Introduction
Alzheimer’s disease is a progressive neurodegenerative disorder that cumulates in mortality on average 4–8 years after the diagnosis, characterized by impairments in the activities of daily functioning and cognitive decline [1]. Since cognitive impairment is a clinical hallmark of Alzheimer’s disease [1] suitable assessments are essential for treatment and research following onset [Reference Robert, Ferris, Gauthier, Ihl, Winblad and Tennigkeit2]. The most widely used and researched cognitive impairment outcome in clinical trials of Alzheimer’s disease is the Alzheimer’s disease Assessment Scale Cognitive Subscale (ADAS-Cog) [Reference Rosen, Mohs and Davis3]. The ADAS-Cog is one of the two primary cognitive outcome measures required by the Food and Drug Administration for clinical drug trials for the treatment of Alzheimer’s disease in the United States [Reference Manning, Ducharme and Lichtenberg4]; however, it is quite long to administer (takes on average 30–35 min to complete).
Early evidence based on traditional psychometric approaches reported that the ADAS-Cog demonstrates acceptable levels of reliability and validity [1, Reference Robert, Ferris, Gauthier, Ihl, Winblad and Tennigkeit2]. Validity was supported based on evidence showing that the different aspects of cognition that constitute the ADAS-Cog are adequately correlated to form a single factor [Reference Rosen, Mohs and Davis3]. However, subsequent research did not replicate the single-factor solution and instead identified two- and three-factor solutions [Reference Manning, Ducharme and Lichtenberg4, Reference Weyer, Erzigkeit, Kanowski, Ihl and Hadler5] and queried the level of reliability of the ADAS-Cog [Reference Cano, Posner, Moline, Hurt, Swartz and Hsu6]. Furthermore, some studies suggest that the ADAS-Cog is appropriate for use only in the moderate stages of cognitive impairment. Namely, the ADAS-Cog demonstrates severe floor (i.e., some items are too easy for patients) and ceiling (i.e., some items are too difficult for patients) effects [Reference Rosen, Mohs and Davis3, Reference Cogo-Moreira, Krance, Black, Herrmann, Lanctôt and MacIntosh7, Reference Grochowalski, Liu and Siedlecki8]. Hence, contentions exist that the ADAS-Cog is inappropriate for mild and severe stage dementia [Reference Rosen, Mohs and Davis3, Reference Cogo-Moreira, Krance, Black, Herrmann, Lanctôt and MacIntosh7, Reference Grochowalski, Liu and Siedlecki8]. In addition, the traditional psychometric approaches to examining the ADAS-Cog cannot examine treatment effects [Reference Rosen, Mohs and Davis3, Reference Levine, Rabinowitz and Rizopoulos9, Reference Wilson, Niu, Nicolson, Levine and Heckers10]. Hence, given these inconsistent findings, examination of the ADAS-Cog using advanced psychometric approaches is warranted.
To improve the ADAS-Cog, advanced psychometric approaches, such as item response theory (IRT), may be helpful [Reference Cano, Posner, Moline, Hurt, Swartz and Hsu6]. Unlike traditional psychometric approaches, like factor analysis, IRT offers ADAS-Cog details at different cognitive impairment levels by item, information (i.e., reliability), and response option. It does so graphically and numerically. Estimates are available to map the ability of an item to discriminate underlying cognitive impairment levels. Also, it is possible to estimate the probability of progressing to a higher cognitive impairment response option rating or not. It is possible to identify which response options are likely, unlikely, and superfluous [Reference Levine and Leucht11]. This feature of IRT is related to identifying items and response options that display ceiling or floor aspects on the ADAS-Cog. This seems of note to clinical trials where a given item may be used as a selection criterion, thereby impacting the response option ratings on the remaining items.
IRT has been implemented in studies to shorten psychiatric [Reference Levine, Rabinowitz and Rizopoulos9–Reference Levine and Leucht11] and cognitive measures in dementia [Reference McGrory, Doherty, Austin, Starr and Shenkin12]. Studies that use IRT to examine the ADAS-Cog highlight that the measure is optimal within the moderate range of cognitive impairment only [Reference Benge, Balsis, Geraci, Massman and Doody13]. However, research has yet to identify an ADAS-Cog IRT-based shortened version that separates treatment and placebo to detect treatment effects.
We aimed to develop an ADAS-Cog short form (ADAS-Cog) using IRT based on individual-level participant clinical trial data and to examine whether it could separate treatment and placebo groups.
Methods
Participants
Study design
Data were accessed on pivotal individual-level participant data of randomized controlled double-blinded trials of donepezil conducted by Eisai Co. Ltd (see Table S1 published as supplementary material online attached to the electronic version of this paper at https://www.cambridge.org/core/journals/european-psychiatry). Data access was granted after the submission of an analytic plan. The data were analyzed on a secure Internet cloud-based platform (http://www.clinicalstudydatarequest.com). Trials were included in which participants with Alzheimer’s disease were assessed with the ADAS-Cog. Individual-level participant data were ascertained from five randomized clinical trials with similar follow-up intervals [Reference Homma, Takeda, Imai, Udaka, Hasegawa and Kameyama14–Reference Burns, Rossor, Hecker, Gauthier, Petit and Moller18]. Institutional review boards approved each trial.
Measures
ADAS-Cog: The ADAS-Cog is a neuropsychological index of cognitive impairment, indicating the severity of cognitive symptoms in Alzheimer’s disease [Reference Mohs and Cohen19]. This measure has been widely used in Alzheimer’s disease clinical trials [Reference Rosen, Mohs and Davis3] and has become as the gold standard for evaluating treatment efficacy [Reference Kueper, Speechley and Montero-Odasso20]. It consists of 11 items to assess memory, language, and praxis functions [Reference Mohs and Cohen19]. The ADAS-Cog total score ranges from 0 to 70, with high scores indicating more severe cognitive impairment.
Analytic plan
First, following the removal of individuals with missing baseline ADAS-Cog item level scores (Table 1), the analytic sample was characterized. Second, items and rating options were removed based on IRT to identify an ADAS-Cog short version. Third, the ADAS-Cog original and short versions were examined with mixed-effects models for repeated-measures analysis (MMRM).
Table 1. Sample characteristics
IRT of the ADAS-Cog at baseline
IRT assumes a single component underlies the data. Hence, principal components analysis was implemented to ascertain the number of components underlying the data. Next, the graded response model (GRM) [Reference Samejima21], a form of IRT, was implemented in the ltm package in R [Reference Rizopoulos22]. The GRM model has been used to shorten measures previously [Reference Levine, Rabinowitz and Rizopoulos9–Reference Levine and Leucht11, Reference Velthorst, Levine, Henquet, de Haan, van Os and Myin-Germeys23]. In IRT, item discrimination parameters (α) map the ability of an item to discriminate impairment levels. Discrimination parameter values for items are considered very low (between 0.01 and 0.24), low (0.25 and 0.64), moderate (0.65 and 1.34), high (1.35 and 1.69), and very high (over 1.7) [Reference Baker24]. Threshold parameters (βs) indicate the point at which there is a probability of endorsing a higher cognitive impairment rating than the previous rating option. If a threshold value exceeds 1.96, it suggests that ratings provide accurate information, and the converse applies to negative values.
Three graphs are used in IRT: item response category characteristic curves (a plot of the probability of endorsing a rating option by the level of underlying cognitive impairment), Item information curves (lines at similar information levels indicate overlapping, namely that the items assess similar information and so there exists a degree of item redundancy). Test information shows the reliability of the cognitive functioning assessment at different impairment levels.
Mixed models to assess treatment effects
We examined change scores, marginal means, and effect sizes differences in the marginal mean with their associated bootstrapped confidence intervals between the donepezil and placebo groups using a three-level MMRM analysis with maximum likelihood estimation. The levels accounted for the data structure such that level 1 represented the visit, level 2 represented the individual, and level 3 represented the trial [Reference Hedeker and Gibbons25]. The covariates were age, sex, baseline ADAS-Cog score, and treatment group, and the outcome was the change score from baseline.
Results
Trial characteristics
After removing 12 participants owing to missing ADAS-Cog item responses, the five trials comprised 2,198 study participants. These formed the basis for the baseline IRT analysis (see Supplementary Table S1).
IRT analysis: Tasks discriminating cognitive impairment levels
A scree plot showed that the data sufficed the unidimensional assumption that IRT requires (see Figure S1 published as supplementary material online attached to the electronic version of this paper at https://www.cambridge.org/core/journals/european-psychiatry). Item discrimination parameters were computed to map the ability of an item to discriminate latent symptom severity levels (see Table 2 alpha values). For example, word recall had the highest ability to discriminate underlying cognitive impairment levels (α=1.92). Four ADAS-Cog tasks (spoken language ability, comprehension of spoken language, remembering test instruction, and word finding difficulty) had low item discrimination parameters (i.e., these tasks lacked the ability to discriminate underlying cognitive impairment levels). Hence, the aforementioned four tasks were considered inappropriate for the IRT-based short-scale, leaving seven possible ADAS-Cog tasks (word recall, commands, naming, constructional praxis, ideational praxis, orientation, word recognition).
Table 2. Item parameters from IRT
Note: Item discrimination parameters (α) map the ability of an item to discriminate latent cognitive impairment levels. Discrimination parameter values (α) that range from 0.01 to 0.24 are very low, 0.25 to 0.64 low, 0.65 to 1.34 moderate, 1.35 to 1.69 high, and over 1.7 are very high (Baker, 2001). βs are standardized estimates of the 0.5 probability of endorsing a higher cognitive impairment rating where negative values indicate progression to the next response is unlikely.
IRT analysis: ADAS-cog information ascertained at different cognitive impairment levels
Task information (reliability) is ascertained by IRT for the total scale and each task. The topmost plot in Figure 1 shows the test information along the vertical axis at different cognitive impairment levels along the horizontal axis for the ADAS-Cog total. Figure 1 (top panel) suggests that the ADAS-Cog is more reliable at moderate and moderately high impairment levels but displays a reliability that is not satisfactory at low and very high cognitive impairment levels. Figure 1 (middle panel) shows that the information ascertained by word recall is moderate across impairment levels up to severe levels of impairment from which the information ascertained is low.
Figure 1. Item response figures. Note: The horizontal axis denotes the underlying latent trait of cognitive impairment.
Of the remaining seven possible ADAS-Cog tasks, the amount of information captured ranged from low to moderate. Word recall captured information at moderate cognitive impairment levels, commands from moderate to high levels, naming at moderate levels, constructional praxis from low to high levels, ideational praxis from moderate to high levels, orientation from moderate to high levels, and word recognition from very low to high levels (for information plots for all tasks, see Figures S2 and S3 published as supplementary material online attached to the electronic version of this paper at https://www.cambridge.org/core/journals/european-psychiatry).
IRT analysis: Response options
Based on item characteristic curves and the probability of a response option being endorsed (Table 2 beta values), we aimed to remove overlapping response options. For instance, the bottom panel of Figure 1 shows that response option 10 is endorsed with a high likelihood at higher impairment levels. All seven possible ADAS-Cog tasks had at least one response option that would likely be required (see Figure S5 published as supplementary material online attached to the electronic version of this paper at https://www.cambridge.org/core/journals/european-psychiatry and Table 2 beta values). However, not all response options appeared to be necessary.
We examined Table 2 (and see Figure S5 published as supplementary material online attached to the electronic version of this paper at https://www.cambridge.org/core/journals/european-psychiatry) to identify and remove superfluous response options. We identified superfluous sources of information for each of the items: word recall (9–10 errors captured severe impairment, and the remaining response options appeared not to capture severe impairment); commands (up to 3 commands incorrect did not appear to have differential utility in capturing impairment, and subsequent commands incorrect slightly superfluous); naming (the options did not capture severe cognitive impairment except five: “9–11 items incorrect”); constructional praxis and ideational praxis (options 0–3 were unlikely to result in a subsequent rating, and 4 and 5 overlapped to moderate to severe capture impairment); orientation (response options 6–8 reflected more severe impairment); and word recognition (12 incorrect responses represented severe impairment, otherwise transition was unlikely and the item responses were quite superfluous).
The ADAS-Cog IRT short-scale scoring key
Based on the above, we recoded the IRT-based ADAS-Cog short version as follows: word recall (0 except 9–10 recoded as 1); commands (up to 3 as 0, otherwise 1); naming (0 except five as 1); constructional praxis and ideational praxis (options 0–3 as 0, and 4 and 5 as 1); orientation (0–5 as 0, 6–8 as 1); and word recognition (0 except 12 as 1). For consistency and ease of future use, dichotomous scoring was implemented.
Mixed models
The bivariate correlation at baseline, at week 12, and week 24 of the short and original ADAS-Cog measures was 0.7 across time points. MRMMs were implemented to contrast the original and IRT-based short ADAS-Cog (Figure 2). The marginal means differed between the original and short ADAS-Cog (original version: donepezil = −1.85, 95% CI = −2.16, −1.53, placebo = −0.38, 95% CI = −0.77, −0.00; short version: donepezil = −0.04, 95% CI = −0.10, −0.02, placebo = 0.11, 95% CI = 0.05, 0.18) were smaller for donepezil than placebo. Based on the marginal means, examination of the effect sizes showed that placebo and donepezil separated more for the original than the short ADAS-Cog version, but the bootstrapped confidence intervals overlapped between versions (ADAS-Cog original ES = 0.33, 95% CI = 0.29, 0.40, ADAS-Cog short ES = 0.25, 95% CI = 0.23, 0.34).
Figure 2. Mixed model modeling changes in the original and short Alzheimer’s Disease Assessment Scale Cognitive Subscale (ADAS-Cog) up to 24 weeks. Note: Upper figure is the original ADAS-Cog and the lower is the short ADAS-Cog based item response theory.
Discussion
Based on five pivotal clinical trials of donepezil compared with placebo for Alzheimer’s disease (N = 2,198), we implemented IRT to shorten the ADAS-Cog and examined whether this short version could separate treatment and placebo groups in a manner similar to the original version. We identified a short ADAS-Cog that consisted of seven items and found that it separated placebo from donepezil in these trials.
IRT identified a short ADAS-Cog consisting of 7 items with dichotomous response options, in contrast to the original, which consists of 11 items with multiple response options. In our estimation, assuming the ADAS-Cog takes 30 min to administer, the test-time for the short version may be approximately 18 min or less, because the short version has seven items (36.37% fewer items than the original ADAS-Cog) and two response options (to ease future administration).
Based on mixed modeling, scores on the ADAS-Cog change short version were separated between placebo from donepezil in these individual participant trial data. Also, mixed modeling to examine ADAS-Cog change showed conclusions concerning efficacy were similar for both the short and original ADAS-Cog scales (i.e., both showed superior efficacy of donepezil compared to placebo). The effect size, however, slightly favored the original compared to the short scale.
Limitations and conclusions
Our study has several primary strengths, such as the use of individual-level participant data. Nonetheless, our study has notable limitations. First, clinical trial selection criteria restrict generalizations from clinical trial data to the general population [Reference Malmivaara26, Reference Canevelli, Bruno, Vanacore, de Lena and Cesari27]. Hence, caution is warranted regarding generalizing from the current results to clinical treatment settings. To inform clinical practice, replicating the results in large-scale naturalistic studies with extended observation periods may be warranted. Second, unmeasured factors (e.g., delusions) may have confounded the study results. Nonetheless, the data common to all the trials did not contain such other information. Hence, our study may suffer from residual confounding, and future research may wish to account for other potential confounders. Third, our results are restricted to donepezil and placebo. Research is warranted to scrutinize the generalizability of these results to other antidementia drugs. Fourth, the study duration was restricted to 24 weeks of follow-up. Given the course of cognitive decline in Alzheimer’s disease, further research is warranted with longer study durations. Fifth, an independent prospective study is warranted to test the validity of the scale.
The clinical trials in our study were completed over a decade ago. Today, a significant proportion of participants would not receive a research diagnosis of Alzheimer’s disease. Specifically, perhaps up to 30% would receive diagnoses for other neurodegenerative disorders, including vascular or mixed dementia, based on current-day research diagnostic criteria that involve biomarkers, such as amyloid PET, to confirm neuropathology in Alzheimer’s disease according to the 2018 NIA-AA Research Framework [Reference Jack, Bennett, Blennow, Carrillo, Dunn and Haeberlein28]. However, the use of biomarkers is yet to translate to daily clinical practice [Reference Frisoni, Boccardi, Barkhof, Blennow, Cappa and Chiotis29]. In current daily clinical practice, the symptomatological diagnostic criteria, including DSM-5 [30] and NINCDS-ADRDA [Reference McKhann, Drachman, Folstein, Katzman, Price and Stadlan31], are the basis for the prescription of donepezil and other antidementia drugs, as were done in the trials included in the current study.
Among the strengths of the current study design are the amount of evidence (five pivotal clinical trials) and the relatively large sample, which make the results robust. These features reinforce our faith in the robustness of the analysis. Clinically, a short ADAS-Cog with a strong correlation with the original offers possibilities in reducing the trial participant burden while keeping reliability intact. In sum, the current study contributes to knowledge on Alzheimer’s disease by identifying a short version of the ADAS-Cog with potential use for treatment monitoring in moderate-stage Alzheimer’s disease.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1192/j.eurpsy.2024.14.
Acknowledgments
Authors Levine and Goldberg contributed equally to this study and are joint first authors. The authors acknowledge Eisai Co. Ltd for providing us with the study data. Eisai Co. Ltd did not provide study design, critical input, or manuscript review for the study. The authors also acknowledge http://www.clinicalstudydatarequest.com for hosting the study data. Data are available based on a request to http://www.clinicalstudydatarequest.com.
Author contribution
Levine: Manuscript drafting, data curation, statistical analysis, data management, study conceptualization.
Goldberg: Study conceptualization, critical manuscript feedback, statistical analysis.
Rotstein: Study conceptualization, interpretation, critical manuscript feedback.
Yoshida: Critical manuscript feedback, data management, statistical analysis.
Samara: Study conceptualization, interpretation, critical manuscript feedback.
Cipriani: Study conceptualization, interpretation, critical manuscript feedback.
Iwatsubo: Study conceptualization, interpretation, critical manuscript feedback.
Leucht: Study conceptualization, interpretation, critical manuscript feedback.
Furukawa: Critical manuscript feedback, statistical review, study conceptualization, mentorship.
Financial support
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Cipriani is supported by the National Institute for Health Research (NIHR) Oxford Cognitive Health Clinical Research Facility, by an NIHR Research Professorship (grant RP-2017-08-ST2-006), by the NIHR Oxford and Thames Valley Applied Research Collaboration and by the NIHR Oxford Health Biomedical Research Centre (grant BRC-1215-20005). The views expressed are those of the authors and not necessarily those of the UK National Health Service, the NIHR, or the UK Department of Health.
Competing interest
Drs Levine, Yoshida, Rotstein, and Goldberg have nothing to disclose. Dr. Samara has received honoraria as a consultant or for lectures for Viatris, Recordati, Lundbeck, and Viatris. Dr. Iwatsubo has served as a consultant of Eisai and Eli Lilly in the last 3 years. Dr. Cipriani has received research and consultancy fees from INCiPiT (Italian Network for Pediatric Trials), CARIPLO Foundation and Angelini Pharma. In the last 3 years, SL has received honoraria for advising/consulting and/or for lectures and/or for educational material from Angelini, Boehringer Ingelheim, Eisai, Ekademia, GedeonRichter, Janssen, Karuna, Kynexis, Lundbeck, Medichem, Medscape, Mitsubishi, Neurotorium, Otsuka, NovoNordisk, Recordati, Rovi, and Teva. Dr. Furukawa reports royalties from Mitsubishi-Tanabe, consulting fees from Boehringer-Ingelheim, DT Axis, Kyoto University Original, Shionogi, SONY, UPTODATE, and Daiichi Sankyo, and a grant from Shionogi, outside the submitted work. In addition, Dr. Furukawa has patents 2020-548587 and 2022-082495 pending, and intellectual properties for Kokoro-app licensed to Mitsubishi-Tanabe.
Open access
Comments
No Comments have been published for this article.