There has been interest by the healthcare science community in recent years in looking closely at how symptoms cluster together within and across clinical conditions as a way to understand the potential shared etiologies, possible preventive strategies, and treatment interventions that may be useful (Miaskowski et al., Reference Miaskowski, Aouizerat, Dodd and Cooper2007; Reference Miaskowski, Barsevick, Berger, Casagrande, Grady, Jacobsen, Kutner, Patrick, Zimmerman, Xiao, Matocha and Marden2017). This has been driven particularly strongly within oncology and nursing research, where so-called “symptom clustering” approaches have been used to understand the patient experience of their outcomes holistically when numerous co-occurring symptoms—such as pain, fatigue, insomnia, and depression—go undetected and unaddressed if the primary clinical focus is elsewhere, as in the case of cancer treatment (Dodd et al., Reference Dodd, Miaskowski and Paul2001; Ho et al., Reference Ho, Rohan, Parent, Tager and McKinley2015; Huang & Lin, Reference Huang and Lin2009; Illi et al., Reference Illi, Miaskowski, Cooper, Levine, Dunn, West, Dodd, Dhruva, Paul, Baggott, Cataldo, Langford, Schmidt and Aouizerat2012; Kim et al., Reference Kim, McGuire, Tulman and Barsevick2005). The clinical challenges posed by complex, comorbid, and chronic conditions are familiar to clinical and health psychologists and neuropsychologists (Ashworth et al., Reference Ashworth, Sletten, Junge, Simpson, Clarke, Cunnington and Rajaratnam2015; Ford, Reference Ford2018; Miles et al., Reference Miles, Silva, Lang, Hoffman, Venkatesan, Sevigny and Nakase-Richardson2021), although the concept of a “symptom cluster” as defined by these allied health research traditions (see Table 1) may be a novel framing (Barsevick, Reference Barsevick2016; Reference Barsevick2007; Dodd et al., Reference Dodd, Miaskowski and Paul2001; Harris et al., Reference Harris, Dodd, Kober, Dhruva, Hammer, Conley and Miaskowski2022; Miaskowski et al., Reference Miaskowski, Barsevick, Berger, Casagrande, Grady, Jacobsen, Kutner, Patrick, Zimmerman, Xiao, Matocha and Marden2017).

Table 1. Definitions of a symptom cluster in the literature

In symptom clusters research, exploratory factor analysis (EFA) has been a preferred approach for detecting clusters of symptoms that may not be expected based on traditional diagnostic groupings or severity delineations (Harris et al., Reference Harris, Dodd, Kober, Dhruva, Hammer, Conley and Miaskowski2022). Considered one of several “variable-centered approaches” of symptom clusters research (in contrast to other “patient-centered approaches,” such as latent profile analysis), EFA allows for statistical understanding of how an array of symptoms can be grouped to reveal the underlying structure of symptoms and how they may be related or caused by shared etiologies (Barsevick, Reference Barsevick2016; Harris et al., Reference Harris, Dodd, Kober, Dhruva, Hammer, Conley and Miaskowski2022; Oh et al., Reference Oh, Thor, Olsson, Skokic, Jörnsten, Alsadius, Pettersson, Steineck and Deasy2016; Skerman et al., Reference Skerman, Yates and Battistutta2012). In the psychometric and neuropsychological assessment traditions, EFA is often used for identifying “domains” for assessment, for psychometric validation of tests or batteries, or for further interpretation of assessments, such as developing composite scores (Ma et al., Reference Ma, Carlsson, Wahoske, Blazel, Chappell, Johnson, Asthana and Gleason2021; Strauss & Fritsch, Reference Strauss and Fritsch2004). While the language used for describing these approaches may differ between these research communities—trading “domains” for “symptom clusters”—the underlying clinical interest is a shared one: to use patient-centered assessments and advanced statistical techniques to better understand complex presentations of symptoms that may offer novel targets for assessment, intervention, and improvement in quality of life.

Thus, in the present study, a symptom clustering framework was chosen to study the experience of people with amnestic mild cognitive impairment (aMCI) or dementia of the Alzheimer type (DAT), as it offers an opportunity to understand patient experiences beyond the lines of traditional domains as typically approached in neuropsychological research. The objective of this research is to detect and examine the presence of possible symptom clusters within and across individuals with a research diagnosis of either aMCI or mild DAT (Weintraub et al., Reference Weintraub, Karpouzian‐Rogers, Peipert, Nowinski, Slotkin, Wortman, Ho, Rogalski, Carlsson, Giordani, Goldstein, Lucas, Manly, Rentz, Salmon, Snitz, Dodge, Riley, Eldes, Ustsinovich and Gershon2022). Given the known wide array of symptoms that can be experienced by patients with these diagnoses, which can span cognitive, emotional, motor, sensory, and social challenges (American Psychiatric Association, 2013; Lezak et al., Reference Lezak, Howieson, Bigler and Tranel2012; Thomas et al., Reference Thomas, Cook, Bondi, Unverzagt, Gross, Willis and Marsiske2020), using an exploratory symptom clustering approach offers the potential to detect groups of symptoms that may cut across domains. While most symptom clusters research has been performed in oncology populations (Harris et al., Reference Harris, Dodd, Kober, Dhruva, Hammer, Conley and Miaskowski2022), this theoretical and methodological approach has been applied in other clinical groups—such as HIV, heart disease, lung disease, and kidney disease—and symptom clusters have been found that correlate with important patient-centered outcomes including quality of life and health care utilization (Miaskowski et al., Reference Miaskowski, Barsevick, Berger, Casagrande, Grady, Jacobsen, Kutner, Patrick, Zimmerman, Xiao, Matocha and Marden2017). Furthermore, the National Institutes of Health (NIH) has recently supported several initiatives to examine other chronic disease groups for relevant symptom clusters (National Institutes of Health, 2017), with the goal to understand what can be learned about these clinical groups through application of these methods. The present study reports on one of these recent projects.

Historically, research on symptom clusters (or domains of functioning) in dementia has primarily focused on using neuropsychological assessments to distinguish similar clinical syndromes with unique neuropathological causes (such as frontotemporal dementia, semantic dementia, and DAT) by comparing neuropsychiatric symptoms that support the diagnosis of one syndrome over another (Kramer et al., Reference Kramer, Jurik, Sha, Rankin, Rosen, Johnson and Miller2003; Marra et al., Reference Marra, Quaranta, Zinno, Misciagna, Bizzarro, Masullo, Daniele and Gainotti2007). More recent research on symptom clusters relevant to MCI and dementia has focused primarily on understanding how patterns of cognitive functioning change with aging—for example, the distinct patterns of fluid cognition seen in adults with and without dementia (Ma et al., Reference Ma, Carlsson, Wahoske, Blazel, Chappell, Johnson, Asthana and Gleason2021; McDonough et al., Reference McDonough, Bischof, Kennedy, Rodrigue, Farrell and Park2016)—and lifestyle predictors of cognitive decline, such as the connections between engaging in cognitive leisure activities and lower rates of dementia (Lee & Chi, Reference Lee and Chi2016). Furthermore, research has shown the relevance of associated symptoms in noncognitive domains—such as motor, sensory, mood, social functioning—to the daily functioning and quality of life of these patients (Dyer et al., Reference Dyer, Lawlor and Kennelly2020; Elovainio et al., Reference Elovainio, Lahti, Pirinen, Pulkki-Råback, Malmberg, Lipsanen, Virtanen, Kivimäki and Hakulinen2022; Insha et al., Reference Insha, Arshad and Fazle2022; Hwang et al., Reference Hwang, Longstreth, Brenowitz, Thielke, Lopez, Francis, DeKosky and Fitzpatrick2020; Rostamzadeh et al., Reference Rostamzadeh, Kahlert, Kalthegener and Jessen2022; Vaingankar et al., Reference Vaingankar, Chong, Abdin, Picco, Jeyagurunathan, Seow, Ng, Prince and Subramaniam2017; van der Linde et al., Reference van der Linde, Stephan, Matthews, Brayne and Savva2010). There has nevertheless been limited research exploring the associations of symptoms across cognitive and noncognitive domains in MCI and dementia. One study, for example, found a connection between vestibular functioning and visuospatial functioning, with worse performance in one domain associated with worse performance in the other (Bigelow et al., Reference Bigelow, Semenov, Trevino, Ferrucci, Resnick, Simonsick, Xue and Agrawal2015). These findings are likely more than correlational; there is increasingly rigorous evidence to indicate that domains of noncognitive functioning (e.g., hearing; physical fitness) may be modifiable risk factors of cognitive decline and dementia (Cohen et al., Reference Cohen, Ryan and Lanzi2021; Livingston et al., Reference Livingston, Huntley, Sommerlad, Ames, Ballard, Banerjee, Brayne, Burns, Cohen-Mansfield, Cooper, Costafreda, Dias, Fox, Gitlin, Howard, Kales, Kivimäki, Larson, Ogunniyi, Orgeta, Ritchie, Rockwood, Sampson, Samus, Schneider, Selbæk, Teri and Mukadam2020). Thus, we believe that exploration of symptom clusters within and across domains in these populations may be fruitful for understanding targets for intervention, prevention, and as possible markers of etiology.

For the purpose of measuring symptoms within and across domains, there have been important test development efforts for standardizing systems of comprehensive symptom assessments, including the NIH Toolbox® (Gershon et al., Reference Gershon, Cella, Fox, Havlik, Hendrie and Wagster2010; Reference Gershon, Wagster, Hendrie, Fox, Cook and Nowinski2013). The NIH Toolbox (NIHTB), which includes tests that span a wide range of cognitive, sensory, motor, emotional, and social domains, appears to have significant potential for exploring symptom clusters. The NIHTB tests are self-contained and conormed, and score corrections can be made based on demographics and/or premorbid ability estimates (Holdnack et al., Reference Holdnack, Tulsky, Slotkin, Tyner, Gershon, Iverson and Heinemann2017; Nitsch et al., Reference Nitsch, Casaletto, Carlozzi, Tulsky, Heinemann and Heaton2017). This widely researched battery benefits from being a computerized assessment by having high precision for measuring reaction times, and automated administration and scoring that can reduce user bias. The NIHTB has been validated previously in healthy aging (Scott et al., Reference Scott, Sorrell and Benitez2019) and dementia research (Ma et al., Reference Ma, Carlsson, Wahoske, Blazel, Chappell, Johnson, Asthana and Gleason2021), including a recent publication supporting the validity of the NIHTB in individuals aged 85 and older (Nolin et al., Reference Nolin, Cowart, Merritt, McInerney, Bharadwaj, Franchetti, Raichlen, Jessup, Hishaw, Van Etten, Trouard, Geldmacher, Wadley, Porges, Woods, Cohen, Levin, Rundek, Alexander and Visscher2023), which further supports its utility for this study. For these reasons, we anticipate that the NIHTB will offer a useful assessment approach for symptom clusters research generally, and specifically in aging research, and we aim to evaluate it for this purpose with the current project. We hypothesize that by using an existing dataset, which was collected as part of the Advancing Reliable Measurement in Alzheimer’s Disease and Cognitive Aging (ARMADA) study (Weintraub et al., Reference Weintraub, Karpouzian‐Rogers, Peipert, Nowinski, Slotkin, Wortman, Ho, Rogalski, Carlsson, Giordani, Goldstein, Lucas, Manly, Rentz, Salmon, Snitz, Dodge, Riley, Eldes, Ustsinovich and Gershon2022), and a variable-oriented symptom clustering approach, we expect we will detect clusters of symptoms within and potentially across domains that may have relevance to understanding potential etiological factors, prevention targets, and intervention strategies for MCI and dementia.

Methods

Participants

Participants for this study were drawn from the ongoing data collection for the ARMADA study and included individuals recruited from nine established Alzheimer’s Disease Research Centers (ADRCs) across the United States. Although the ARMADA study includes several participant groups, only data from 165 individuals aged ≥ 60 with a research diagnosis of aMCI (single or multidomain) or mild DAT were included in the present study. Diagnoses of aMCI and mild DAT were based on the 2011 NIA–Alzheimer’s Association criteria (Albert et al., Reference Albert, DeKosky, Dickson, Dubois, Feldman, Fox, Gamst, Holtzman, Jagust, Petersen, Snyder, Carrillo, Thies and Phelps2011; McKhann et al., Reference McKhann, Knopman, Chertkow, Hyman, Jack, Kawas, Klunk, Koroshetz, Manly, Mayeux, Mohs, Morris, Rossor, Scheltens, Carrillo, Thies, Weintraub and Phelps2011) following the procedures of the National Alzheimer’s Coordinating Center (Morris et al., Reference Morris, Weintraub, Chui, Cummings, DeCarli, Ferris, Foster, Galasko, Graff-Radford, Peskind, Beekly, Ramos and Kukull2006; Weintraub et al., Reference Weintraub, Besser, Dodge, Teylan, Ferris, Goldstein, Giordani, Kramer, Loewenstein, Marson, Mungas, Salmon, Welsh-Bohmer, Zhou, Shirk, Atri, Kukull, Phelps and Morris2018; Reference Weintraub, Salmon, Mercaldo, Ferris, Graff-Radford, Chui, Cummings, DeCarli, Foster, Galasko, Peskind, Dietrich, Beekly, Kukull and Morris2009), with the Clinical Dementia Rating Scale (CDR; Morris, Reference Morris1993) global score of 0.5 required for aMCI and 1.0 for mild DAT. Participants were excluded for acute neurological disorders that could lead to cognitive impairment, a history of major psychiatric illness, or substance use disorder (see Weintraub et al., Reference Weintraub, Karpouzian‐Rogers, Peipert, Nowinski, Slotkin, Wortman, Ho, Rogalski, Carlsson, Giordani, Goldstein, Lucas, Manly, Rentz, Salmon, Snitz, Dodge, Riley, Eldes, Ustsinovich and Gershon2022, for full details on sample recruitment, inclusion/exclusion criteria, and diagnostic classification procedures). The study was approved by the institutional review boards at each of the participating data collection sites and was completed in accordance with Helsinki declaration.

Study design, measures, and procedures

Data used for these analyses come from the baseline assessment timepoint of the ARMADA project and included administration of the NIHTB (English language version) and a detailed inventory of demographics and health history (in many cases this was the Uniform Data Set [UDS] assessment). Table 2 presents details on the NIHTB assessments administered in the baseline assessment of the ARMADA study, including scoring metrics and directions. Analyses were conducted on all measures in NIHTB version 2.0 except the Standing Balance test, which was only completed by a minority of participants. Only cases in which the NIHTB and demographic/health history data were collected within a 130-day window were included in the present analyses.

Table 2. NIH Toolbox measures by battery and domain

Notes. Uncorrected SS = uncorrected standardized scores (M = 100, SD = 15) and T-scores (M = 50, SD = 10) are weighted to the 2010 Census; Better Ear Threshold for the Words-In-Noise Test is defined as the lowest threshold score observed for either ear, in the unit decibels of signal-to-noise ratio (dB S/N).

Data analyses

EFA was used to analyze the data. All EFA models were fit in R using the lavaan package (version 0.6-16; Rosseel, Reference Rosseel2012). Rates of missing data were generally low (< 5%), although a small number of assessments had slightly higher rates of missingness (e.g., Picture Sequence Memory = 18%). Full-information maximum likelihood (FIML) was used for model estimation and used all available information in the dataset. We examined EFA solutions that varied between 1 and 9 factors being extracted. Several criteria were used for factor extraction, as employing multiple criteria simultaneously may be optimal, compared to reliance on any single criterion alone (Auerswald & Moshagen, Reference Auerswald and Moshagen2019). First, we conducted a parallel analysis (Horn, Reference Horn1965). Second, we compared global model fit between adjacent solutions, as indicated by the chi-square test statistic χ ², the root-mean-square error of approximation (RMSEA; Steiger & Lind, Reference Steiger and Lind1980), and the Tucker-Lewis index (TLI; Tucker & Lewis, Reference Tucker and Lewis1973). Conventional fit index cutoffs—despite their known limitations (Marsh et al., Reference Marsh, Hau and Wen2004)—were used to pinpoint solutions that exhibited acceptable fit to the data: nonsignificant χ ², RMSEA values below .06, TLI values greater than .95 (Hu & Bentler, Reference Hu and Bentler1999). We also noted the solution in which the lower bound of the 90% confidence interval for the RMSEA dipped below .05 (Preacher et al., Reference Preacher, Zhang, Kim and Mels2013). Comparative model fit was indexed by the Bayesian information criterion (BIC; Schwarz, Reference Schwarz1978), with the lowest BIC value indicating an optimal balance between model fit and parsimony. Finally, the interpretability of each EFA solution was given the largest weight when deciding how many factors to extract. Following extraction, factor loadings were rotated using oblimin, a common oblique rotation method, as we expected the extracted factors to be correlated. Following extraction, factor scores were computed using empirical Bayes estimation and compared (t-test, Cohen’s d) between the aMCI and DAT subsamples.

Results

Demographic characteristics of the analysis sample are shown in Table 3. Participant age was approximately normally distributed between 60 and 94 years of age. Participant gender was roughly evenly distributed, with a slighter higher proportion of males in each diagnostic category. The sample was predominantly non-Hispanic and White. A majority of the sample reported high levels of educational attainment (4-year college degree or higher). Descriptive statistics for the primary outcome variables are shown in Table 4 (descriptive statistics within each diagnostic category are available in the Supplementary Material). Apart from the 9-Hole Pegboard Test and Odor Identification tests, all variables exhibited skewness and kurtosis indices between 2 and 2. Moderate floor effects were also observed for three measures: Picture Sequence Memory, Pain Interference, and Friendship.

Table 3. Demographic characteristics

Table 4. Descriptive statistics

Parallel analysis suggested the extraction of four factors (Fig. 1). Model fit results, shown in Table 5, were mixed. The chi-square test statistic was significantly different from zero for all solutions except the 7-, 8-, and 9-factor solutions. The RMSEA index fell below .06 in the 5-factor solution, and the lower bound of the 90% RMSEA confidence interval fell below .05 in this solution as well. The BIC was lowest for the 4-factor solution, although the TLI index did not show acceptable fit until the 7-factor solution. Given these conflicting answers, factor loading patterns for the 5-, 6-, 7-, 8-, and 9-factor solutions were inspected, and the 6-factor solution was considered the most interpretable solution. This solution exhibited acceptable fit for the RMSEA statistic and nearly met the conventional .95 threshold for the TLI.

Figure 1. Parallel analysis scree plot. In this figure, the triangles represent eigenvalues obtained across the 25 NIHTB measures. The dotted line represents average eigenvalues obtained from randomly generated datasets. Four of the observed eigenvalues were greater than the average of random samples, and thus the parallel analysis suggests retention of four factors, although a 6-factor structure was ultimately chosen as a more interpretable solution.

Table 5. Model fit

Note. BIC = Bayesian information criterion, CI = confidence interval, df = degrees of freedom, p = p-value, RMSEA = root-mean-square error of approximation, TLI = Tucker-Lewis index, χ ² = chi-square goodness of fit test statistic.

Rotated factor loadings for the 6-factor solution are shown in Table 6—a second oblique rotation method (geomin) was inspected and produced similar rotated loading values—and interfactor correlations are provided in Table 7. The factor loadings largely adhered to a simple structure pattern, with very few secondary loadings greater than .3. communalities, which represent the proportion of variance accounted for in each indicator variable by all extracted factors (symbolized as h ² ) were largest (> .5) for emotional, social, and cognitive performance outcomes, and smallest for motor and sensory outcomes—in particular, the tests of visual acuity (Visual Acuity Test; h ² = .14) and auditory function (Words-in-Noise; h ² = .17) had the lowest communalities.

Table 6. Factor loadings for 6−factor solution

Note. h ² = communality: collective proportion of variance in NIH Toolbox test accounted for by all extracted factors. u ² = uniqueness: proportion of variance in NIH Toolbox test not accounted for by extracted factors. Factor loading absolute values > .3 are highlighted in boldface.

Table 7. Factor correlations for 6-factor solution

Note. Interfactor correlations are shown in the lower diagonal. The proportion of variance accounted for in the entire variable set by each factor is shown on the diagonal.

Measures from the Cognition battery loaded onto two factors reflecting Fluid Intelligence (Factor 1) and Crystallized Intelligence (Factor 2). Along with the five cognitive tests of fluid intelligence, one motor and one sensory test (9-Hole Pegboard, Odor Identification) also exhibited weak/moderate loadings on Fluid Intelligence (Factor 1). The relationships between olfaction with cognitive decline in aging and with frontal-lobe structural integrity appear potentially relevant for understanding this factor (Bathini et al., Reference Bathini, Brai and Auber2019; Cynthia et al., Reference Felix, Chahine, Hengenius, Chen, Rosso, Zhu, Cao and Rosano2021; Roberts et al., Reference Roberts, Christianson, Kremers, Mielke, Machulda, Vassilaki, Alhurani, Geda, Knopman and Petersen2016; Sohrabi et al., Reference Sohrabi, Bates, Weinborn, Johnston, Bahramian, Taddei, Laws, Rodrigues, Morici, Howard, Martins, Mackay-Sim, Gandy and Martins2012; Yap et al., Reference Yap, Mahendran, Kua, Zhou and Wang2022). Regarding Crystallized Intelligence (Factor 2), the loading for Picture Vocabulary dominated the factor (rounded–1.00); a similar factor loading pattern was observed for this factor by Ma et al. (Reference Ma, Carlsson, Wahoske, Blazel, Chappell, Johnson, Asthana and Gleason2021) in individuals with MCI/dementia.

Measures from the Emotion battery loaded onto three factors that were interpreted as indicators of Negative Affect (Factor 3), Positive Affect/Life Satisfaction (Factor 4), and Social Health (Factor 5). Although these factors were quite homogeneous (i.e., indicated by variables with strong primary loadings and weak secondary loadings), both Pain Interference and Loneliness exhibited moderate loadings on Negative Affect (Factor 3). Loneliness also exhibited a moderate loading on Social Health (Factor 5). The remaining sixth factor included several Physical Function outcomes across motor (gait speed/ambulation endurance, manual motor strength) and sensory (visual acuity) domains (Factor 6), although the factor loading pattern suggests the factor is strongly indicated by lower extremity function (ambulation endurance). The auditory perception outcome measure, Words-in-Noise Test, did not load clearly on any factor, although the highest loading (absolute value = .25) observed for this test was for the Physical Function factor (Factor 6).

Interfactor correlations ranged between .51 and .41 (see Table 7); the strongest correlations were observed among the three factors indicated by the Emotion battery measures (Factors 1, 4, and 5). The proportion of variance accounted for in the NIHTB tests by each factor is shown in the diagonal in Table 7. These proportions ranged between .07 and .15, with the highest value of .15 observed for Negative Affect and the lowest value of .07 observed for Crystallized Intelligence and Positive Affect/Life Satisfaction. Overall, the factor loading patterns and communality estimates suggest the outcomes from the Emotion and Cognition batteries form rather tight-knit clusters, whereas the Motor and Sensory outcome measures form less cohesive clusters that may be dominated by a single measure. Nevertheless, we believe that the cross-domain factor loadings may be useful for understanding patterns of symptoms that cluster together in meaningful ways in these populations. Finally, factor score comparisons between the aMCI and DAT subsamples are shown in Table 8. The DAT group exhibited significantly worse functioning compared to the aMCI group on three factors: Fluid Intelligence (Factor 1), Crystallized Intelligence (Factor 2), and Physical Function (Factor 6); effect sizes (absolute values) ranged from .01 to 1.14.

Table 8. Mean comparisons

Note. Negative values for d indicate lower average factor scores for the DAT group compared to the aMCI group and vice versa.