Computerized Games versus Crosswords Training in Mild Cognitive Impairment

The study design and methods were published previously.¹⁹ The study was conducted at two sites: Columbia University/New York State Psychiatric Institute (New York; Columbia University was the lead coordinating site) and Duke University Medical Center (Durham, NC). The sites’ institutional review boards approved the study, which was funded by the National Institute on Aging. Lumos Labs provided specific games and crossword training modules with technical support from their Web-based platform without cost and had no role in study design, data interpretation, or publication. Participants were not charged but had no poststudy commitment. A data and safety monitoring board provided oversight. The protocol is available with the full text of this article at evidence.nejm.org.

Participants were recruited from clinical referrals, supplemented by advertising. All participants signed the informed consent form, which indicated random assignment to one of two cognitively stimulating exercises — Web-based cognitive games or computerized crossword puzzles — without stating which treatment might be better. Key inclusion criteria were aged 55 to 95 years, English-speaking ability, and meeting Alzheimer’s Disease Neuroimaging Initiative criteria for early MCI or late MCI. Early MCI was defined by a Wechsler Memory Scale-III Logical Memory delayed recall score (scores range from 0 to 25, with higher scores indicating better verbal recall) of 3 to 6 with 0 to 7 years of education, score of 5 to 9 with 8 to 15 years of education, and score of 9 to 11 with 16 or more years of education. Late MCI was defined by a Wechsler Memory Scale-III Logical Memory delayed recall score of ≤2 with 0 to 7 years of education, score of ≤4 with 8 to 15 years of education, and score of ≤8 with 16 or more years of education. Additional inclusion criteria were a Folstein Mini-Mental State Examination score of ≥23 of 30 (range, 0 to 30; a higher score indicates better cognition) and availability of an informant, such as a family member, to provide information about the participant’s functioning. Participants were required to have a home computer with Internet connection to access the study website. Key exclusion criteria were current major psychiatric or neurologic disorder, dementia, contraindication to magnetic resonance imaging (MRI; conducted on 3.0 T scanners), and use of online cognitive games or crossword puzzles two times per week or more in the past year. Prescribed cholinesterase inhibitors and memantine were continued. Use of high-dose opioids, anticholinergics, and/or benzodiazepines in lorazepam equivalents of ≥1 mg per day were also exclusion criteria. Inclusion and exclusion criteria are described in more detail in Table S1 of the Supplementary Appendix available with the full text of this article at evidence.nejm.org.

Patients were randomly assigned 1:1 to games or crosswords, stratified by site, age (≤70 and >70 years), and early MCI versus late MCI. An unblinded research coordinator conducted training sessions, and a blinded research coordinator administered cognitive and functional assessments. Participants were unblinded.

Lumos Labs provided Web-based cognitive games and crossword puzzles. Lumos Labs account credentials included a study-specific email address and password. Each games session was composed of 6 modules randomly selected from 18 available modules that included memory tasks, matching tasks, spatial recognition tasks, and processing speed tasks. For games, difficulty was scaled over time using the participant’s Lumosity Performance Index, which considered three variables: games performance, cognitive area performance, and overall cognitive performance.¹⁹ Participants received their overall games performance score at the end of each session. Lumos Labs also provided computerized crossword puzzles of medium difficulty — intended to be equivalent to the New York Times’ Thursday crossword puzzles — without performance-based scaling over time. If the puzzle was completed within 15 of the allotted 30 minutes, a second puzzle was presented. The participant could view the correct answers at the end of the session but did not receive a score.

Participants were evaluated in person at five scheduled visits (weeks 0, 12, 32, 52, and 78), and research staff conducted three additional scheduled phone calls (weeks 20, 42, and 64). Figure S1 describes the timeline of the study. Initial intensive, home-based computerized training for games or crosswords consisted of four 30-minute training sessions per week for 12 weeks. Subsequent booster training was composed of four 30-minute sessions, completed over 1 week and occurring at weeks 20, 32, 42, 52, 64, and 78. During weeks 32, 52, and 78, participants completed three sessions at home and the fourth in clinic. During weeks 20, 42, and 64, participants completed all four sessions at home. Unblinded study coordinators received weekly electronic reports of completed sessions and contacted participants who had not completed sessions to attempt to improve adherence.

The primary outcome was change from baseline in the 11-item Alzheimer’s Disease Assessment Scale-Cognitive (ADAS-Cog) total score²⁰ at 78 weeks, adjusted for baseline. Scores on ADAS-Cog range from 0 to 70, with higher scores indicating greater cognitive impairment. The secondary cognitive outcome was the change from baseline in composite z score of neuropsychological tests at 78 weeks, adjusted for baseline. Lumosity software support for the Neurocognitive Performance Test (NCPT),²¹ used for one of the secondary cognitive outcomes, was discontinued halfway through the study; therefore, this measure was not analyzed.

Additional secondary outcomes at 78 weeks included the short-form University of California, San Diego Performance-Based Skills Assessment (UPSA) score (range, 0 to 100; higher scores indicate better functional performance), a performance-based measure of skills to organize, plan, and perform such tasks as counting physically presented change with bills and coins or planning a trip²²; the informant-reported Functional Activities Questionnaire (FAQ) score (range, 0 to 30; higher scores indicate greater impairment in instrumental activities of daily living) with items that include paying bills, shopping, remembering appointments, and taking medications²³; and the proportion of participants transitioning from MCI to clinical dementia.

Hypothesized moderators of outcome were apolipoprotein E e4 genotype, baseline MRI hippocampal volume, and baseline University of Pennsylvania Smell Identification Test.²⁴ Exploratory analyses were conducted for change from baseline to week 78 in hippocampal volume (mean total volume), calculated as the summation of right and left hippocampal volumes, and cortical thickness, evaluated with Freesurfer version 6.0 from the MRI three-dimensional spoiled gradient recalled acquisition. Resting-state functional MRI results with changes in default mode network connectivity²⁵ and related measures will be reported separately from this article.

Demographics and baseline values were compared between treatment arms using Wilcoxon’s rank-sum test for continuous variables and Fisher’s exact test for categorical variables. We conducted linear mixed-effects model repeated-measures analysis to assess the effect of treatment on cognitive and functional outcomes. For each outcome, the change in the measure (baseline minus study time point) was the dependent measure, and treatment, study time point, and their interaction were the predictors, adjusting for the baseline value of the outcome measure. The model was then further adjusted for the stratification factors of site, age group, and MCI type. Each model included all time points, with a focus on 12 weeks (end of intensive training sessions) and 78 weeks (primary end point). Effects of treatment on change in MRI outcomes (week 0 to week 78) were evaluated in linear regression models.

Hypothesized moderators were assessed individually by including the moderator variable and the associated moderator–interaction terms in regression models (week 0 to week 78). Fisher’s exact test was used to test the association between the binary outcome of dementia and treatment group, and the associations between adverse events and treatment group.

A sample size of 100 participants was projected to detect a Cohen’s d effect size of 0.58 (medium effect size) with 80% power for the primary outcome, assuming dropout was uniformly distributed over time to reach 20% by 78 weeks. On the basis of the data and safety monitoring board recommendation to balance site demographics, the projected sample size was increased to 110.

Analyses were conducted on the intent-to-treat sample (i.e., all randomly assigned participants) according to the assigned treatment. Missing data on outcome variables were addressed by linear mixed-effects models, which do not require complete measurements under the assumption that the data were missing at random. Two-sided α<0.05 was the criterion for significance for the primary outcome.¹⁹ For secondary outcomes, there was no adjustment for multiple statistical comparisons; point estimates with unadjusted 95% confidence intervals (CIs) are reported.

Between November 2017 and February 2020, 168 participants were screened for eligibility. Enrollment ended in March 2020 because of pandemic restrictions. Of 109 enrolled participants, 2 dropped out before random assignment (Fig. 1), and the final sample included 107 randomly assigned participants. The last study visit occurred in September 2021. The mean age of participants was 71.2 years (SD, 8.8 years), and 58% were female. Demographics were similar to the general population of people with MCI and included 25% minority populations (Table S2). Participants assigned to games and crosswords did not differ in demographics and baseline measures (Table 1); therefore, only the three stratification factors were examined as covariates. During the study, 9 (17.6%) of 51 participants dropped out in the games group compared with 7 (12.5%) of 56 participants in the crosswords group. Among late MCI participants, 6 (20.0%) of 30 dropped out in the games group and 5 (15.2%) of 33 dropped out in the crosswords group.

Over 78 weeks, ADAS-Cog score, which was adjusted for baseline ADAS-Cog score, showed a small decline for games and an improvement for crosswords (LS means difference, −1.44 points; 95% CI, −2.83 to −0.06; P=0.04; Table 2). ADAS-Cog scores worsened by an average of 0.4 points in the games group (9.53 to 9.93) and improved by nearly 1 point in the crosswords group (9.59 to 8.61; Table S3). At week 12, ADAS-Cog score showed a similar LS means difference of −1.35 points (95% CI, −2.71 to 0.00). Compared with the games group, the Cohen’s d effect size for crossword puzzles for change in ADAS-Cog scores from baseline was 0.43 (consistent with a medium effect size) at week 12 and 0.34 (consistent with a small effect size) at week 78 (Fig. 2).

Improvement of ≥2 points in ADAS-Cog score was observed in 25.0% of the games group and in 37.3% of the crosswords group at 78 weeks. Within the late MCI group, change in ADAS-Cog score worsened for games and improved for crosswords at week 12 (LS means difference, −1.92; SE, 0.85; 95% CI, −3.6 to −0.23) and week 78 (LS means difference, −2.45; SE, 0.89; 95% CI, −4.21 to −0.70), but this effect was not observed in early MCI (change in ADAS-Cog score at week 12: LS means difference, −0.26; 95% CI, −2.31 to 1.80; and at week 78: LS means difference, 0.28; 95% CI, −1.78 to 2.35). There was no treatment by site interaction for the primary outcome. Change in composite neuropsychological test score was similar in the two groups (Table 2 and Fig. 3).

In a sensitivity analysis using analysis of covariance, results (point estimates) were similar to the primary mixed-model analysis, although the 95% CIs for ADAS-Cog included the null, with slight inflation of the SE (Table S5). Furthermore, in analyses that assessed the adequacy of the fitted mixed-effects models, the residual plots did not show a systematic pattern, indicating no obvious violation from the underlying assumptions in mixed-effects models.

None of the three prespecified potential moderators — apolipoprotein E e4 genotype, baseline hippocampal volume, and baseline University of Pennsylvania Smell Identification Test — affected the change in ADAS-Cog score.

Change in UPSA score, adjusted for baseline UPSA score, did not differ between the games and crosswords groups (Table 2 and Fig. 3). Change in FAQ score, adjusted for baseline FAQ score, was greater for the games group than for the crosswords group at week 78 (LS means difference, −1.08 points; 95% CI, −1.97 to −0.18; Table 2 and Fig. 3), even after adjusting for covariates in the model (LS means difference, −1.18 points; 95% CI, −2.04 to −0.33; Table 2). For change in FAQ score, effects were similar for young versus old participants, early MCI versus late MCI, and the two sites. FAQ raw scores are described in Table S3.

Some participants exceeded the targeted number of enrolled sessions (intended maximum, 72; actual maximum observed, 116) because of the expanded time windows between in-person assessments, particularly during pandemic-related delays. The total number of enrolled sessions was similar for games versus crosswords, young versus old participants, and early MCI versus late MCI. Participants at Duke enrolled in more sessions than at Columbia (median, 72; interquartile range, 64 to 74, vs. median, 68; interquartile range, 66 to 75 sessions per participant; Hodges-Lehmann estimate of location difference, 3; 95% CI, 0 to 6). The coronavirus disease 2019 pandemic led to delays of ≥60 days for in-person testing during the trial in 6.3% of participants (7.3% games; 5% crosswords).

Progression to a diagnosis of dementia occurred in 6 (10.7%) of 56 participants in the crosswords group and 8 (15.7%) of 51 in the games group (odds ratio, 0.65; 95% CI, 0.17 to 2.32). Reversion from MCI to normal cognition occurred in 17 (30.4%) of 56 participants in the crosswords group and 12 (23.5%) of 51 in the games group (odds ratio, 1.41; 95% CI, 0.55 to 3.71).

Decline in hippocampal volume at week 78, adjusted for baseline hippocampal volume, was greater for the games group than for the crosswords group after adjusting for the three stratification variables (LS means difference, 34.07; SE, 17.12; 95% CI, 0.51 to 67.63). Decline in cortical thickness at week 78, adjusted for baseline cortical thickness, was greater for the games group after adjusting for the three stratification variables (LS means difference, 0.02; SE, 0.01; 95% CI, 0.00 to 0.04).

The number and type of adverse events, including serious adverse events, for the treatment groups are described in Table S4. None was considered to be related to treatment conditions. Coronavirus disease 2019 diagnosis occurred in 5 of 51 participants in the games group and 0 of 56 participants in the crosswords group; its inclusion as a covariate did not change ADAS-Cog score results.