Design
SCD-Well was a European multicenter, observer-blind RCT with two intervention arms: an MBI named the Caring Mindfulness-based Approach for Seniors (CMBAS) and an HSMP. The study was conducted across four sites (London, Cologne, Lyon, and Barcelona). The trial was registered on ClinicalTrials.gov (NCT03005652). SCD-Well was sponsored by the French National Institute of Health and Medical Research (INSERM), and ethical approval and regulatory authorizations were obtained at each site. Written informed consent was obtained from all participants (please see the “Declarations” section for further details). Further details pertaining to the study’s eligibility criteria, interventions, and assessments are available in the trial protocol [21], as well as the primary outcome report, which focuses on trait anxiety [22].
Procedure
Due to the group-based nature of the interventions, participants were recruited in two waves at each site. Briefly, participants fulfilling the eligibility criteria were invited to the baseline visit (week 0) for cognitive and behavioral assessments. They were then randomized with a 1:1 allocation, using permuted block sizes of 4 and 6, stratified by site and centralized via a secure electronic case report form. Participants were invited to meet their intervention facilitator at a pre-class meeting, during which their trial allocation was revealed. The assessments were repeated at both post-intervention (week 8) and 6-month (week 24) follow-up visits. The size of each intervention group ranged from 7 to 13 participants.
Participants
Recruitment took place from March 2017 through January 2018. For study inclusion, participants were required to fulfill the research criteria for SCD [1]. Briefly, these require an individual to self-report a decline in cognitive function but to score normally on standardized cognitive tests used to screen for MCI and/or dementia. The SCD criteria exclude neurodegenerative diseases (except Alzheimer’s disease), psychiatric disorders, and clinically significant affective symptoms. However, subclinical affective symptoms are not exclusionary. All participants were recruited from memory clinics, and the minimum age for study eligibility was 60 years; these characteristics are associated with an increased risk of incident dementia in SCD [4].
Interventions
Caring Mindfulness-based Approach for Seniors (CMBAS)
The CMBAS followed the general format of a mindfulness-based stress reduction program, consisting of a pre-class interview, eight weekly group-based sessions of 2 h, and a half-day of meditation practice in the sixth week of the program to help consolidate learning. In addition to standard MBI practices [11], CMBAS participants were also taught compassion meditation practices focusing on cultivating wholesome attitudes toward oneself and others. Additional modifications included the provision of psychoeducation designed to help participants with SCD deal more adaptively with cognitive concerns and a tendency to worry, building on earlier work by Zellner Keller et al. [23]. Participants were asked to engage in home practice for approximately 1 h per day on 6 days per week and to record whether they engaged in these practices in a diary. Home practice consisted of formal practices (e.g., following guided meditation audio recordings), as well as informal practices designed to help participants apply mindfulness skills to their daily lives (e.g., mindful eating—bringing awareness to the taste, smell, and texture of a meal).
Health Self-Management Program (HSMP)
The HSMP followed the same format and structure as CMBAS and was matched in administration, dosage, and duration (including a half-day review with a healthy lunch and a discussion in the sixth week of the program). The intervention was based on a manual for living with chronic health conditions [24]; the manual was available in English, French, Spanish, and German. A previous RCT of an MBI which included older adults with neurocognitive difficulties adapted the manual to be delivered as a group psychoeducation intervention [16]; the adapted program was used to equalize treatment expectancy between arms and control for the “non-specific” components of the MBI (e.g., social interaction, input from a professional facilitator and light physical activity). In the current trial, the topics taught in the HSMP included self-management, problem-solving, sleep, stress, exercise, managing medicines, communicating with family and healthcare professionals, eating, weight management, and planning for the future. To promote engagement, participants were asked to plan, undertake, and report back on weekly “action plans.” Implementation of “action plans” was recorded by participants in a diary.
Intervention facilitators and psychometrists
Each site had two clinically trained facilitators experienced in leading group-based programs, one for each intervention. Facilitators received their respective intervention manual, instructions, and intervention-specific training prior to the start of the study. After each class, facilitators completed a self-report checklist [25] to indicate the extent to which they adhered to the session as outlined in the manual. They also received ongoing supervision to promote standardization of delivery across sites. All psychometrists were blind to participants’ allocation and completed the study-specific training in order to standardize the administration and scoring of outcome measures.
Composite cognitive outcomes
We calculated three composite measures of cognition from the broad battery of tests that were administered (see Additional file 1: Supplementary Methods for details). Schneider and Goldberg [26] summarized the potential advantages of composite over individual cognitive measures, including greater sensitivity to detect cognitive changes and avoidance of ceiling and floor effects, improved test-retest reliability, and reduced statistical multiplicity. Furthermore, the wider breadth of composite (versus individual) cognitive measures reduces the chance that any performance gains simply reflect similarities between the intervention activities and outcome measures (primarily a concern for cognitive training interventions). Schneider and Goldberg noted that scores across various cognitive domains are correlated, and this justifies the creation of “global” composites; nevertheless, they also emphasized that the measurement of individual cognitive domains remains crucial [26]. We thus specified both a global and two domain-specific composites. The same statistical approach was used to create each composite (described in detail below for the global composite). Composite scores were only calculable for time points where participants had data available for all of the necessary constituent tests (for details of how missingness was handled, see the “Statistical analyses” section). For each of the three composites, higher scores reflect better performance. Following the calculation of the composites (see below), each had a mean of 0 but a standard deviation (SD) less than 1; composites were thus “re-standardized” prior to analyses.
Abridged Preclinical Alzheimer Cognitive Composite 5
Donohue and colleagues [27] devised a global composite comprising four cognitive tests (two episodic memory, one attention, and one dementia screening measure); the authors demonstrated that this measure was sensitive to amyloid-beta (Aβ)-related cognitive decline in four cohorts over a 36-month period. The composite was named the Preclinical Alzheimer’s Cognitive Composite (PACC) [27]. Subsequently, Papp and colleagues [28] demonstrated that the sensitivity of the PACC could be increased through the addition of a category fluency score; the revised five-item measure was designated the PACC5. We produced an “abridged,” four-item version of the PACC5 (PACC5Abridged) in SCD-Well, as only one episodic memory measure was available. The tests constituting the PACC5Abridged were the Rey Auditory Verbal Learning Test (delayed recall), the WAIS-IV Coding subtest (raw score), category fluency for animals (total correct), and the Mattis Dementia Rating Scale-2 (total score). The primary cognitive functions assessed by these measures are episodic memory, attention, semantic fluency, and global neuropsychological status, respectively. To create the global composite, each constituent score was first standardized, by subtracting the baseline pooled sample mean from each individual’s score at each available time point, and the result was divided by the baseline pooled standard deviation. We then took the average of these four scores, yielding the PACC5Abridged.
Attention cognitive composite
We also calculated an attention cognitive composite (“attention composite”). To calculate this measure, we first standardized scores from the Trail-Making Test Part A (TMT-A; completion time in seconds), a “naming” condition from the Stroop requiring participants to name the color of rectangular stimuli arranged in a grid (completion time in seconds), and WAIS-IV Coding (raw score). TMT-A and Stroop scores were multiplied by minus one, so that higher scores reflected better performance. We took the average of these three standardized scores, yielding the attention composite.
Executive function cognitive composite
Lastly, we calculated an executive function cognitive composite (“executive composite”). To calculate this measure, we first standardized scores from the TMT-B (completion time in seconds), letter fluency for “P” (total correct), and a Stroop “interference” score (time in seconds). The Stroop interference score was calculated by subtracting the completion time of the Stroop naming condition (see the previous paragraph) from the completion time of a Stroop “incongruent” condition requiring participants to name the ink color of color words, where the ink color was incongruent with the word itself. TMT-B and Stroop interference scores were multiplied by minus one, so that higher scores reflected better performance. We took the average of these three standardized scores, yielding the executive composite.
Additional measures
Depressive symptoms were assessed using the 15-item Geriatric Depression Scale (GDS-15; range 0–15); higher scores reflect greater depressive symptoms [29]. Anxiety was measured using the State-Trait Anxiety Inventory-State subscale (STAI-A; range 20–80); higher scores reflect greater anxiety [30]. After the first intervention session, each participant also completed the Credibility/Expectancy Questionnaire (CEQ), which measures participants’ perception of their assigned intervention’s credibility, as well as their associated degree of expectancy [31]. Responses were used to compare participants’ expectations and perceptions of interventional credibility between arms. At the final visit (V3) participants were asked whether they had continued practice during the preceding four weeks of the follow-up period.
Statistical analyses
Sample size calculations were based on the expected effect size with 80% power and a two-sided type 1 error of 5% for the mean change in the SCD-Well primary outcome (i.e., STAI-Trait subscale) from pre- to post-intervention between intervention arms. This resulted in a minimum total number of 128 (64 per group) [21], which the trial exceeded (n = 147). For the present analyses, we calculated the statistical power for the PACC5Abridged only, as this was considered the main outcome. For an effect size of 0.25 on the PACC5Abridged, the power achieved by the study was 33%; for an effect size of 0.50, the power was 87% (for the rationale for selecting these effect sizes and further details of the approach used to calculate power, please see Additional file 1: Supplementary Methods). Descriptive statistics were calculated for the sample’s demographics and baseline measures. Given participants were randomized to interventions, we did not test for demographic or baseline differences between arms [32]. Here, we focus on the results for the three cognitive composites; data and models for individual cognitive tests are provided in Additional file 1: Tables S2-S3 and Figs. S2-S3. Linear mixed models (LMMs) were used to assess the effect of intervention assignment on outcomes over time. For each LMM, all participants who had at least one score for the respective outcome were included. All models included fixed effects for age at baseline (years), education level (years), baseline STAI-A score, baseline GDS-15 score, sex, and study site, as well as random participant intercepts. The parameters time (in weeks), trial arm, and the interaction between time and trial arm were also included to ascertain whether (a) outcome scores changed during the trial and (b) any observed change differed by arm. The use of a continuous-time metric (i.e., weeks) linearly constrained the modeled cognitive trajectories. Given other trajectories were plausible (e.g., improvement during the intervention period, but not during follow-up), we also analyzed the outcomes using LMMs with a factorial time metric (i.e., using the visit structure: baseline, post-intervention, and follow-up).
Analyses were conducted according to both per-protocol (PP) and intention-to-treat (ITT). In the “Results” section we report PP analyses and note where these differ from ITT. The PP analyses included all available (i.e., non-missing) cognitive test data; the main reasons for missingness were participants not attending the post-intervention and/or follow-up visits due to dropping out or being lost to follow-up (see Fig. 1 for the CONSORT flow diagram). In addition, a sensitivity analysis re-estimated all PP LMMs using only the subset of participants who attended ≥ 4 intervention sessions; these analyses were motivated by previous research adjudging four MBI sessions to be an adequate minimal dose [33]. A series of PP linear regression analyses were conducted to determine the strength of association between participant baseline characteristics (i.e., predictors) and change on each composite in each trial arm separately. The outcome (i.e., dependent variable) for analyses was the follow-up (week 24) minus the baseline (week 0) score. The candidate predictors included in separate regression models were age, sex, education, site, GDS-15, STAI-A, CEQ-credibility, CEQ-expectancy, and the baseline score on the respective composite. All models controlled for age, sex, education, and site (either through the inclusion of these as the predictor of interest or as covariates).
For the ITT analyses, missing outcome data (for participants who dropped out or were lost to follow-up) were multiply-imputed using chained equations (the missing data pattern is presented in Additional file 1: Fig. S1). Given participants were randomized after their first cognitive assessment, virtually all baseline data were available for inclusion in the imputation models. Five datasets were “completed” using multiple imputation, and the LMM for each outcome was estimated using each of these five datasets. Finally, the five iterations of each LMM were pooled to yield a single ITT model for each outcome (for full details, see Additional file 1: Supplementary Methods).
Analyses were conducted in R v.4.0.2 under RStudio v.1.3.1073. LMMs were fit using the package lme4 v.1.1-27.1; p-values for LMMs were obtained via lmerTest v.3.1-3. LMM-adjusted means and 95% confidence intervals (CIs) for each arm/outcome/time point, as well as change (Δ) in composite scores within and across groups, were produced using emmeans v.1.7.0. Multivariate imputation by chained equations was performed using mice v.3.14.0. For all analyses, uncorrected p-values are reported and were deemed statistically significant at < 0.05.
Cognitive retest effects
Individuals undergoing repeated cognitive testing on the same measures are likely to learn task characteristics, which may result in improved performance over time. This study did not include an inactive comparator condition, and thus, cognitive retest effects could not be quantified empirically; we thus adjusted for these in statistical analyses. Cognitive retest effects were modeled based on recommendations [34]. Among the three strategies available, we utilized the first approach (referred to by the authors as “jump”); this specification was selected as the two alternatives were highly collinear with time (see Additional file 1: Supplementary Methods and Table S1 for details). This approach engenders the inclusion of a time-varying LMM covariate taking the value of “0” at baseline and “1” at the two subsequent visits. This coding represents participants’ lack of prior experience with the cognitive tests at baseline and their increased familiarity with these at weeks 8 and 24. The process of deciding which of the three cognitive retest effect specifications to use is described in Additional file 1: Supplementary Methods. The chosen cognitive retest effect covariate (coded as “0,” “1,” “1”) was only included in LMMs using linear time (i.e., weeks 0, 8, 24); both the effects of time and cognitive retesting could be estimated in these models. However, the cognitive retest effect parameter was not estimable (and thus not included) in LMMs using factorial time (i.e., according to visit).