A randomized trial predicting response to cognitive rehabilitation in multiple sclerosis: Is there a window of opportunity?

Background: Cognitive training elicits mild-to-moderate improvements in cognitive functioning in people with multiple sclerosis (PwMS), although response heterogeneity limits overall effectiveness. Objective: To identify patient characteristics associated with response and non-response to cognitive training. Methods: Eighty-two PwMS were randomized into a 7-week attention training (n = 58, age = 48.4 ± 10.2 years) or a waiting-list control group (n = 24, age = 48.5 ± 9.4 years). Structural and functional magnetic resonance imaging (MRI) was obtained at baseline and post-intervention. Twenty-one healthy controls (HCs, age = 50.27 ± 10.15 years) were included at baseline. Responders were defined with a reliable change index of 1.64 on at least 2/6 cognitive domains. General linear models and logistic regression were applied. Results: Responders (n = 36) and non-responders (n = 22) did not differ on demographics, clinical variables and baseline cognition and structural MRI. However, non-responders exhibited a higher baseline functional connectivity (FC) between the default-mode network (DMN) and the ventral attention network (VAN), compared with responders (p = 0.018) and HCs (p = 0.001). Conversely, responders exhibited no significant baseline differences in FC compared with HCs. Response to cognitive training was predicted by lower DMN-VAN FC (p = 0.004) and DMN-frontoparietal FC (p = 0.029) (Nagelkerke R2 = 0.25). Conclusion: An intact pre-intervention FC is associated with cognitive training responsivity in pwMS, suggesting a window of opportunity for successful cognitive interventions.


Introduction
Multiple sclerosis (MS) is the most frequent demyelinating, inflammatory and neurodegenerative disease of the central nervous system (CNS) in young adults. Next to the well-known physical symptoms of the disease, cognitive impairment is present in up to 65% of the people with multiple sclerosis (PwMS). 1 Cognitive decline is characterized by a slowed information processing speed, impaired memory function as well as problems with attention and executive function, causing (severe) problems in patient's daily lives (e.g. unemployment). 1,2 Given the prevalence and impact of cognitive dysfunction in PwMS, effective cognitive training strategies are urgently needed. Restorative, non-pharmacologic interventions, especially computerized cognitive training, show potential for improvements in cognitive functioning in PwMS, 3,4 although these improvements are at best mild-to-moderate on a group level (effect sizes ranging from 0.06 to 0.23 standardized mean difference). 3 The high variability in individual response to cognitive training (i.e. substantial improvements in some patients and no improvement or even decline in others) limits the overall group effect of restorative interventions. 5 This might hamper our enthusiasm for such interventions, obscuring the fact that for a subset of patients these interventions may be highly beneficial. This introduces a challenge to identify, beforehand, which patients will benefit from such interventions and which patients will not.
So far, only a few exploratory studies have hinted towards specific patient characteristics being beneficial for cognitive training, that is having a high likelihood of successful treatment response. For example, less grey and white matter atrophy, a relapsing-remitting disease course, and higher processing speed were predictive of better response to training, 6,7 as were different profiles of default-mode network connectivity. 8 In this study, we investigated the effect of a restorative computerized cognitive training programme (attention training) on cognition in a large sample of PwMS. Next, we compared baseline demographic, clinical and magnetic resonance imaging (MRI) characteristics between responders and non-responders to identify clinical and MRI-based characteristics of response (and non-response).

Materials and methods
This study was approved by the Medical Ethical Committee of the VUmc. Informed consent was obtained prior to participation.

Participants
Eighty-two PwMS were included. Inclusion criteria were: MS diagnosis according to the McDonald criteria, 9 18-68 years of age, and ability to safely undergo an MRI examination. Patients were screened for motor and visual skills to ensure cognitive training participation ( Figure 1). Exclusion criteria included drug abuse, neurological and psychiatric diseases, prior cognitive training participation, and relapses or steroid use 4 weeks prior to examination. To assess disease severity, a validated Expanded Disability Status Scale (EDSS)-based questionnaire was used. 10 Patients underwent visual and motor screening to ensure intervention participation. Patients were randomized (by means of computer-generated tables) into an intervention group (n = 58) or a waiting-list control group (n = 24). Treatment allocation was not concealed, and there was no blinding. Twenty-one age-, sex-and education-matched healthy controls (HCs) were included at baseline.

Intervention
The intervention consisted of the C-Car computer programme, previously used in the field of neurooncology. 11 C-Car combined with compensatory training has been effective in a similarly sized glioma patient sample. 11 Glioma patients and PwMS exhibit comparable, mostly subtle cognitive impairment. The focus on attention is based on the fact that attention is key for proper functioning in other cognitive domains such as memory and information processing speed. In MS, deficits in attention have been linked to deficits both in working memory and processing speed, 12 making C-Car an interesting programme for this study.
C-Car simulates driving a car, with tasks designed to train sustained, selective, alternating and divided attention. C-Car simulates driving a car while presenting information processing tasks. Tasks include forming words out of two consecutive road signs (which each present two letters), counting the number of letters in a word, performing basic arithmetic operations (addition and subtraction), and ranking words in alphabetical order ( Figure 2). With increasing difficulty, distractions are added: distracting noise to ignore, and a moving pointer of the petrol gauge to which attention should also be paid. The programme is adaptive; patients practice at their own level, and difficulty is increased throughout the sessions (e.g. faster stimulus rate and longer exercise duration, addition of aforementioned distractions). A score of at least 90% is needed to progress to the next difficulty level. Patients were provided with a laptop, and were required to train for 7 weeks (once a week, 45 minutes per session) at home. Researchers kept weekly contact with patients to ensure compliance (defined as the total time spent training being 75% or more).

Neuropsychological assessment and questionnaires
All patients underwent neuropsychological assessment at baseline (T 0 ), post-intervention (T 1 ) and 3 months follow-up (T 2 ). HC underwent neuropsychological assessment only at baseline (T 0 ). The following six cognitive domains, relevant to MS-specific cognitive decline, 4

Defining response
To differentiate responders from non-responders, a reliable change index (RCI) from T 0 to T 1 was calculated for each cognitive test score. 24 To correct for practice effects, scores of the waiting-list control group were used to calculate the standard error of difference. If patients reached a post-intervention change in test score which corresponded to an RCI threshold of 1.64 (90% confidence interval), a reliable improvement (>+1.64) or decline (<−1.64) was designated. Responders were defined as scoring above the RCI threshold on at least two out of six (33%) of the cognitive domains measured (see above), on at least one test per domain.

MRI protocol
All patients underwent brain MRI scanning at baseline (T 0 ) and post-intervention (T 1 ). HCs underwent MRI scanning only at baseline (T 0 ). All subjects were scanned on a 1.5 T whole-body magnetic resonance system (Siemens Magnetom Avanto Syngo, Erlangen, Germany), using an eight-channel phased-array head coil). The details of the MRI protocol are described in the Supplementary Material.

Grey matter, white matter and lesion volumes
White matter lesions were automatically segmented on the PD/T2 images using k-nearest neighbour classification with tissue-type priors, 25 which was also used to compute lesion volume. The lesion segmentations were visually inspected and manually corrected where necessary. Subsequently, white matter lesion masks were registered to the 3D T1-weighted images to enable lesion filling. 26 Whole-brain, grey and white matter volumes were calculated on the lesionfilled images using SIENAX, following previously published pipelines 27 and deep grey matter volumes were obtained using FIRST.

Diffusion-weighted imaging processing
Diffusion tensor imaging (DTI) data were pre-processed using motion and eddy current correction on images and gradient vectors followed by diffusion tensor fitting (in FMRIB Software Library FSL). To assess white matter integrity, fractional anisotropy (FA) maps were computed and non-linearly registered to the FMRIB58_ FA brain. Next, FA maps were averaged across subjects and skeletonized to obtain the main white matter tracts common to the entire sample using the standard Tract-Based Spatial Statistics pipeline (part of FSL).
To obtain individual measures of whole-brain white matter integrity damage, the severity and extent of white matter damage was quantified based on FA (see Supplementary Material).

Resting-state functional connectivity
To define which resting-state network each region belonged to, the Yeo resting-state network atlas 28 was overlaid on the Brainnetome atlas, 29

Statistical analyses
Statistical analyses were performed in SPSS, version 26. Independent samples t-tests and chi-square tests were used to assess baseline group differences, and nonparametric tests were used for variables that were non-normally distributed. To assess the effects of the intervention, linear mixed models were used with the intercept as a random factor, and group (intervention vs. waiting-list) and time (T 0 vs. T 1 vs. T 2 ) as fixed factors. After the intervention group was divided into responders and non-responders, a multivariate general linear model was used to assess baseline differences in (1) demographic variables, (2) cognitive test scores, (3) structural MRI outcomes and (4) functional MRI outcomes. Finally, a backward logistic regression was performed, with responder/nonresponder classification as dependent outcome, to identify the strongest independent baseline predictors of response. The Bonferroni post hoc test was used to examine differences between groups. Variables were

Baseline characteristics
The  a more passive coping style on average compared with HC (see Table 2). Compared with HCs, PwMS had significantly less grey matter (GM), white matter (WM) and deep grey matter (DGM) volume, as well as significantly higher FC between DMN-DAN (p = 0.012), DMN-VAN (p = 0.011) and DMN-FPN (p = 0.03) ( Table 1).
PwMS in the intervention group and in the waitinglist control condition did not differ in age, sex, educational level, cognitive performance and self-perceived cognition and levels of fatigue, anxiety and depression at baseline. Imaging measures were not different between the two groups (Tables 1-3).

Effects of C-Car
The pre-and post-intervention differences investigated on a group level on short-and long-term followup neuropsychological testing are presented here.
Immediate effects. A significant group × time effect was found for working memory (Letter-Number-Sequencing) between T 0 and T 1 (Figure 4).  (Table 2), nor for measures of white matter damage and FC (Table 3).

Predicting response
The backward logistic regression identified two independent predictors for response: lower FC between Table 3. Microstructural white matter changes and functional connectivity in the intervention group, waiting-list group, and healthy controls.

Discussion
The main aim of this study was to gain insight into the effectiveness of computerized attention training in PwMS and into the clinical and MRI-based characteristics that are associated with successful cognitive training response. Therefore, we performed the conventional analysis of pre-post effects of C-Car on a group level, which yielded results and effect sizes similar to those found in previous cognitive training studies. 3 Unexpectedly, we saw no significant postintervention improvement in attention. However, patients in the intervention group showed a modest immediate improvement in working memory (standardized mean difference = 0.37, p = 0.038) with no sustained effects on cognitive performance at 3 months follow-up, which is in line with previous results 5, 31 and might indicate the need for a longer duration of cognitive training programmes. Interestingly, we observed a long-term effect for fatigue, with patients in the intervention group showing stable fatigue at the 3 month follow-up compared with the increased fatigue of the waiting-list control group, consistent with previous findings of the C-Car intervention in glioma patients. 11 Comparing responders and non-responders provided important information. Responders and non-responders did not differ in demographic and clinical variables, baseline cognitive performance, nor in the amount of atrophy and microstructural damage. In previous work, lower grey and white matter volume and a relapsing-remitting disease course were predictive of better response. 6 The fact that we were unable to reproduce these results might be explained by sample differences. Specifically, patients in our study had a higher cortical GM volume and shorter disease duration (13.52 vs. 21.6 years) compared with patients in the study of Fuchs et al. 6 Interestingly, regarding fMRI findings, non-responders exhibited higher FC between the DMN and attention networks compared with responders and HC. More precisely, responders showed no differences in FC compared with HC, suggesting an intact connectivity ( Figure 5). This relationship between alterations in DMN connectivity and treatment response is in line with previous results 8 and may be explained by the anti-correlation between DMN and attention networks both in tasks and during rest. 32 The similarity of our results to those of previous studies with different  non-responders, the DMN insufficiently deactivates when needed, and as such shows an increased connectivity with attention networks. Indeed, in healthy individuals, an increased FC between the DMN and attention networks has been related to poorer attention. 30 In addition, alterations in DMN connectivity and network dynamics are rather common in PwMS and relate to cognitive impairment. 33,34 As such, it may well be that an intact FC of the DMN is a prerequisite for successful cognitive training response, regardless of the intervention used. In our multivariate prediction model, lower FC between DMN-VAN and between DMN-FPN (i.e. 'normal FC') were both identified as predictors of response. This indicates that the fewer deviations there are from HC-like FC, the higher the chance for successful cognitive training.
Consequently, it seems that the timing of cognitive training in PwMS is of utmost importance. One could argue that a mind-set shift from symptom management towards preventive intervention aimed at preserving cognition is needed (i.e. enhancement of network functioning rather than restoring it, since the latter might be impossible).
Our study is not without limitations. The use of a waiting-list control group is not as optimal as an active control condition. 35 Also, cognitive impairment was not an inclusion criterion. As a result, the group is heterogeneous in terms of cognitive performance at baseline. That being said, 57.3% of all PwMS were impaired (z < −2.0SD) at baseline on at least one test. Moreover, although the definition of cognitive decline is well-established in the literature, the definition for response is less clear. We thus decided to rely on a conservative statistical approach (reliable change index). Another limitation of our RCI approach is that multiple tests were included in the domains of attention and working memory. Response in at least one test was calculated as response in the cognitive domain, hence making response in working memory and attention perhaps slightly more likely.
To conclude, our results demonstrated a mild-tomoderate overall short-term working memory effect of a computerized attention training for PwMS. Despite the lack of significant improvement in attention on a group level, we demonstrated that by investigating individual responses to treatment almost 40% of PwMS do improve after training, an effect that would have gone unnoticed in group-level statistics. Response seems to depend on a window of opportunity defined by an intact FC between the DMN and attention networks, allowing the brain to be receptive for the effects of cognitive training.
Given the heterogeneity of MS progression, disease course and observed differential response to cognitive training, 5 it is evident that future studies in the field now need to start exploring individualized (selection) approaches to maximize the effectiveness of cognitive training programmes. 6