Does Computerized Cognitive Training Improve Diabetes Self-Management and Cognition? A Randomized Control Trial of Middle-Aged and Older Veterans with Type 2 Diabetes

Abstract

1. Aims.

This randomized control trial compared an adaptive computerized cognitive training intervention with a non-adaptive version. The primary hypothesis predicted better diabetes self-management in type 2 diabetes patients at 6 months post-intervention than baseline, with seven secondary outcomes.

2. Methods.

Intent-to-treat analysis of veterans without dementia aged 55+ from the Bronx, NY and Ann Arbor, MI (N=90/per arm) used linear mixed model analyses.

3. Results.

Contrary to the hypothesis, only memory showed more improvement in the adaptive arm (p<0.01). Post-hoc analyses combined the two arms; self-management improved at six-months post-intervention (p<0.001). Memory, executive functions/attention, prospective memory, diastolic blood pressure, and systolic blood pressure improved (p < 0.05); hemoglobin A1c and medication adherence did not improve significantly.

4. Conclusions.

The adaptive computerized cognitive training was not substantially better than non-adaptive, but may improve memory. Post-hoc results for the combined arms suggest computer-related activities may improve diabetes self-management and other outcomes for middle-aged and older patients with type 2 diabetes. Practice effects or awareness of being studied cannot be ruled out.

1. INTRODUCTION

Type 2 diabetes mellitus (hereafter, diabetes) has been associated with increased risk for cognitive decline, mild cognitive impairment (MCI), and dementia[1]. Individuals with both diabetes and cognitive impairment tend to have greater difficulties in managing their diabetes[2]. Even relatively low cognition in cognitively intact individuals can impair adherence to medications and overall disease self-management[3, 4]. In turn, deficits in disease self-management are associated with poor glycemic and blood pressure control in diabetes, which then increase the risk of impaired cognition[5], provoking a downward spiral of co-morbid disease severity[6]. Finding interventions that enhance cognitive function or limit its decline in patients with diabetes thus hold the potential for breaking this spiral. Improved diabetes self-management can provide better control over the disease process.

Interventions for cognitive remediation have been defined by the Cognitive Remediation Expert Working Group as those “targeting cognitive deficit (attention, memory, executive function, social cognition, or metacognition) using scientific principles of learning with the ultimate goal of improving functional outcomes”[7]. In older adults without dementia, computerized cognitive training (CCT) is a relatively easily implemented, inexpensive component of cognitive remediation[8]. Although still controversial[9], CCT has shown promise as a tool for improving cognitive functioning or limiting its decline over time[9, 10]. For studies assessing the utility of CCT, four increasing levels of efficacy have been offered by Harvey et al.[9]: 1. The training tasks themselves; 2. Nontrained cognitive tasks (near transfer); 3. Cognitively demanding functional tasks (far transfer); 4. Everyday function (environmental transfer). We extend this hierarchy to a fifth level, modifiable biological functions (somatic transfer), indicating everyday physiological function. Most randomized controlled trials (RCT) of CCT have focused primarily on near transfer, usually after only a brief post-intervention interval, and often with a relatively modest sample[11].

The present RCT (#NCT 01736124) involved older veterans, a population with relatively high rates of diabetes[12]. Veterans with diabetes and demonstrable issues with disease self-management were randomly assigned to an adaptive-CCT program or an active control group, measured at baseline and at six months post-intervention, the primary endpoint. The primary outcome was post-treatment change from baseline in diabetes self-management, with seven secondary outcomes. The primary linear mixed model analysis used the intent-to-treat sample. We hypothesized that outcomes would be better in participants who were assigned to an adaptive-CCT program compared with those assigned to the active control group.

2. METHODS

2.1. Settings.

This RCT recruited veterans with diabetes from two Veterans Affairs Medical Centers (VAMCs) – the James J. Peters VAMC in the Bronx, NY and the Ann Arbor VAMC in Washtenaw County, Michigan. These sites were distinct not only geographically, but in urbanicity and demographics; drawing participants from both would potentially provide greater generalizability of our results. The Bronx, with a population exceeding 1.4 million people, has a median household income of $38,085, and a large minority population (43.6% African American; 56.4% Hispanic, of any race)[13]. Ann Arbor, with a population of about 367,000, has a median household income of $69,434 and a small minority population (12.3% African American; 4.8% Hispanic, of any race)[13].

2.2. Participants.

The eligibility and recruitment details for this RCT have been described in detail elsewhere[14]. Briefly, we recruited United States military veterans at least 55 years of age with a type 2 diabetes diagnosis and evidence of less-than-optimal diabetes self-management as defined by a score of 18 or less out 20 on the Diabetes Self-Management Questionnaire (DSMQ; see Screening). To facilitate computer-experienced recruitment the initial age threshold was decreased from 65 to 60 and finally to 55. We excluded participants with dementia, any major psychiatric or neurological diagnosis that affects cognition (e.g. Parkinson’s disease, stroke), a significant vision or hearing impairment, and those prescribed an Alzheimer’s disease medication. Eligibility also required familiarity with and access to a computer and internet and identification of a cooperative informant who could provide independent information on the participant’s cognitive functioning and diabetes self-management. We amended the protocol to purchase and lend laptops with mobile WiFi to otherwise eligible participants.

At each site, veterans with diabetes were recruited by mailings, using lists supplied by the VAMC medical informatics staff, based on the Computerized Patient Record System (CPRS), the VAMC electronic medical record. Information from the CPRS allowed for the identification of veterans at least age 55, with a high (>8.0%) hemoglobin A1c (HbA1c) observation in the past year, and none of the exclusionary diagnoses. Additional Bronx veterans were recruited through face-to-face encounters at VAMC diabetes-related clinics and open houses, and flyers at the VAMC. At enrollment, all participants signed informed consent. The study was approved by the VA Central Institutional Review Board and the Research and Development committees of both VAMCs. The study was conducted from April 2015 to January 2018 and registered at clinicaltrials.gov (NCT 01736124).

2.3. Screening.

Following informed consent, eligibility criteria were assessed by direct interview of participants and CPRS review. The DSMQ[15], described below, was administered to prospective participants and a score >18 of a possible 20 excluded veterans with near optimal management. The Clinical Dementia Rating Scale (CDR)[16] was administered to both the participant and an informant, excluding scores ≥1, indicating a definite dementia. To further identify those without dementia for participation, the Mini-Mental State Exam (MMSE)[17] was administered, excluding those who scored less than 25, out of 30. Basic competence in computer and internet use was required. This was assessed first by asking potential participants about their computer experience, followed by project staff observing the participant in a computer task. In addition to assessing their ability to use the computer, the task required color discrimination, assessing color blindness, the absence of which was required.

2.4. Randomization and Double-Blind Assessments.

At each site, a computer program randomized the order of assignment to adaptive-CCT and active control groups within pairs of consecutive enrolled and eligible participants. All staff members who conducted assessments were blind to group membership, as were the participants themselves. Data analysis was conducted blind to group status.

2.5. Adaptive-CCT Intervention and Active Control.

Both the adaptive-CCT and the active control are multimedia, interactive, and can be performed at flexible hours without a live instructor, typically at the subject’s residence. The adaptive-CCT program used in this RCT was Cognifit^™[18], a commercially available, web-accessed training program designed to improve cognition by targeting the user’s weaker cognitive functions. The challenge at each session was adapted according to the user’s prior performance. Participants were instructed to have sessions three times per week for eight weeks, with at least one day of rest between sessions, for a total of 24 sessions. Each 20-minute session included a unique combination of three games accessing a variety of cognitive abilities. The program utilized a wide variety of games encompassing a broad range of cognitive functions, to keep the user interested and engaged. The first two games of each session were training tasks intended to enhance specific cognitive functions. The third game of each session was a non-adaptive assessment task, yielding an overall cognitive score provided to the user.

The active control program used the same Cognifit^™ games; it differed only by not having personalized adjustment of challenge of training games, and cognitive score feedback. As opposed to unrelated computer games or a passive control, the use of a non-adaptive CCT as an active control provided a clear test of whether adjusting the challenge according to the performance of the participant would lead to better outcomes.

Two months after the end of training, participants in each group were asked to engage in a booster week of three sessions, correspondingly adaptive or non-adaptive.

2.6. Participant Timeline.

Some eligibility criteria were assessed in a brief telephone interview before enrollment, others afterwards at the baseline assessment (Visit 1)[14]. Randomization to the adaptive-CCT or active control followed completion of baseline assessment and determination of eligibility. Both CCT programs were initiated at Visit 2, soon after randomization. Visits 3 was an outcome assessment scheduled for immediately after the intervention. The booster intervention was scheduled two months after the completion of training, called “Visit 4” to indicate its position, although usually initiated by a telephone conversation, rather than face-to-face. Visits 5 and 6 were outcome assessments scheduled six- and twelve-months post-intervention. Since the primary and many secondary outcomes of this study were for changes beyond the immediate intervention period, the primary efficacy evaluations were change from baseline (Visit 1) to the primary endpoint time, six months post-treatment (Visit 5).

To strengthen participation rates and adherence, staff members called each active participant once a week, to monitor their activity with the program and address any technical difficulties. Contact was attempted to identify the reason if a participant dropped out. In addition, efforts were made to bring back the participant for the outcome assessments.

2.7. Primary Outcome Measure: DSMQ.

The primary outcome for this RCT was specified a priori as the DSMQ, an environmental transfer (level 4). The DSMQ has been shown to be significantly associated with HbA1c and with receiving diabetes-related services[15]. It assesses five areas of diabetes self-care (medication adherence, exercising, following an eating plan, blood glucose monitoring, and foot care). At baseline, participants and their informants were independently asked “Over the past year, how difficult has it been for you/the patient to do each of the following, exactly as the doctor who takes care of your diabetes suggested?” At subsequent assessments, we used the initial phrase “Since the last time we met….” Respondents rated each category on a five-point scale ranging from 0 (“So difficult that I/[he or she] couldn’t do it at all”) to 4 (“Not difficult, I/[he or she] got it exactly right”), with a “not applicable” option, as well. Category scores are summed, ranging from 0 to 20, with higher scores representing better self-management. The total score was prorated for “not applicable” categories. The average DSMQ score between participant and informant was used to manage doubt about the validity of each.

To see whether there was a discrepancy between the perspectives of the participant and informant a matched sample t-test of the baseline DSMQ scores showed minimal discrepancy (t = −0.4353, d.f. = 161, p = 0.6639). The modest/moderate intra-class correlation between the participant and the informant was 0.46, which is only an approximate measure of inter-rater reliability because the participant and informant do not have the same perspective on self-management.

2.8. Secondary Outcome Measures.

All seven secondary outcomes were assessed at baseline (Visit 1). All except medication adherence (see below), were assessed at three outcome assessment visits. These outcomes had four levels of efficacy – 2 through 5 – described in the Introduction; performance in the training task itself (level 1) was not a secondary outcome.

2.8.1. Near Transfer (Level 2): Memory and Executive Function/Attention.

As described in detail elsewhere[19], an experienced neuropsychologist (EG-B) trained and certified research associates at each site, who administered the neuropsychological battery of 9 tests[14] yielding 16 scores. For each test, the same version was employed at each assessment. Questions and issues about administration were reviewed routinely and discussed at bi-weekly conference calls. Tests of memory were Word List Memory and Logical Memory (Story A), each providing three scores: immediate recall, delayed recall, and recognition. Tests of executive function/attention were Target Cancellation Tests: TMX and Diamond, Trail Making Test: A and B, Digit Symbol Substitution Test. And Digit Span: Forward and Backward. (Tests of the language domain – not a secondary outcome – were Fluency, Similarities, and Boston Naming.)

Cognitive domain scores were derived in a three-step procedure using the sixteen neuropsychological test scores at each assessment. 1) Since the failure of an otherwise willing participant may be a serious indication of substantial impairment in cognition, for a participant with more than two missing scores out of 16, all 16 were considered missing – there was only one at baseline. 2) After reversing (multiplying by −1) the test scores for the two Trail Making Tests – the only tests where lower scores indicated better performance – z-scores for each test were calculated from the non-missing values. For a participant’s missing z-score, we imputed the mean of the participant’s non-missing z-scores in that cognitive domain. 3) At the baseline assessment, the mean (mean_b) and standard deviation (SD_b) were calculated for each test using both directly calculated and imputed step 2 results. At each subsequent assessment, the baseline means and SDs were used to standardize the step 2 results for each test: standardized score = (score – mean_b) / SD_b. Thus, for baseline, the standardized score was the z-score, but this was not the case for the other assessments, since their standardized score was not based on the mean and SD for that assessment. By calculating standardized scores this same way at every subsequent assessment, the participant’s change score between a pair of assessments was simply the same as the change for the neuropsychological test divided by SD_b for that test. At every assessment, the domain scores were the means of the relevant standardized scores.

2.8.2. Far Transfer (Level 3): Prospective Memory.

Since diabetes medication is taken at specified times, the Appointments subtest in the Rivermead Behavioral Memory Test (3^rd Edition)[20], a time-based test of prospective memory, was used.

2.8.3. Environmental Transfer (Level 4): Medication Adherence.

In contrast to the other outcome measures, which for each visit could be measured at a single day, medication adherence was assessed by the Continuous Multiple interval measure of Gaps in therapy (CMG[21]), using diabetes medication refill data from the CPRS. We identified diabetes associated medication refills and their dates occurring within one year prior to baseline (Visit 1) as pre-intervention and refills occurring within one year after completion of intervention (Visit 3) as post-intervention. To calculate pre-intervention and post-intervention CMG, the total number of days in treatment gaps was divided by the total number of days within the time period of interest. Multiplying this fraction by 100 provide the CMP percentage – the percentage of the time period with gaps in treatment – where higher a percentage indicates worse medication adherence.

2.8.4. Somatic Transfer (Level 5): HbA1c, Systolic Blood Pressure, and Diastolic Blood Pressure.

The aim of diabetes self-management is to control factors that increase the risk of diabetes complications, so glycemic control (HbA1c) and blood pressures (systolic and diastolic) were assessed at baseline and each outcome assessment visit. HbA1c was measured at each assessment visit with a fingerstick of blood using a Siemens DCA Vantage Analyzer. Blood pressures were averages of two measures, five minutes apart, on resting participants in a seated position using an Omron HEM-780 device, with an approximately 6-inch cuff.

2.9. Other Participant Characteristics Assessed.

Age, sex, race/ethnicity, years of education, MMSE, and CDR were collected on all subjects at baseline. To further characterize the sample, health literacy was measured by the Short-Test of Functional Health Literacy in Adults[22], depressive symptoms by the Geriatric Depression Scale[23], and subjective cognitive complaints by a cognitive self-report questionnaire[24].

2.10. Statistical Procedures.

Primary analyses were conducted on the ITT sample – all eligible participants who were randomized and had complete baseline assessment data for the primary outcome analysis. Secondary analyses were conducted using the ITT participants who also completed Visit 5 (Retained).

Demographic and clinical variables were compared between adaptive-CCT and active control groups in the ITT sample, and between retained vs. non-retained at Visit 5. Between-group comparisons used t-tests and chi-square tests, as appropriate.

Efficacies of the adaptive-CCT intervention vs. the active control group on the primary and secondary outcomes were tested using linear mixed models (LMM) procedures. Specifically, the estimated equation was:

$$\text{outcome}=\text{b}0+\text{b}1*\left(\text{intervention group indicator}\right)+\text{b}2*\left(\text{visit indicator}\right)+\text{b}3*\left(\text{intervention group indicator}\right)*\left(\text{visit indicator}\right),$$

where the intervention group indicator was defined as active control CCT = 0 and adaptive CCT = 1, and the visit indicator was defined as Visit 1 = 0 (reference), Visit 5 = 1. These main effects and their interaction, intervention group X visit, were entered as fixed factors, and participants as a random factor. Defined as such, b0 estimates the active control group at Visit 1, b1 estimates the difference in the outcome between adaptive CCT group compared to active control at Visit 1, b2 estimates the difference in the outcome between Visit 5 compared to Visit 1 in the active control group. The main test of interest, intervention group X visit interaction, estimated by b3, is the discrepancy between the adaptive CCT group and the active control in the change from Visit 1 to Visit 5. This interaction is the residual of outcome after removing the other three effects.

The usual LMM analysis also estimates an equation for a constant (the grand mean of the four outcomes), a main effect of group (with outcomes averaged across visits), a main effect of time (with outcomes averaged across groups), and group x time interaction (the corresponding residual). Its different units of measurement of the constant and the group and time main effects do not affect the analysis of the fitted overall model for those three variables, so they also do not affect the analysis of the residual interaction. The advantage of the analysis of this study over the usual LMM analysis is that b1 and b2 test hypotheses relevant the study design, a baseline difference between groups (attributable to unlucky randomization) and an active control change between Visit 1 and Visit 5 (attributable to an appropriate choice of the active control intervention).

Although hypothesized demonstration of benefit of adaptive-CCT beyond that of the active control was directional, tests of significance were set a priori to be two-sided at the 0.05 level, to include a test for the possibility that the adaptive-CCT is inferior to the active control.

Except for medication adherence, the primary endpoint that was compared with baseline (Visit 1) was six months post-intervention (Visit 5); secondary endpoints were immediately after the intervention (Visit 3) and one-year post-intervention (Visit 6). Secondary analysis included intermediate outcome assessment visits and their interactions with group as covariates. Medication adherence compared only two periods, one-year pre-intervention and one-year post-intervention.

3. RESULTS

3.1. Sample characteristics.

A total of 199 participants were screened, of whom 7 were ineligible and 12 others who dropped out prior to completing Visit 1 (Consort Diagram). Thus, there were 180 in the ITT sample (90 adaptive-CCT, 90 active control). There were no significant differences in age, gender, or race/ethnicity between participants who met eligibility requirements compared with those who did not (data not shown).

Table 1 compares participant characteristics by adaptive-CCT intervention vs. active control group, and by retained vs non-retained at Visit 5. There were no statistically significant differences in characteristics between intervention and active control groups, including the baseline outcomes, attributable to successful randomization. The retained differed from non-retained by ethnicity (more non-Hispanic whites and fewer non-Hispanic blacks) and site (more from Ann Arbor), but not by age, sex, and education – the usual demographic predictors of cognition in older adults. Other significant differences were that retained were more likely to have CDR=0, higher MMSE, and higher health literacy scores, but they did not significantly differ on baseline outcomes.

Table 1.

Baseline Characteristics: Comparison by Group and by Retained Status

	ITT Sample			Visit 5 Retained Sample
Variable	Active control	Adaptive-CCT	p-value	Non-Retained	Retained	p-value
	Demographic and Other Sample Characteristics
Total N	90	90		65	115
Age, mean (SD)	70.7 (6.9)	69.6 (6.5)	0.36	70.9 (7.5)	69.7 (5.9)	0.33
male,%	96.7	98.9	0.62	95.4	99.1	0.14
Race/ethnic group,%			0.47			<0.001
non-Hispanic white	47.8	53.3		30.8	61.7
non-Hispanic black	30.0	27.8		41.5	21.7
Hispanic	21.1	16.7		23.1	16.5
Other ethnic group	1.1	2.2		4.6	0.0
years of education, mean(SD)	14.3 (2.7)	14.4 (2.3)	0.99	13.9 (2.1)	14.6 (2.7)	0.10
Site=Bronx(%)	57.8	58.9	0.88	80.0	46.1	<0.001
CDR=0 (%)	82.2	85.6		72.3	90.4	0.001
MMSE, mean(SD)	28.3 (1.4)	28.5 (1.3)	0.25	28.0 (1.3)	28.6 (1.4)	0.003
CSRQ, mean(SD)	62.3 (14.1)	63.4 (12.8)	0.44	65.0 (11.5)	61.9 (14.2)	0.18
GDS, mean(SD)	6.3 (6.4)	7.0 (6.2)	0.28	8.0 (6.8)	6.0 (5.9)	0.06
STOFHLA, mean(SD)	33.5 (3.5)	32.3 (6.0)	0.35	31.4 (6.6)	33.8 (3.5)	<0.001
Follow-up outcome assessments, mean(SD)	3.0 (1.2)	2.8 (1.3)	0.43	-	-
	Outcome Variables at Baseline
DSMQ, N, mean(SD)	90, 14.3 (2.4)	90, 14.2 (2.2)	0.78	65, 14.6 (2.3)	115, 14.1 (2.3)	0.28
Memory, N. mean(SD)	84, 0.01 (0.79)	85, −0.01 (0.72)	0.63	56, −0.34 (0.74)	113, 0.17 (0.71)	<0.001
Attention/Executive Functions, N, mean(SD)	84, 0.02 (0.63)	85, −0.02 (0.61)	0.64	56, −0.31 (0.58)	113, 0.15 (0.58)	<0.001
Prospective Memory, N, mean(SD)	85, 2.7 (1.3)	83, 2.8 (1.3)	0.84	58, 2.6 (1.38)	110, 2.8 (1.3)	0.25
Medication Adherence, N, mean(SD)	75, 14.6 (14.2)	75, 15.0 (10.7)	0.96	56, 17.4 (12.0)	94, 13.2 (12.6)	0.0504
Hemoglobin A1c, N, %, SD (mean mmol/mo)	88, 7.5, 1.3 (58)	88, 7.7, 1.5 (61)	0.45	64, 7.9, 1.6 (63)	112, 7.4, 1.2 (57)	0.14
Systolic Blood Pressure, N. mean(SD)	88, 131.8 (17.7)	88, 133.5 (18.7)	0.38	64, 133.5 (20.1)	112, 132.3 (17.2)	0.92
Diastolic Blood Pressure, N, mean(SD)	88, 73.6 (10.8)	88, 74.7 (10.3)	0.30	64, 74.7 (10.5)	112, 73.8 (10.6)	0.66

CDR = Clinical Dementia Rating Scale; MMSE = Mini-Mental Status Exam; CSRQ = Cognitive Self Report Questionnaire; GDS = Geriatric Depression Scale; STOFHLA = Short Test of Functional Health Literacy; DSMQ = Diabetes Self-Management Questionnaire

3.2. Primary Analyses at Visit 5.

Results from the primary endpoint (Visit 5) analyses on the primary and secondary outcomes, except medication adherence, are summarized in Table 2A. The first line (GROUPS) indicates that there were no differences between groups at Visit 1 for any outcomes, attributable to successful randomization. The second line (VISITS (ALL): 5 v. 1) indicates “collective” (without regard to group status) improvement for all outcomes from Visit 1 except for HbA1c, which did not change. Thus, behavioral and cognitive measures increased and both blood pressures measures decreased. The last line (GROUPS x 5 v. 1) indicates that there were no additional effects of the adaptive-CCT intervention compared with the active control group for the primary outcome, DSMQ, or most of the other secondary outcomes. The memory score was the one exception; the adaptive-CCT group showed significantly greater improvement in memory at Visit 5 than the active control group (b ± SE = 0.205 ± 0.078, p<0.01). Medication adherence (comparing one year pre-baseline to one year post-intervention) did not differ significantly by GROUPS (0.003 ± 0.019, n.s.), ALL: 5 v. 1 (−0.002 ± 0.017, n.s.), or the interaction of GROUPS x 5 v. 1 (−0.026 ± 0.025, n.s). Finally, the five DSMQ items were assessed individually. As with the total DSMQ score, neither GROUPS (baseline differences) nor GROUPS x 5 v. 1 (adaptive CCT effects) were significant for any single item (data not shown); for ALL: 5 v. 1 (collective change), significant improvement was found for taking medications (0.172 ± 0.063, p<0.01), checking glucose (0.240 ± 0.114, p<0.05), and checking feet (0.316 ± 0.099, p<0.01), but not exercise (0.048 ± 0.109, n.s.) or diet (0.166 ± 0.098, n.s.).

Table 2.

LMM ESTIMATION ON PRIMARY AND SECONDARY OUTCOMES FOR VISIT 5 (PRIMARY ENDPOINT)

A. VISIT 5 VS. VISIT 1
	DSMQ b (SE)	HbA1c b (SE)	Systolic BP b (SE)	Diastolic BP b (SE)	Memory b (SE)	Attention/Executive b (SE)	Prospective Memory b (SE)
GROUPS (AT VISIT 1)	−0.060 (0.346)	0.178 (0.217)	1.852 (2.583)	1.181 (1.588)	−0.019 (0.113)	−0.035 (0.094)	0.054 (0.177)
VISITS (ALL): 5 v. 1	0.944 *** (0.243)	0.077 (0.179)	−5.688* (2.273)	−3.322* (1.362)	0.311 ** (0.055)	0.207 ** (0.046)	0.676 ** (0.165)
GROUPS x 5 v. 1	−0.160 (0.345)	−0.122 (0.254)	2.125 (3.209)	2.918 (1.923)	0.205 ** (0.078)	−0.056 (0.065)	−0.043 (0.236)
B. All VISITS UP TO VISIT 5 VS. VISIT 1
GROUPS (AT VISIT 1)	−0.060 (0.347)	0.179 (0.210)	1.843 (2.547)	1.177 (1.570)	−0.019 (0.113)	−0.035 (0.097)	0.054 (0.173)
VISITS (All): 3 v. 1	0.773 *** (0.215)	−0.032 (0.145)	0.933 (1.959)	0.563 (1.203)	0.272** (0.051)	0.124 ** (0.040)	0.396 * (0.155)
VISITS (All): 5 v. 1	1.040 *** (0.229)	0.058 (0.152)	−5.641** (2.070)	−3.043* (1.271)	0.316 ** (0.054)	0.197 ** (0.043)	0.683 ** (0.165)
GROUPS x 3 v. 1	−0.162 (0.314)	0.007 (0.210)	−3.858 (2.822)	−1.766 (1.733)	0.077 (0.074)	0.017 (0.058)	0.012 (0.226)
GROUPS x 5 v. 1	−0.291 (0.326)	−0.091 (0.217)	2.070 (2.928)	2.569 (1.798)	0.186 * (0.076)	−0.057 (0.061)	−0.047 (0.237)

DSMQ = Diabetes Self-Management Questionnaire; HbA1c = hemoglobin A1c; BP = blood pressure;

Visit indicators defined as visit 1 = 0 (reference), intervention group indicator defined as adaptive CCT = 1, active control CCT = 0 (reference)

Coefficient on group indicator estimates the difference in the outcome between adaptive CCT group compared to active control (reference) at visit 1. Coefficient on visit estimates the difference in the outcome between visit 5 compared to visit 1 in the active control group (reference). The main test of interest, intervention group X visit interaction, estimates whether the change from visit 1 in the outcome differed between adaptive CCT group compared to active control.

^*p<0.05;

^**p<0.01;

^***p<0.001

3.3. Analyses with/at Other Endpoints.

Including intermediate Visit 3 in the model and its interaction with group (Panel 2B) did not meaningfully change any of the results at Visit 5. Table 3 shows results of comparisons with Visit 1 – for Visit 3 (Panel 3A), Visit 6 (Panel 3B), and including the intermediate assessment visits (Panel 3C). For both Visits 3 and 6 significant improvement from baseline was observed for the DSMQ and each of the neuropsychological test outcomes, but not HbA1c or the two measures of blood pressure. There were no significant interactions by group for either of these two visits.

Table 3.

LMM ESTIMATION ON PRIMARY AND SECONDARY OUTCOMES FOR VISIT 3 AND VISIT 6

A. VISIT 3 VS. VISIT 1
	DSMQ b (SE)	HbA1c b (SE)	Systolic BP b (SE)	Diastolic BP b (SE)	Memory b (SE)	Attention/Executive b (SE)	Prospective Memory b (SE)
GROUPS (AT VISIT 1)	−0.060 (0.346)	0.182 (0.201)	1.738 (2.611)	1.123 (1.553)	−0.019 (0.114)	−0.035 (0.097)	0.054 (0.186)
VISITS (ALL): 3 v. 1	0.772 *** (0.217)	−0.059 (0.119)	0.570 (1.949)	0.306 (1.138)	0.275 ** (0.055)	0.123 ** (0.038)	0.396 * (0.169)
GROUPS x 3 v. 1	−0.193 (0.318)	0.059 (0.172)	−3.900 (2.807)	−1.654 (1.640)	0.082 (0.080)	0.017 (0.056)	0.011 (0.247)
B. VISIT 6 VS. VISIT 1
GROUPS (AT VISIT 1)	−0.060 (0.341)	0.183 (0.217)	1.693 (2.715)	1.091 (1.598)	−0.019 (0.117)	−0.035 (0.094)	0.054 (0.189)
VISITS (ALL): 6 v. 1	0.844 ** (0.286)	0.221 (0.194)	−3.434 (2.497)	−1.375 (1.540)	0.369 ** (0.065)	0.120 * (0.047)	0.571 ** (0.159)
GROUPS x 6 v. 1	0.270 (0.412)	−0.354 (0.278)	−0.085 (3.548)	−0.557 (2.178)	−0.002 (0.093)	0.047 (0.068)	−0.179 (0.231)
C. All VISITS UP TO VISIT 6
GROUPS (AT VISIT 1)	−0.060 (0.347)	0.179 (0.216)	1.838 (2.572)	1.174 (1.587)	−0.019 (0.113)	−0.035 (0.097)	0.054 (0.171)
VISITS (ALL): 3 v. 1	0.789 *** (0.227)	0.003 (0.143)	1.038 (2.044)	0.571 (1.245)	0.270** (0.052)	0.123 ** (0.041)	0.394 ** (0.144)
VISITS (ALL): 5 v. 1	1.040 *** (0.242)	0.095 (0.150)	−5.273* (2.157)	−2.987* (1.314)	0.312 ** (0.055)	0.193 ** (0.043)	0.689 ** (0.153)
VISITS (ALL): 6 v. 1	1.004 *** (0.255)	0.190 (0.158)	−2.839 (2.292)	−0.705 (1.408)	0.356 ** (0.058)	0.115 * (0.046)	0.569 ** (0.164)
GROUPS x 3 v. 1	−0.185 (0.332)	−0.020 (0.207)	−4.037 (2.944)	−1.780 (1.793)	0.081 (0.075)	0.022 (0.059)	−0.002 (0.210)
GROUPS x 5 v. 1	−0.282 (0.345)	−0.127 (0.214)	1.765 (3.053)	2.554 (1.860)	0.189 * (0.078)	−0.054 (0.061)	−0.062 (0.221)
GROUPS x 6 v. 1	0.059 (0.368)	−0.276 (0.227)	−1.084 (3.261)	−1.610 (1.994)	0.011 (0.084)	0.030 (0.065)	−0.166 (0.238)

DSMQ = Diabetes Self-Management Questionnaire; HbA1c = hemoglobin A1c; BP = blood pressure;

Visit indicators defined as visit 1 = 0 (reference), intervention group indicator defined as adaptive CCT = 1, active control CCT = 0 (reference)

^*p<0.05;

^**p<0.01;

^***p<0.001

3.4. Secondary Analyses on Retained Sample.

Results from the Secondary Analyses on every outcome using the Retained Sample was similar to the result from the ITT sample (Supplementary Table 1).

3.5. Cognitive Domains and Diabetes Self-Management.

It was anticipated that the adaptive-CCT intervention would be particularly efficacious compared to the active control on cognitive measures which in turn would affect diabetes self-management. In post-hoc analyses, we examined the relationship between the cognitive domain score changes from baseline to Visit 5 in the two cognitive domain scores with change in the DSMQ (all of which were significant). Change in memory domain score over this period was significantly, albeit modestly, associated with change in DSMQ (r = 0.11, p = 0.017); change in attention/executive functions domain score was not (r = 0.05, p = 0.305).

3.6. Retention and Adherence.

Among the 24 assigned training sessions over the training period preceding all follow-ups, the number of sessions was a measure of adherence to instructions. An effort was made to retain for follow-up as many participants as possible regardless of their adherence. Among the 180 participants in the ITT analyses, 132 were retained in some follow-up (Consort Diagram). Retention was not significantly different between the adaptive CCT (62 of 90, 69%) and active control (70 of 90, 78%) groups (Chi-square [1] = 1.82, P = 0.18), but adherence distributions were very different. The 48 non-retained participants had a unimodal adherence distribution with 24 having either having no (15 participants) or one (9 participants) training sessions and a maximum of 18 (3.95 ± 5.53). The retained participants had a bimodal adherence distribution with the biggest gap between 16 and 20 sessions. A vast majority (113 of 132, 86%) had 20 or more training sessions, 109 (83%) had all 24. Among the 19 participants with 16 or fewer training sessions, 11 had none or one (3.11 ± 4.15), a similar adherence distribution to all the non-retained participants.

When adherence and its interaction with group were added to the LMM analysis as covariates for those outcomes assessed at Visit 5 (Supplemental Table 2), and for medication adherence, there were no significant changes of LMM results of the primary analyses, group by visit interaction. Adherence was significantly associated with memory and attention/executive functions meeting the Bonferroni criterion for eight outcome measures. For the adherence by group interaction, only prospective memory was significant.

4. DISCUSSION

This RCT examined whether, compared with an active control program, an adaptive-CCT program improved diabetes self-management – environmental transfer (level 4)[9] – for older veterans with suboptimal self-management of their diabetes. Contrary to our hypothesis, the adaptive-CCT group did not differentially show better self-management, as measured by the DSMQ (the primary outcome) at the six-month post-intervention (the primary endpoint time). Combining both the adaptive-CCT and active controls groups, participants collectively showed significantly better self-management at six-month post-intervention assessment compared with baseline, a small effect size, 0.29.

For only one secondary outcome, memory, the adaptive-CCT group showed significantly greater improvement at six-month post-intervention than the active control group. Collectively, the participants improved in memory as was seen for most of the other outcomes. Thus, the single instance of differential efficacy for the adaptive-CCT program was for a near transfer (level 2) outcome – the primary objective of CCT programs.

Four of the six other secondary outcomes – attention/executive functions (level 2), prospective memory (level 3), and systolic and diastolic blood pressures (level 5) – did not show better improvement at six-month post-intervention for the adaptive-CCT group than the active control group, but as with the DSMQ and memory, participants collectively improved. For medication adherence (level 4) and HbA1c (level 5) there was no differential improvement between the groups or collective improvement.

The primary endpoint time of this RCT was the six-month post-intervention assessment, determined a priori, to assess moderately enduring efficacy of an adaptive-CCT compared with the active control. At both secondary endpoint times (the post-intervention and twelve-month post-intervention), the DSMQ (level 4) and all three cognitive outcomes – memory and attention/executive functions (level 2) and prospective memory (level 3) – showed collective improvements from baseline, without detecting any differential change by group. Thus, the improvements in diabetes management and the cognitive outcomes in the full sample emerged early and persisted a year after the intervention. Both blood pressure measures (level 5), which significantly improved at the six-month post-intervention, were not improved either earlier at post-intervention or sustained at twelve-month post-intervention. There were no differences from baseline or between groups for medication adherence (level 4) and HbA1c (level 5) at any endpoint.

In contrast to most other studies of CCT, which have focused on cognitive benefits, our study has primarily focused on self-management of diabetes. A significant result with adaptive-CCT was not found. There were also seven additional secondary hypotheses representing four types of efficacy at increasing distances from the cognitive training. The two closest hypotheses, at level 2, were a priori cognitive domains – memory and attention/executive function. For these two cognitive hypotheses, the test of memory (P=0.008) would be significant with a Bonferroni correction, but not for all eight primary and secondary hypotheses. None of the other secondary hypotheses were significant.

An earlier RCT on 84 similarly aged Israeli patients with diabetes examined four cognitive outcomes – global cognition, delayed memory, memory and learning, and non-memory component – and the DSMQ using the same adaptive-CCT and active control programs[25]. The sample was very different demographically, with 40% women, better educated, and ethnically homogeneous. It found no significant differences between the adaptive-CCT (N=44) and active control (N=40) groups for any of the five outcomes assessed from baseline to six months post-intervention. The present study, with a larger sample, found that memory differentially improved in the adaptive-CCT group compared with the active control group. Assessed collectively, immediately after the intervention, the Israeli sample showed significant improvement in all five outcomes, and at their six-month post-intervention assessment, all but the non-memory component showed further statistically significant improvement. The significant improvement in these cognitive and self-management outcomes regardless of treatment arm was similar to findings in the present study. We know of no other RCT using CCT with diabetes-related outcomes. The sample size of the study is larger than most other CCT studies[11, 26].

There was an effort to retain participants regardless of their adherence. The distribution of adherence was bimodal with many having no training sessions or only one, or 24. Non-retention was limited to those with poor adherence, but some with poor adherence were retained. Including adherence in the analysis did not substantially change the results for primary and secondary outcomes.

A limitation of this study is that the primary outcome, DSMQ, is based on reports from the participant and usually but not always an informant. Ancillary analyses of the DSMQ score from the participants or from the informants did not show differences in statistical significance. For participants that also had an informant Supplementary Table 3 provides the means and standard deviations of DSMQ scores of the participants, the informants, and their average for the two arms of the study and for both the baseline visit and the six-month follow-up. Also included in the table are the same statistics in a subsample who were retained. In all cases, the values of participants and informants can readily be seen to be similar. An intrinsic limitation of the DSMQ is that it is not an objective measure.

The absence of a passive control group is a limitation that prevents ruling out practice effects as explaining the changes observed in the full sample. Improved cognitive test scores might be attributable to practice effects, but practice does not easily explain improvements in the DSMQ or the two blood pressure measures. For these latter outcomes, however, a Hawthorne effect[27] – improvements attributable to an individual’s awareness that one’s activities are being monitored – or placebo effect[28] cannot be ruled out in this design.

The RCT sample was limited to military veterans, almost all male, from two sites. Age and education, two demographic characteristics associated with cognitive performance, were well balanced between groups and so these were not included as covariates. Exploratory analyses controlling for age and education did not change the significance status of any of the outcomes.

In contrast to lack of significance between the two arms, non-retained participants differed significantly between the Bronx site (33%) and the Ann Arbor (17%) site (chi-square[1]= 5.73, p = 0.017).The more urban Bronx site had a larger proportion of minority and lower income participants; the burden of participating in clinical trials and longitudinal research is heavier for such individuals, making them harder to retain[29]. Controlling for site did not affect the significance status of the findings. Only for diastolic blood pressure did one site – Ann Arbor – drive the significant collective change from baseline to six-month post-intervention.

This RCT was premised on the idea that for patients with diabetes without dementia but with sub-optimal self-management of their disease, an intervention that enhanced cognition (level 2) would in turn improve disease self-management (level4). While no causal relationship can be inferred from an association, the significant, albeit modest, post-hoc association between baseline-to-six-month post-intervention change scores in the memory domain and the DSMQ is consistent with this initial premise. The results of this RCT did not support our primary hypothesis; the adaptive-CCT did not differentially show improved diabetes self-management (level 4). Nonetheless, along with the findings in the Israeli RCT[25], the secondary analyses of this RCT suggests that engagement in computer games and activities requiring cognitive skills per se are associated with a range of improved outcomes for older patients with diabetes at many levels of efficacy: near (memory and attention/executive functions), far (prospective memory), environmental (diabetes self-management), and even somatic (systolic and diastolic blood pressure) levels of transfer. Furthermore, our results raise the possibility that for patients with diabetes, adaptive-CCT may be especially useful for improving memory.

DATA AVAILABILITY

The data supporting the results reported in this article are available at https://clinicaltrials.gov/ct2/show/NCT01736124?term=silverman&cond=Diabetes+Type+2&draw=2&rank=1icalTrials.gov

The study sponsor/funder was not involved in the design of the study; the collection, analysis, and interpretation of data; writing the report; and did not impose any restrictions regarding publication of the report.