It is believed that one person dies from a stroke every 4 min (Benjamin et al., 2017). For stroke survivors, cognitive impairment is one of the common dysfunctions that results in a poorer quality of life and is one of the main reasons preventing patients from returning to family and society (Hochstenbach et al., 2001; Van Der Flier et al., 2018). It has been reported that the most common cognitive symptoms following stroke are memory loss (90%), attention deficit (82%), and executive dysfunction (75%) (Nakajima, 2006). Working memory (WM), an important part of cognitive function, refers to a memory system with limited ability to process and store information temporarily (Baddeley, 2003). It is now generally believed that there exist age-related changes in WM capacity in the normal population: verbal WM peaks at ~age 26 and visual-spatial WM peaks at ~age 18, and normal older adults suffer from WM capacity decreases (Hale et al., 2011; Swanson, 2017). WM is vital for tasks that involve the goal-oriented use of immediate memory, the storage and manipulation of recently assimilated information, and the conversion and scheduling of task priorities in multitasking situations (Redick and Lindsey, 2013; Kumar et al., 2016). The incidence of WM impairment among post-stroke patients is reported to be as high as 87.6% (Jaillard et al., 2009). WM deficits remain prominent in the chronic stage of stroke (Kant et al., 2014). Lesion studies have revealed that the fronto-parietal cortex is related to WM deficits in patients with stroke (Baldo and Dronkers, 2006; van Asselen et al., 2006). A meta-analysis and systematic review study suggested that patients with stroke showed decrements of moderate magnitude in both low-load and high-load WM tasks and in all subsystems of WM (including central executive, episodic buffer, visual-spatial sketchpad, and phonological loop) (Baddeley, 2012; Lugtmeijer et al., 2021). Deficits in the WM system may affect executive function along with shifting and inhibition, episodic memory formation and retrieval, general intelligence, and syntactic processing (Conway et al., 2003; Bergmann et al., 2013; Friedman and Miyake, 2017; Alatorre-Cruz et al., 2019).

While the conventional concept was that the WM capacity of adults was constant, subsequent studies have found that WM capacity could be improved through training. WM training (WMT) is a cognitive training method that focuses on WM tasks operated on a computer in an adaptive way. The difficulty of the training task can be adjusted to maximize the WM capacity of the trainee. The methods of WMT can be divided into WM span tasks (e.g., number forward and backward, spatial forward and backward), updating tasks (e.g., n-back), and complex WM tasks (e.g., verbal WM tasks combined with spatial WM tasks) according to the training tasks used. Studies have shown that WMT could significantly improve the WM ability of healthy subjects (Olesen et al., 2004), as well as the WM ability of acquired brain injury patients with WM disorder (Richter et al., 2018). In addition, it was suggested that the efficacy of WMT could be transferred to other untrained cognitive functions, such as attention (Jaeggi et al., 2008), fluid intelligence (Westerberg et al., 2007), reading (Loosli et al., 2011), inhibition (Holmes et al., 2009), and mathematics (Kelly et al., 2006).

It has been suggested that WMT is an effective cognitive rehabilitation strategy in stroke patients for improving individual WM capacity, as well as the performance of untrained WM tasks and tasks that show far-transfer effect (Peers et al., 2020; Eschweiler et al., 2021; Nikravesh et al., 2021). However, the neural effects following WMT in stroke patients remain largely unclear and have been poorly studied. Using the blood oxygenation level dependent (BOLD) technique, functional magnetic resonance imaging (fMRI) can accurately detect the brain regions that are closely related to individual WM, as well as the brain regions that show functional changes after WMT. Thus, the present study aimed to explore the behavior and neural effects of WMT on chronic stroke patients using neuropsychological scales and task- and resting-state fMRI. We hypothesized that patients with stroke in the present study might have improved WM capacity and altered brain activation and functional connectivity after the WMT.

The neuropsychological results of a verbal n-back test that we used to examine the near-transfer effect of WMT revealed performance improvements in both 1-back and 2-back conditions. Patients in the WMT group made fewer mistakes and reacted quicker in an untrained verbal WM task 4 weeks after WMT. This kind of near-transfer effect was also reported in previous WMT studies (Vermeij et al., 2015; Hardy et al., 2021). The RSPM test used to assess the far-transfer effect of WMT indicated that training improved fluid intelligence (a non-verbal reasoning ability) in ischemic stroke patients in the WMT group. This finding was in agreement with previous studies that reported improved fluid intelligence after WMT (Jaeggi et al., 2010; Rudebeck et al., 2012). This improvement in fluid intelligence may be attributed to the training tasks that require the engagement of reasoning and executive processing, which are closely related to fluid intelligence. Our results also showed improvement in attention ability 4 weeks after the WMT in the WMT group. WM and attention are closely associated with each other; some previous studies have observed far-transfer effects on the tests of attention because tasks of WMT included regulation of attention and reduction of interference (Richmond et al., 2011; Brehmer et al., 2012).

Before training, neuroimaging results in the present study showed that verbal WM tasks evoked functional activation in a WM network, mainly including bilateral frontal-parietal and cerebellum regions, specifically, the bilateral MFG, IPL, and CPL. Our results were similar to previous fMRI studies that specifically investigated the brain areas involved in verbal WM (Paulesu et al., 1993; Chai et al., 2018), and were in keeping with previous studies that related specific lesion locations to WM performance in stroke patients (van Asselen et al., 2006; Baier et al., 2014).

Training can generate changes in brain functional activation modes. According to previous literature reports, training-induced brain activation modes can be grouped into the following four categories. First, is that the same brain regions are activated before and after training, but the activation intensity is increased in these regions after training (Westerberg and Klingberg, 2007). This “increase mode” may reflect an extensive recruitment of neural structures involved in the processing or the existing cortical neurons are more strongly responding to the task. Second, is that the same brain regions are activated at baseline and after training, but the activation intensity is decreased in these brain regions after intervention (Schneiders et al., 2011). This “decreased mode” may indicate the improved neural efficiency as a consequence of training, leaving more neurons for other task processing. The third mode is a combination of the above two modes, that is, the same brain regions are activated before and after training, but the activation is decreased in some of these brain regions while it is increased in other areas (Dahlin et al., 2008). This “redistribution mode” may suggest that training can lead to a quantitative shift in brain functional activation. The final mode is different from the other three modes. It is characterized by the activation of additional brain regions after training as compared to baseline (Opitz et al., 2014). This “network reorganization mode” may manifest that new neural routes are involved in the processing of tasks. After training, a similar brain network was activated during tasks of verbal WM in the WMT group, but WMT resulted in increased activation in the left MOG and MFG. This WMT-induced brain activation conforms to the “increase mode” described above. The increased activation in left MOG indicated increased compensatory ability to the impaired frontal-parietal regions and could be interpreted as more recruitment of neurons in the MOG into the WM network due to repetitive training. The increased activation in the left MFG might be attributed to enhanced neuronal responsiveness or the increased number of neurons within the left MFG that could result in an increase in regional fMRI signal underlying enhanced signal observed in the left MFG after intervention.

It was reported that healthy subjects maintained their ability to increase functional activation in the WM network with each increase in WM load (McAllister et al., 2001; Chen et al., 2012). We found that stroke patients in the current study were impaired in their abilities to elevate functional activation in the WM network as a response to the increasing WM load (2-back>1-back). This phenomenon suggested that with task difficulty increasing, no additional neurons within the WM network could be recruited to play compensatory roles for the completion of more difficult tasks. However, in the WMT group after training, increased activation was observed in the left CAL, right CPL, and MFG with increasing task difficulty. These results revealed that patients displayed improved abilities to increase activation in their WM network in response to the increasing WM load, and potentially indicated that WMT resulted in the left MFG and bilateral cerebellum regions being recruited to complete the challenging tasks.

After WMT, functional connectivity between the left MFG and left IPL, and between bilateral IPL and right CPL was increased. The enhancement of functional connectivity between these brain regions occurred within the WM network. In other words, WMT increased the resting-state functional connectivity within the WM network. It has been reported that the intensity of intrinsic resting-state activity in MFG and IPL is positively correlated with the intensity of functional activation in task-state activity (Zou et al., 2013), and, in general, the MFG and IPL are significantly activated in WM tasks. That is to say, the strength of the internal resting activity of MFG and IPL may be closely related to WM capacity. Deserno et al. (2012) reported that decreased functional connectivity between the MFG and IPL was closely related to the decline of WM capacity in patients with schizophrenia. In patients with acute stroke, it was suggested that impairment of WM was associated with damage to frontal-parietal regions (Martin et al., 2021). The MFG has been recognized to be crucial for maintaining information and supporting executive function during WM, especially high-load WM tasks (Volle et al., 2008; Bokde et al., 2010). The IPL was believed to play a critical role in maintaining phonological representations during phonological WM (Yue and Martin, 2022). It has been demonstrated that the cerebellum was included in the WM network and played a relevant role in subvocal rehearsal (Pleger and Timmann, 2018). It was also reported that cerebellum regions were responsible for executive control and for preventing irrelevant information from entering the WM network (Baier et al., 2014). However, the specific contribution of the cerebellum to the various processes involved in WM remains largely indistinct. As for why WMT led to resting-state functional connectivity changes within the WM network, we hypothesized that functional connectivity reflected past use, and those brain regions that were frequently used and activated simultaneously in the past have stronger functional connectivity. Thus, we suggested that the enhancement of resting-state functional connectivity within the WM network might be due to frequent and repetitive co-activation of these brain regions during WMT.

Several potential limitations involved in the present study should be considered. First, we did not assess the duration of training effects on behavior and brain as we did not set up a follow-up design. Furthermore, it is still not clear whether functional plasticity after WMT is based on structural plasticity, exploring structural changes after WMT in patients with stroke is therefore strongly recommended for subsequent studies. In addition, although comparisons of sample characteristics at baseline exhibited no statistical difference, the WMT group had larger mean lesion volumes and had more patients with left lesions than the CG, which might influence the outcomes more or less. Moreover, exploring the neural basis of the far-transfer effect following WMT in the future is desirable. Finally, future studies should compare the results of verbal WM and visual–spatial WM separately, as we only evaluated verbal WM during fMRI after WMT.

Our study supported that WMT potentially improved WM capacity in ischemic stroke patients during the chronic stage, and that the training effects might transfer to fluid intelligence and attention ability. Our results also demonstrated that repeated WMT potentially increased brain activation and resting-state functional connectivity within the WM network in patients with ischemic stroke. These findings provided robust evidence supporting WMT as an effective intervention to enhance cognitive rehabilitation and also shed light on the functional neuroplasticity mechanism of WMT on cognitive recovery after ischemic stroke.

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding authors.

The studies involving humans were approved by the Medical Ethics Committee of the CRRC. The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

ZC designed the study, performed the experiments, analyzed the data, and wrote the manuscript. XY did the financial support, review, and final approval of the article to be published. Both authors read and approved the final manuscript.