ARMADA participants. There were no missing data in the demographic variables. The mean age of this cohort was 76.3 ± 7.7 years. In terms of sex distribution, 51.3% are women. Additionally, 42.3% had completed some or all of their college education, while 50.3% had attained a graduate level of education. A majority of the cohort (88.9%) identified as White. Within this cohort, 33.2% (N = 66) of the participants displayed positive AD biomarkers. There was no significant difference in age, sex, ethnicity, race, handedness, or education between the amyloid positive and amyloid negative groups (measured by statistical differences, ps > 0.05).

Regarding neuropsychological test scores, there were 2.41% missing data in the cognition module, 0.21% missing data in the emotion module, 5.6% missing data in the motor module, and 2.31% missing data in the sensation module. 61 people, accounting for 30.7% of our sample, had at least one feature with missing data.

The amyloid-positive group scored lower on all tests within the cognition module and in most tests within the motor and sensation modules (SMDs < −0.1). Most of the difference in the scores in the cognition module (i.e., cognition fluid composite score, cognition total composite score, domain dimensional change card sort test, list sorting working memory, pattern comparison processing speed, picture sequence memory, and picture vocabulary test) and 2 scores in the sensation module (i.e., odor identification test, words-in-noise test-right) were statistically significant and no tests in the motor module were significant, after controlling for age, sex, and education (ps > 0.05). Thus, despite SMDs indicating lower performance in most of the motor and sensation modules, these did not translate into statistical significance post-adjustment for age, sex, and education. This suggests that while the effect size points to a potential trend, the variations within these groups may not be adequately captured by significance testing alone, potentially due to sample size or variability. Similarly, within the emotion module, the amyloid-positive group exhibited both higher scores in certain tests (anger -physical aggression, general life satisfaction, instrumental support, meaning and purpose, positive affect, and sadness; SMDs > 0.1) and lower scores in loneliness (SMD = −0.11) and self-efficacy (SMD = −0.17), although none of them were statistically significant after controlling for age, sex, and education (ps > 0.05). These findings underscore the necessity to interpret SMDs in conjunction with p-values, as they provide complementary insights into the clinical relevance of the test scores. The lack of statistical significance, despite notable SMDs, suggests that larger sample sizes or more sensitive measures may be required to fully understand the implications of amyloid positivity on motor, sensory, and emotional functions. For a detailed breakdown of demographic information and NIHTB scores, please refer to Table 1.

Table 1.

Summary Statistics of the Demographics and NIHTB Scores of the ARMADA Participants

Characteristics	Total (n = 199)	Missing, N (%)	Amyloid Negative (n = 133; 66.8%)	Amyloid Positive (n = 66; 33.2%)	Standardized Mean Difference	Adjusted Statistical Difference
Age, mean (SD), years
	76.31 (7.74)	0 (0)	75.78 (7.57)	77.36 (8.03)	0.2	0.293
Sex, N (%)
Female	102 (51.3)	0 (0)	72 (54.1)	30 (45.5)	−0.09	0.234
Male	97 (48.7)	0 (0)	61 (45.9)	36 (54.5)	0.09	0.234
Ethnicity, N (%)
Not Hispanic or Latino	192 (96.5)	0 (0)	61 (92.4)	131 (98.5)	−0.06	0.058
Hispanic or Latino	7 (3.5)	0 (0)	5 (7.6)	2 (1.5)	0.06	0.058
Race, N (%)
White	177 (88.9)	0 (0)	118 (88.7)	59 (89.4)	0.01	0.412
Black or African American	15 (7.5)	0 (0)	12 (9.0)	3 (4.6)	−0.04	0.431
Asian	2 (1.0)	0 (0)	1 (0.8)	1 (1.5)	0.01	0.688
American Indian or Alaska Native	1 (5.0)	0 (0)	1 (0.8)	0 (0.0)	−0.01	0.992
Other	2 (1.0)	0 (0)	0 (0.0)	2 (3.0)	0.03	0.987
More than one race	2 (1.0)	0 (0)	1 (0.8)	1 (1.5)	0.01	0.678
Education, N (%)
High school or less	15 (7.5)	0 (0)	8 (6.0)	7 (10.6)	0.05	0.333
Some college	21 (10.6)	0 (0)	10 (7.5)	11 (16.7)	0.09	0.502
College	63 (31.7)	0 (0)	40 (30.1)	23 (34.8)	0.05	0.73
Graduate	100 (50.3)	0 (0)	75 (56.4)	25 (37.9)	−0.19	0.159
Handedness, N (%)
Right	182 (91.5)	0 (0)	124 (93.2)	58 (87.9)	−0.05	0.301
Left	17 (8.5)	0 (0)	9 (6.8)	8 (12.1)	0.05	0.301
Cognition, mean (SD)
Cognition Crystallized Composite	114.85 (7.68)	2 (1.01)	116.11 (7.73)	112.31 (6.99)	−0.52	0.061
Cognition Fluid Composite	86.48 (12.75)	13 (6.53)	89.82 (11.25)	79.31 (12.93)	−0.87	<0.001
Cognition Total Composite	100.5 (10.23)	13 (6.53)	103.16 (9.4)	94.78 (9.65)	−0.88	<0.001
Domain Dimensional Change Card Sort Test	28.24 (3.67)	1 (0.5)	28.91 (2.6)	26.89 (4.94)	−0.51	0.035
Flanker	19.56 (1.85)	0 (0)	19.86 (0.85)	18.95 (2.91)	−0.42	0.061
List Sorting Working Memory	14.7 (3.77)	6 (3.01)	15.56 (3.31)	12.9 (4.07)	−0.72	0.001
Oral Reading Recognition	111.38 (6)	2 (1.01)	112.05 (5.81)	110.02 (6.2)	−0.34	0.241
Pattern Comparison Processing Speed	34.33 (7.87)	0 (0)	36 (7.03)	30.95 (8.42)	−0.65	0.003
Picture Sequence Memory	8.33 (6.61)	11 (5.53)	9.68 (6.71)	5.45 (5.4)	−0.69	0.003
Picture Vocabulary	116.97 (9.72)	0 (0)	118.67 (9.87)	113.55 (8.47)	−0.56	0.044
Emotion,mean (SD)
Anger - Affect	12.83 (2.26)	1 (0.5)	12.81 (2.2)	12.88 (2.4)	0.03	0.816
Anger - Hostility	7.28 (3.38)	1 (0.5)	7.29 (3.41)	7.27 (3.36)	0.00	0.848
Anger - Physical Aggression	5.59 (1.64)	1 (0.5)	5.49 (1.16)	5.77 (2.33)	0.15	0.822
Emotional Support	33.2 (5.44)	0 (0)	33.29 (5.27)	33.02 (5.8)	−0.05	0.832
Fear - Affect	8.93 (2.29)	1 (0.5)	8.92 (2.33)	8.95 (2.21)	0.02	0.973
Fear - Somatic Arousal	8.22 (2.33)	1 (0.5)	8.2 (2.12)	8.24 (2.71)	0.02	0.880
Friendship	32 (6.2)	0 (0)	31.97 (6.26)	32.06 (6.12)	0.01	0.822
General Life Satisfaction	24.48 (10.64)	0 (0)	23.98 (9.83)	25.48 (12.14)	0.14	0.269
Instrumental Support	32.42 (7.97)	0 (0)	31.67 (8.41)	33.94 (6.8)	0.30	0.083
Loneliness	8.53 (3.2)	0 (0)	8.64 (3.31)	8.3 (2.97)	−0.11	0.704
Meaning and Purpose	26.69 (14.11)	0 (0)	25.59 (13.37)	28.91 (15.38)	0.23	0.168
Negative Affect	45.86 (8.3)	1 (0.5)	45.71 (8.09)	46.17 (8.76)	0.05	0.880
Perceived Hostility	14.03 (4.18)	0 (0)	14.04 (4.15)	14 (4.28)	−0.01	0.848
Positive Affect	18.61 (11.4)	0 (0)	17.95 (10.22)	19.94 (13.45)	0.17	0.364
Perceived Rejection	12.79 (4.23)	0 (0)	12.69 (4.33)	13 (4.03)	0.07	0.880
Perceived Stress	21.25 (5)	1 (0.5)	21.27 (4.81)	21.2 (5.4)	−0.01	0.822
Sadness	8.99 (2.6)	1 (0.5)	8.89 (2.56)	9.2 (2.68)	0.12	0.596
Self-Efficacy	20.23 (9.58)	0 (0)	20.75 (10)	19.17 (8.64)	−0.17	0.704
Social Satisfaction	50.1 (8.97)	0 (0)	49.89 (8.8)	50.52 (9.36)	0.07	0.589
Motor, mean (SD)
Standing Balance Test	93.91 (10.94)	47 (23.62)	95.06 (10.58)	90.98 (11.42)	−0.37	0.209
Grip Strength Test - Dominant Hand	96.28 (9.33)	2 (1.01)	96.26 (9.13)	96.32 (9.81)	0.01	0.879
Grip Strength Test - Non-dominant Hand	96.01 (9.45)	2 (1.01)	96.44 (9.37)	95.14 (9.63)	−0.14	0.263
9-Hole Pegboard Dexterity Test - Dominant Hand	90.8 (14.19)	1 (0.5)	92.33 (13.61)	87.66 (14.91)	−0.33	0.269
9-Hole Pegboard Dexterity Test - Non-dominant Hand	93.22 (11.67)	1 (0.5)	94.44 (10.9)	90.71 (12.85)	−0.31	0.263
2-Minute Walk Endurance Test	515.14 (101.87)	14 (7.04)	526.41 (99.39)	491.65 (103.8)	−0.34	0.263
4-Meter Walk Gait Speed Test	3.81 (0.87)	11 (5.53)	3.8 (0.84)	3.82 (0.93)	0.03	0.704
Sensation, mean (SD)
Odor Identification Test	6.36 (1.98)	8 (4.02)	6.78 (1.76)	5.5 (2.14)	−0.65	0.003
Pain Interference	8.78 (3.06)	1 (0.5)	8.73 (3)	8.89 (3.18)	0.05	0.809
Visual Acuity Test	79.72 (6.77)	4 (2.01)	79.92 (5.74)	79.31 (8.52)	−0.08	0.973
Words-In-Noise Test - Left	14.42 (7.6)	5 (2.51)	15.75 (7.1)	11.77 (7.91)	−0.53	0.051
Words-In-Noise Test - Right	16.09 (7.47)	5 (2.51)	17.7 (6.29)	12.89 (8.59)	−0.64	0.004

A standardized mean difference (SMD) threshold of 0.1 was employed to assess the equilibrium of each metric. Measurements with an SMD exceeding 0.1, highlighted in bold, signify a comparative lack of balance. The amyloid positive group had lower scores in half of the tests from the cognition and two tests from the sensation modules (SMDs < −0.1, ps < 0.05). In terms of the emotion module, the amyloid positive group had higher scores in some tests (Anger - Physical Aggression, General Life Satisfaction, Instrumental Support, Meaning and Purpose, Positive Affect, and Sadness; SMDs > 0.1) and lower scores in loneliness (SMD = −0.11) and self-efficacy (SMD = −0.17) although none of them were statistically significant (ps > 0.05). All statistical tests were reported after controlling for age, sex, and education. Please note that the beginning line for each demographics and module have been intentionally left blank. This formatting choice is designed to enhance the readability of the table by clearly distinguishing between different demographic features and modules.

Measurements of cognition, emotion, motor, and sensation

Cognition

The correlation analyses revealed relationships between various cognitive measures. Notably, the cognition total composite score, which exhibited greater reliability in gauging overall cognitive strength compared to individual tests, demonstrated the strongest correlation with fluid cognition (cognition fluid composite score) (r = 0.9), as expected. It also displayed robust correlations with crystallized cognition (cognition crystallized composite score), language (picture vocabulary), working memory (list sorting working memory), and episodic memory (picture sequence memory) (rs = 0.7). The crystallized cognition (cognition crystallized composite), representing intellectual abilities acquired over time, exhibited high correlations (rs = 0.9) with language abilities, encompassing both vocabulary (picture vocabulary) and reading (oral reading cognition).

It is worth noting that two subdomains of executive function: set shifting (domain dimensional change card sort test) and inhibitory control and attention (Flanker test) primarily showed a substantial correlation with each other (r = 0.6), but demonstrated comparatively weaker associations with the third subdomain of executive function - working memory (r = 0.2 for set shifting and r = 0.4 for inhibitory control) and other cognitive functions, especially with language (rs = 0.2 for reading and rs = 0.3 for vocabulary). Detailed information can be found in Figure 1a.

Figure 1.

Correlation and PCA contribution of the cognition module

1a shows positive correlation among tests in the cognition module. Particularly, cognition crystallized composite score was highly correlated with picture vocabulary score (r = 0.9) and oral reading cognition score (r = 0.9). Cognition total composite score was also highly correlated with cognition fluid composite score (r = 0.9). 1b shows Flanker test score and domain dimensional change card sort test score had the highest principal component contribution in the cognition module. The loading values represent the contribution of each variable to the principal components, and they are expressed as a proportion of the total variance explained by each component. These loading values are depicted visually through both the size and color intensity of the points on the plot. Specifically, the color of the points transitions from yellow (indicating a lower contribution) to red (indicating a higher contribution), and the size of the points increases with the contribution value. The scale for both color and size ranges from 1 to 60, with specified breaks at 20, 40, and 60 to aid in interpretation. It's important to note that these values do not reflect absolute magnitudes but rather relative contributions within our dataset. Therefore, a point with a value of 60 represents the variable with the highest contribution to the principal component within this dataset, rather than a fixed upper limit or maximum absolute contribution.

In order to grasp the underlying patterns shaping the data within the cognition module, we examined the distinct contributions of each cognitive ability to the overall variance of the dataset (refer to Figure 1b). Three factors explained the cognitive battery. While the first and second principal components (PC1 and PC2) showed low loadings for each variable, PC3 was notably characterized by robust loadings from both inhibitory control and set shifting subdomains of executive functions (Flanker test and domain dimensional change card sort test).

Motor

As anticipated, the two upper extremity strength measurements exhibited the highest correlation within the motor domain (correlation between grip strength in the dominant and non-dominant hands = 0.9). This was followed by the correlation between the two dexterity measurements (correlation between 9-hole pegboard dexterity test in the dominant and non-dominant hands = 0.8). As expected, the 4-meter walk gait speed test (measured in time), which evaluates locomotion, displayed inverse correlations with all other motor abilities, indicating that individuals with slower gait speed tended to demonstrate lower performance in endurance, strength, dexterity, and balance. Moreover, strength displayed negligible or no correlation with dexterity (9-hole pegboard dexterity test rs = 0 and 0.1) and balance (standing balance test rs = 0 and 0.1), while endurance exhibited moderately positive correlations with all other motor skills (strength, dexterity, and balance; rs = 0.3 and 0.4) except for locomotion. Detailed information is available in Figure 2a.

Figure 2.

Correlation and PCA contribution of the motor module and the sensation module

2a shows high correlations between dominant hand and non-dominant hand in grip strength test scores (r = 0.9) and 9-hole pegboard dexterity test scores (r = 0.8). The 4-meter walk gait speed test score (measured in time) was negatively correlated with all other tests in the motor module (r ranges between −0.2 and −0.6). 2b shows 2-minute walk endurance test, grip strength test, and 9-hole pegboard dexterity test had the highest principal component contribution in the motor module. 2c shows relatively high correlation between words-in-noise test left and words-in-noise test right scores (r = 0.7) while correlations among other tests were all fairly low (r ranges between −0.1 and 0.3). 2d shows words-in-noise tests and pain interference scores had the highest principal component contribution in the sensation module. In both 2b and 2d, the loading values represent the contribution of each variable to the principal components, and they are expressed as a proportion of the total variance explained by each component. These loading values are depicted visually through both the size and color intensity of the points on the plot. Specifically, the color of the points transitions from yellow (indicating a lower contribution) to red (indicating a higher contribution), and the size of the points increases with the contribution value. The scale for both color and size ranges from 1 to 60, with specified breaks at 20, 40, and 60 to aid in interpretation. It's important to note that these values do not reflect absolute magnitudes but rather relative contributions within our dataset. Therefore, a point with a value of 60 represents the variable with the highest contribution to the principal component within this dataset, rather than a fixed upper limit or maximum absolute contribution.

In the assessment of the motor modules, our PCA revealed 2 principal components that effectively captured the variability within the data. PC1 was characterized by relatively substantial loadings of the 2-minute walk endurance test while PC2 was characterized by relatively high loadings of the grip strength test (both dominant and non-dominant hands) and the 9-hole pegboard dexterity test (dominant hand). Refer to Figure 2b for a comprehensive overview.

Sensation

As anticipated, the two auditory measurements exhibited the strongest correlation within the sensory domain, with a correlation coefficient of 0.7 observed between the left and right sides in the words-in-noise test. However, when examining scores across different sensory modalities, encompassing audition, vision, olfaction, and pain, the correlations were notably low (rs < 0.4). These findings imply that separate underlying mechanisms are influencing the decline of different sensory perceptions. Detailed information can be found in Figure 2c.

Regarding the underlying pattern within the sensory module, two factors were identified. Our observations primarily revealed a substantial loading of the pain score, as assessed by the pain interference test, on PC2. In addition, PC1 was characterized by relatively high loadings of words-in-noise tests (both left and right). Details are available in Figure 2d.

Emotion

The measurements of negative emotions, encompassing negative affect scores (fear-somatic arousal, anger-physical aggression, anger hostility, and general negative affect), stress scores (perceived stress), and social relationships (perceived rejection, perceived hostility, and loneliness), exhibited positive correlations with each other. Notably, the composite score of negative affect and perceived stress exhibited the highest correlation at r = 0.8, while the correlation scores between other pairs ranged from low to moderate (correlation scores ranging from 0.1 to 0.6).

Similarly, measurements of positive emotions, including psychological well-being (positive affect, general life satisfaction, meaning and purpose), self-efficacy, and social relationships (emotional support, instrumental support, social satisfaction, and friendship), displayed positive correlations across the board. Notably, tests evaluating social relationships— specifically social satisfaction and friendship, as well as social satisfaction and emotional support—exhibited the strongest correlations (rs = 0.8). In contrast, correlation scores between other pairs remained low to moderate (correlation scores ranging from 0.1 to 0.6).

Furthermore, negative types of emotions exhibited a general negative correlation with positive types of emotions. Notably, tests measuring negative aspects of social relationships and positive aspects of social relationships demonstrated the highest negative correlations (rs = −0.8), while other pairs exhibited low to moderate correlations (correlation scores ranging from −0.1 to −0.6). However, three tests assessing negative affect (fear-affect, anger-affect, and sadness) displayed fairly low to negligible correlations with any other tests (correlation coefficients ranging from -0.3 to 0.3). Detailed information can be found in Figure 3a.

Figure 3.

Correlation and PCA contribution of the emotion module

3a shows both positive and negative correlations among tests in the emotion module. Among negative emotion tests, negative affect score and perceived stress score were highly correlated (r = 0.8); among positive emotion tests, social satisfaction score was highly correlated with friendship score (r = 0.8) and emotional support score (r = 0.8). In contrast, social satisfaction score was highly negatively correlated with perceived rejection (r = −0.8) and loneliness (r = −0.8). 3b shows sadness, general life satisfaction, fear, anger - affect, perceived hostility, instrumental support, and anger-physical aggression scores had the highest principal component contribution in the emotion module. The loading values represent the contribution of each variable to the principal components, and they are expressed as a proportion of the total variance explained by each component. These loading values are depicted visually through both the size and color intensity of the points on the plot. Specifically, the color of the points transitions from yellow (indicating a lower contribution) to red (indicating a higher contribution), and the size of the points increases with the contribution value. The scale for both color and size ranges from 1 to 60, with specified breaks at 20, 40, and 60 to aid in interpretation. It's important to note that these values do not reflect absolute magnitudes but rather relative contributions within our dataset. Therefore, a point with a value of 60 represents the variable with the highest contribution to the principal component within this dataset, rather than a fixed upper limit or maximum absolute contribution.

In terms of the underlying pattern within the emotional module, 5 factors were identified. PC3 was characterized by high loadings of sadness score (an item within the negative affect category), PC4 exhibited relatively high loadings of fear-affect and anger-affect (items within the negative affect category) and psychological well-being (general life satisfaction) inversely, and PC5 exhibited relatively high loadings of social relationships (perceived hostility and instrumental support) and anger-physical aggression (an item within the negative affect category). Further details are available in Figure 3b.

Predictions of amyloid status

Classifier performance

In the evaluation of amyloid status prediction, the random forest model achieved an AUROC of 0.743 (95% CI: 0.728 – 0.757), exhibiting high specificity (0.88) and low sensitivity (0.50) on the test set. The corresponding ROC curve is depicted in Figure 4a. In contrast, the logistic LASSO regression model yielded an AUROC of 0.682 (95% CI: 0.676 – 0.687), displaying high specificity (0.85) but lower sensitivity (0.25). In summary, the random forest model outperformed the logistic LASSO regression model in amyloid status prediction.

Figure 4.

Random forest classification results

4a shows AUROC curves for both the random forest classification and the LASSO logistic regression models. Random forest model had better performance than the LASSO logistic model. 4b shows the feature importance ranking from the model with superior performance (i.e., random forest).

Feature importance

Given the superior performance of the random forest model, our focus for data interpretation was directed towards this model. A comprehensive list of feature importance rankings can be found in Figure 4b. Here, we emphasize the top 10 features with the highest importance: episodic memory (picture sequence memory), cognition total composite score, fluid cognition (cognition fluid composite), working memory (list sorting working memory), audition (words-in-noise test – right), processing speed (pattern comparison processing speed), olfaction (odor identification), endurance (2-minutes-walk endurance), locomotion (4-meter walk gait speed), and language (picture vocabulary). Notably, cognitive scores comprise more than half of the top 10 rankings, aligning with the prominent cognitive decline associated with aging and neurodegeneration. Following cognitive scores, the sensation domain (audition and olfaction) ranked second in importance, while motor domain features (endurance and locomotion) ranked last. Interestingly, none of the emotional scores emerged as high-ranking features in predicting amyloid status.

In addressing the concerns regarding class imbalance, we also explored several remediation strategies: down sampling, up sampling, class weighting, and Synthetic Minority Over-sampling Technique (SMOTE). Each method was rigorously evaluated for its impact on model performance metrics, maintaining consistency with the same seed for cross-validation splits. Despite observing marginal improvements in sensitivity with these approaches, we concluded that the original model—with an AUC of 0.74—provides the most appropriate balance of sensitivity and specificity for our clinical objectives. This decision is substantiated by the clinical implications of false positives and negatives, as well as the model's robustness across various performance metrics. We've documented these findings in the supplementary materials (see Supplement Table S1), ensuring transparency and reproducibility of our analyses.

Discussion

The primary objective of this study was to develop predictive models that discern brain amyloid status using results from cost-effective and efficient neuropsychological tests. We analyzed data from 199 participants in the ARMADA study who underwent NIHTB tests across four domains and had documented amyloid status. The model achieved acceptable accuracy (AUC of 0.74) and displayed high specificity but low sensitivity. This suggests that the NIHTB is appropriate for broad-based initial screenings to pinpoint individuals who may not require extensive evaluation of in-vivo biomarker or ongoing monitoring for AD risk. Furthermore, we highlighted the significance of various NIHTB test scores and demographic factors in classifying amyloid status, offering valuable information for the clinical interpretation of the classifiers.

As anticipated, cognition emerges as the primary domain correlating with amyloid status, a conclusion underscored by the high ranking of cognitive tests in feature importance. Notably, episodic memory stands out as the foremost cognitive skill linked with amyloid status. This aligns with the understanding that amyloid is a hallmark of AD pathology (15) and that amnestic MCI and dementia are typical clinical syndromes in AD4 (61). Our findings are in line with prior research emphasizing the distinction in episodic memory between groups distinguished by amyloid-positive and -negative biomarkers (18). Multiple studies have also corroborated the connection between episodic memory and amyloid accumulation across various clinical stages of MCI and dementia associated with AD (21, 22, 28, 62).

Additionally, fluid cognition—reflecting an individual's ability to process information, act, and address new challenges—proved more associated with amyloid status than crystallized cognition, which represents accumulated knowledge. This distinction is anticipated since crystallized intelligence generally rises with age (63), thereby exerting a lesser impact on AD biomarkers, especially when participants fall within a similar age range. Conversely, since fluid intelligence typically declines from early adulthood (64), it more strongly correlates with cognitive decline pathologies. It's logical, then, that other test scores like a composite cognitive score, working memory, and processing speed—which are highly correlated with the fluid cognitive composite score—were important features in relation to amyloid status.

Language, particularly picture vocabulary, emerged as a significant predictor of amyloid status. This supports prior studies linking language and semantic memory with amyloid accumulation (18, 19, 20, 24). Notably, vocabulary is more important than oral reading recognition in its association with AD. This distinction is consistent with studies that regard reading as a ‘hold test', indicative of cognitive reserve—a capacity that remains relatively resilient in the face of cognitive decline (65). Vocabulary, on the other hand, demands the retrieval of specific information, such as word definitions, which may render it more vulnerable to the neurodegenerative effects seen in AD (66).

Although executive functions—particularly set-shifting, inhibitory control, and attention—weren't prominent predictors for amyloid status, they remain noteworthy in the cognitive module data due to their strong PCA loadings. This suggests a potential linkage of these functions with other AD pathologies, such as tau (29).

Auditory assessment, specifically the «words in noise test - right,» emerged as a top predictor and the primary sensory factor linked to amyloid status. While prior research on the interplay between audition and AD pathology has yielded varied outcomes (39, 40, 67), our findings underscore the connection between auditory ability and amyloid deposition. The significance of audition in our findings aligns with research listing midlife hearing loss as one of the 12 risk factors for dementia (68). This hearing impairment can lead to social isolation, further increasing dementia risk in later life stages. Moreover, our results highlight the «words in noise test - right» as a more predictive factor compared to its left counterpart. This observation resonates with existing literature, indicating a pronounced discrepancy in hearing loss between the left and right ears in AD patients, potentially attributed neurodegenerative challenges impacting interhemispheric communication (69, 70)5 .

The odor identification test, which evaluates olfaction, ranked as the second most vital sensory predictor for amyloid status. The finding that patients with a positive AD biomarker had lower odor identification scores is consistent with the results published in another recent study from the ARMADA dataset (48). This resonates with previous studies that highlighted differences in amyloid density within the human olfactory system (45, 47). Furthermore, in vivo studies have consistently observed a decline in olfaction among AD patients (71, 72).

As for other sensory predictors, while certain visual impairments have been identified within specific AD subgroups, they might not be as prominent in the current diverse sample, with a relatively small proportion of AD participants (17). Additionally, despite «pain interference» holding significant PCA loadings in the sensory module, it emerged as the least impactful sensation related to amyloid status. This suggests that pain might be more of a common symptom of aging rather than one unique to dementia.

Regarding motor function, locomotion (4-meter walk gait speed test) and endurance (2-minute walk endurance test) both emerged as top predictors. This underscores the role of motor dysfunction in differentiating groups based on AD pathology. Such findings align with previous postmortem studies linking AD pathology to motor functions, especially gait speed (32, 33), and findings of AD pathology in brain regions related to motor functions (35, 36). Other evidence comes from recent research which draws connections between gait speed and cognitive decline (73). Moreover, endurance training like aerobic exercise has been shown to improve cognitive function in patients with dementia of the AD type (74), potentially through neuroplasticity enhancement. These findings accentuate the prospective advantages of exercises centered on locomotion and endurance for countering cognitive decline in patients with dementia of the AD type.

Surprisingly, even though age is a recognized risk factor for dementia (75), it was not among the top influential variables in our analysis. This might be attributed to the narrow age range within our study sample (refer to Table 1). Similarly, since a significant portion of ARMADA participants possess advanced educational backgrounds, education did not emerge as a critical predictor either. Other demographic factors such as gender, race, and ethnicity were also not significant, possibly due to the limited size of our study sample.

Our analysis of the emotional module yielded interesting findings. A prominent observation was the strong correlation between negative affect and perceived stress. Moreover, multiple positive attributes of social relationships were closely interrelated, while the negative aspects exhibited the strongest inverse correlations with these positive attributes. This trend can be explained by the cascade of negative emotions, like loneliness, that often follow societal-induced stressors and events, such as a pandemic, leading to heightened social isolation. This is corroborated by research indicating that individuals with preclinical AD and higher amyloid deposition reported increased loneliness (76). On the other hand, enhanced social support—be it emotional, instrumental, or through friendship—can bolster an individual's psychological well-being, underscoring the importance of minimizing social isolation for the aging population. Further, the emotional test battery in the original study was designed to measure 3 distinct underlying factors (negative affect, social satisfaction, and psychological well-being) (77), but it turns out there were 5 principal components that best fit the data in the current study. This indicates a more complex structure in the empirical data than the original design of the NIHTB emotional test battery. PC4 and PC5 each represents one of the originally intended factors separately (i.e., psychological well-being and social satisfaction), implying that these two factors were distinctive in the current cohort. Yet, the factor of negative affect was represented by 3 PCs (PC3, PC4, and PC5), which suggest that either the factor was multifaceted or the test items were interrelated in measuring the current cohort. The latter explanation was supported by the correlation analyses that negative affect were highly correlated with some positive attributes. Moreover, what is particularly intriguing is that while diverse emotions have been observed in AD patients (78, 79, 80), no emotional tests stood out in predicting amyloid status. This paves the way to explore the potential relationships between emotions and other AD pathologies, such as tau deposition (52, 56).

In the broader research context, our ARMADA study's machine learning models, utilizing behavioral measures from the NIHTB, demonstrated an AUROC of 0.74, situating our findings within the diverse methodologies of current amyloid-beta (Aβ) status prediction research. A study employing demographics and cognitive assessments reported an AUROC of 0.73 (81), illustrating the utility of behavioral measures. Similarly, research using domain-specific cognitive test scores achieved high accuracy in certain subpopulations (82), emphasizing the value of detailed cognitive profiling. On the non-behavioral front, a biomarker-based approach achieved a notable AUROC of 0.933 (83), while another study integrating GFAP, Aβ42/Aβ40 ratio, and APOE-ε4 status reported a moderate AUROC of 0.732 (84). Additionally, a neuroimaging-based study utilizing MRI parameters demonstrated an accuracy of 81.1% (85), highlighting the efficacy of imaging biomarkers. Our study's high specificity in discerning amyloid-negative cases adds to the growing body of evidence supporting precise, non-invasive prediction models and illustrates the benefits of combining both behavioral and non-behavioral measures for Aβ status classification in AD research.

In the future, we intend to draw on the advancements in remote cognitive assessments, such as the promising findings from the Mobile Toolbox (MTB) project (86), to inform the development of a shorter, focused version of the NIH toolbox for amyloid detection. This will involve selecting tasks with demonstrated predictive validity for amyloid status, supported by psychometric evaluation and validation against established biomarkers. Our collaboration with established studies will support cross-validation and enhancement of this tool. Ultimately, our goal is to deliver a validated, accessible assessment tool for early detection and ongoing monitoring of AD-related cognitive changes.

Conclusion

The current machine learning model, which predicts amyloid status based on NIHTB measurements and demographic data, achieved an AUC of 0.74. Since our sample size is relatively small (N = 199), more data need to be collected to validate our findings and enrich the study population. Our findings underscore the significance of cognitive, motor, and sensory functions in relation to AD biomarkers and affirms that NIHTB is an appropriate tool for widespread screening for individuals susceptible to AD.

Supplement

Supplementary material is available for this article at https://doi.org/10.14283/jpad.2024.77 and is accessible for authorized users.

Table S1. Comparison of Model Performance Metrics Across Different Class Balancing Techniques

Authors' contributions: YC processed the data, ran all analyses, and wrote the manuscript. EH, SW, DR, RG contributed to the result interpretation and revised the manuscript. HD and SD supervised all work and revised the manuscript. All authors read and approved the final manuscript.

Funding: This work was supported by the ‘Assessing Reliable Measurement in Alzheimer's Disease and Cognitive Aging' project (ARMADA), T5U2CAG057441, sponsored by the NIA (MPIs Richard Gershon and Sandra Weintraub), NIA P30AG062421, and the MGH Neurology Transformative Scholar Award (PI Sudeshna Das).

Availability of data and materials: The datasets analyzed during the current study are available from the National Alzheimer's Coordinating Center (NACC) but are not publicly accessible. Access to the data is granted following a formal application and review process to ensure the privacy and confidentiality of research participants. Information on the application process can be found at NACC's data request page (https://naccdata.org/).

Ethics approval and consent to participate: The study was approved by the institutional review boards of all participating institutions including Northwestern University (NU), University of Michigan (UM), University of Wisconsin-Madison (UW-M), Mayo Clinic-Jacksonville, Florida (MCF), University of Pittsburgh (U Pitt), Emory University, University of California-San Diego (UCSD), Columbia University, and Massachusetts General Hospital (MGH). Written informed consent was obtained from all participants or their authorized representatives.

Consent for publication: Not applicable.

Competing interests: On behalf of all authors, the corresponding author states that there is no conflict of interest.