The Patient-Intervention-Comparison-Outcome-Study Design (PICOS) framework was used to organize the inclusion criteria.

(P) Studies recruited patients diagnosed with DoC by coma recovery scale-revised (CRS-R) or Glasgow coma scale (GCS). (I) Studies using tDCS, TMS, MNS, or any other type of NINT to investigate its effectiveness on level of consciousness in patients with DoC. (C) Studies adopting control conditions such as sham stimulation, no stimulation, or any active control intervention. (O) Studies adopted at least one of the clinical behavioral scales, neurophysiology evaluation, neuroimaging evaluation, or any other measures to assess level of consciousness. (S) Randomized controlled trials (RCTs) with parallel or cross-over design.

In addition to the above criteria, we only included studies reported in English. Review articles, conference abstracts, expert opinion papers, and editorials were excluded. Meanwhile, studies reporting on less than five patients or providing only one session intervention were excluded.

Electronic databases were searched from their inception through November 2022, including PubMed, Embase, Web of Science, Scopus, and Cochrane central register of controlled trials. Meanwhile, we searched the reference lists of included studies to identify further studies. The search was performed using the following keywords: (“disorders of consciousness” OR DoC OR Coma OR “Vegetative State” OR VS OR “unresponsive wakefulness syndrome” OR UWS or “minimally conscious state” OR MCS) AND (neuromodulation OR “non-invasive brain” OR “transcranial electrical current stimulation” OR TES OR “transcranial Direct Current Stimulation” OR tDCS OR “transcranial alternating current stimulation” OR tACS OR “transcranial random noise stimulation” OR tRNS OR “transcranial magnetic stimulation” OR TMS OR “theta burst stimulation” OR TBS OR “low-intensity focused ultrasound” OR LIFU OR “transcutaneous auricular vagus nerve stimulation” OR taVNS OR “Near-infrared laser stimulation” OR N-LT OR “focused shock wave therapy” OR F-SWT OR “median nerve stimulation” OR MNS). The detailed searching strategies were shown in Appendix S1.

After removing duplicates, two authors (B.B.Y, and J.Y.W.) independently screened all eligible articles and cross-checked the information. In a case where the eligibility of inclusion was conflicted, a third senior author (X.L.) was consulted to solve the dispute, and a final decision was made by consensus.

Two authors (Z.Y.L. and X.T.Z.) independently extracted data using a standardized form for each eligible study. This form includes information concerning general information (e.g. title, first author, year of publication), methodology (e.g. study design, duration of the study), demographics (e.g. age, gender, time from injury to intervention), interventions (e.g. type of intervention, stimulation parameter), outcomes (e.g. outcome measures, evaluation timepoint) and adverse events. Disagreements between the two authors were resolved by discussion with a third senior author (X.L.).

We assessed the risk of bias in all included studies using revised Cochrane risk-of-bias tool (ROB version 2.0) (Sterne et al., 2019). Aspects of randomization process, deviations from the intended interventions, missing outcome data, measurement of the outcomes, and selection of the reported result were evaluated. Bias arising from period and carryover effects were additionally evaluated for cross-over studies. Each domain was assessed as “low risks”, “some concerns”, or “high risks”. Two authors (Z.Y.L. and X.T.Z.) independently accomplished assessment and resolved any discrepancy through discussion with a third senior author (X.L.).

With the assistance of R statistical software, library “meta” was used to perform meta-analysis if at least two studies assessed one specific outcome. For outcomes based on continuous data obtained at the end of the intervention, we adopted mean difference (MD) with 95% confidence interval (CI) as effect size. For dichotomous outcomes obtained at the end of the intervention, we adopted risk ratio (RR) with 95% CI. For ordinal outcomes, if there was a cut-off point that could be obtained, we transformed the data into dichotomous data. Otherwise, it was calculated as continuous data.

Meta-analyses were calculated following the methods suggested by the Cochrane Review (Higgins et al., 2011). Combined design meta-analytic formula, using the methods suggested by Elbourne et al. (Curtin et al., 2002; Elbourne et al., 2002), were used to combine parallel and cross-over trial results. With no carry-over effects in cross-over studies, trial results were combined with parallel studies by the combined design meta-analytic formula or included in qualitative analysis, depending on whether the study reported the order of crossover and individual specific different time-point results. Otherwise, only data from the first phase in cross-over studies was included and combined with parallel studies.

Statistical heterogeneity of included trials was evaluated with I² statistic and between-study variance (τ²) (Higgins and Thompson, 2002). Studies with an I² of 0 to 24% were considered as low heterogeneity; I² of 25% to 49% as moderate heterogeneity; I² of 50%-74% as substantial heterogeneity and 75%-100% as high heterogeneity (Higgins et al., 2003). When I² was greater than 50%, it was assumed that there was considerable heterogeneity between studies, therefore random-effects model was applied. Sensitivity analyses were conducted by the “leave-one-out” method and omitting studies with high risk of bias.

To examine the differential effects of confounders, pre-planned subgroup analyses were conducted by certain parameters if data was available, including etiology of Doc (TBI, CVA or HIE), initial level of consciousness (Coma, VS/UWS or MCS), phase of DoC (Acute, Subacute or Chronic), and stimulation site (primary motor cortex, dorsolateral prefrontal cortex or others).

Moderator effects were examined by meta-regression using stimulation parameters as predictor variables. For tDCS, these parameters included number of total sessions and total stimulation time. For TMS, these parameters included frequency of stimulation, number of sessions, number of pulses per session, and total stimulation number of pulses.

Reporting biases were assessed by a contour-enhanced funnel plot (Peters et al., 2008). Based on the effect sizes and standard errors of included studies, the significance of any effect size could be calculated and relevant study could be distributed to special color regions representing different significance levels. An asymmetrical appearance of the plots represents the existence of bias. When bias was detected through funnel plot, we used a trim and fill algorithm to adjust (Duval and Tweedie, 2000). The adjusted results obtained by the algorithm could balance the bias in the overall results of studies that were unpublished due to insignificant effects. Adjusted results combined with primary results were used to determine whether the bias remarkably affected effect size estimation.

Process of study selection was summarized in Figure 1. A total of 5,956 studies were retrieved from the databases. After duplication elimination, 3,878 studies were obtained. Fifteen studies met the eligibility criteria and were finally included in the systematic review. Two studies were excluded from quantitative synthesis due to lack of sufficient data for obtaining effect sizes. A PDF document of the comprehensive search results including all the records was shown in Appendix S2.

Table 1 summarized characteristics of studies included in this review. A total of 15 randomized controlled studies with sample size ranging from 6 to 50 patients were included in this review. Regarding patients' initial level of consciousness, three studies included only patients with coma, four included patients with VS/UWS, three included patients with MCS, and the remaining five did not specify the type of DoC in their eligibility criteria.

As for the intervention of studies, seven studies used tDCS. Five studies used TMS. The remaining three studies used MNS. In addition, in terms of the design of studies, we included seven randomized cross-over controlled studies and eight randomized parallel controlled studies.

Seven tDCS studies enrolled 154 patients with VS/UWS or MCS. Clinical behavioral scales like CRS-R, Western Neurosensory Stimulation Profile (WNSSP), or neurophysiology evaluation like electroencephalogram (EEG), event-related potential (ERP) were evaluated as the outcome. Among them, Huang et al. (2017) and Thibaut et al. (2017) reported significant effects (p < 0.05) on CRS-R in patients with MCS. Both studies used a unilateral monopolar montage protocol and selected right supraorbital region as the reference cathode. It is worth noting that the former selected posterior parietal cortex (PPC) as the anodic stimulation site, whereas the latter chose the left dorsolateral prefrontal cortex (DLPFC). Cavinato et al. (2019) reported that only a proportion of patients with MCS had significant effects on WNSSP when the left DLPFC and the right deltoid muscle were selected as the anode and cathode respectively (p < 0.05). None of the remaining four studies reported any significant effects (Estraneo et al., 2017; Zhang et al., 2017; Martens et al., 2018; Wu et al., 2019).

Five TMS studies recruited 155 patients with VS/UWS or MCS. CRS-R was the only commonly used outcome. In addition, neurophysiological evaluation such as EEG, somatosensory evoked potential (SEP), or brainstem auditory evoked potential (BAEP) was also used to assess patients' level of consciousness. Cincotta et al. (2015) and He et al. (2018) applied TMS on left primary motor cortex (M1) and reported no significant effect on CRS-R and EEG. Zhang et al. (2021) investigated the effects of TMS with left DLPFC on CRSR, EEG, and BAEP in patients with DoC and reported a significant effect (p < 0.05). Similarly, Chen et al. (2022) and Fan et al. (2022) selected left DLPFC as stimulation target and showed significant effects (p < 0.05) of TMS on CRS-R either.

Three MNS studies enrolled 36 patients with coma. Cooper et al. (1999) and Nekkanti et al. (2016) both applied MNS on right median nerve and reported significant effects on GCS. However, Peri et al. (2001) used MNS on unilateral median nerve according to patients' injured brain hemisphere and reported no significant effect on GCS.

Results from assessment of bias using revised Cochrane risk-of-bias tool for parallel and cross-over studies were presented in Figures 2, 3. Only three studies were assessed as “low risks” (Peri et al., 2001; Chen et al., 2022; Fan et al., 2022). In addition, the majority of included studies (n = 7) were assessed as “some concerns” because of indistinct illustration of randomization process or other relatively rare reasons (Cooper et al., 1999; Huang et al., 2017; Thibaut et al., 2017; Zhang et al., 2017; He et al., 2018; Martens et al., 2018; Wu et al., 2019). Five studies were assessed as “high risks” on account of significant bias in at least one domain (Cincotta et al., 2015; Nekkanti et al., 2016; Estraneo et al., 2017; Cavinato et al., 2019; Zhang et al., 2021).

Thirteen studies were included in the meta-analysis. CRS-R was the only commonly used outcome in six tDCS studies and five TMS studies. Similarly, GCS was the only outcome that could be extracted from two MNS studies. Separate meta-analyses were conducted for tDCS, TMS, and MNS studies.

Meta-analysis of effectiveness of tDCS on CRS-R of six studies was presented in Figure 4. There was a small but significant effect size (MD 0.71 [95% CI 0.28, 1.13], p < 0.01) without significant heterogeneity (I² = 0.0%, p = 0.53). Furthermore, result was stable when we adopted sensitivity analysis (Table 2). The contour-enhanced funnel plot (Figure 5) with the trim and fill method did not show evidence of reporting bias (MD 0.68 [95% CI 0.26, 1.11], p < 0.01).

To examine the differential effects of confounders, subgroup analyses were conducted. As shown in Table 3, the subgroup analysis by the etiology of DoC revealed that among patients with TBI, tDCS showed a positive and significant effect size on CRS-R (MD 1.09 [95%CI 0.37, 1.82], p = 0.003), while patients with CVA had a positive but insignificant effect size (MD 0.53 [95%CI−0.10, 1.163], p = 0.10) and patients with HIE only showed a negative and insignificant effect size (MD−0.30 [95%CI−1.50, 0.91], p = 0.63).

Factors that showed a significant effect size favoring tDCS include initial level of consciousness among patients with MCS (MD 1.08 [95%CI 0.40, 1.77], p = 0.004), subacute phase of DoC (MD 0.97 [95%CI 0.13, 1.81], p = 0.02) and DLPFC as the stimulation target (MD 0.92 [95%CI 0.20, 1.64], p = 0.01).

As shown in Table 4, the meta-regression analysis showed that none of the between-study variables significantly predicted the effects of tDCS (number of total sessions: β = 0.01, p = 0.71; total stimulation time: β = 0.00, p = 0.71).

Meta-analysis of effectiveness of TMS on CRS-R of five studies was presented in Figure 6. There was a small but significant effect size (MD 1.59 [95% CI 1.01, 2.18], p < 0.01). Non-significant level of heterogeneity (I² = 0.0%, p = 0.71) was found. The result was further confirmed by using sensitivity analysis (Table 2). The contour-enhanced funnel plot (Figure 7) with the trim and fill method did not show evidence of reporting bias (MD 1.51 [95% CI 0.96, 2.06], p < 0.01).

To examine the differential effects of confounders, subgroup analyses were conducted. None of the four included TMS studies specified subjects from their etiology, initial level of consciousness, and duration of DoC. Individual patient data related to the above variables could not be extracted either. As a result, subgroup analysis was only conducted for stimulation site. As shown in Table 3, only patients who applied TMS on DLPFC showed a positive and significant effect size (MD 1.75 [95%CI 1.09, 2.40], p < 0.01), while patients who applied TMS on M1 had a small but insignificant effect size (MD 1.01 [95%CI−0.28, 2.30], p = 0.13).

As shown in Table 4, the meta-regression analysis showed that none of the between-study variables significantly predicted the effects of TMS (frequency of stimulation: β = −0.01, p = 0.88; number of sessions: β = 0.01, p = 0.63; number of pulses per session: β = 0.00, p = 0.74; total stimulation number of pulses: β = 0.00, p = 0.80).

Meta-analysis of effects of MNS on GCS was presented in Figure 8. Only two MNS studies were included in the meta-analysis of effects of MNS on GCS. There was a significant effect size in GCS (MD 3.20 [95%CI: 1.45, 4.96], p < 0.001) favoring the MNS group. Sensitivity analysis, reporting bias, subgroup analysis, and meta-regression were not conducted due to the limited number of studies.

The meta-analysis of the effect of tDCS on level of consciousness in patients with prolonged DoC indicated a positive, albeit, small significant effect size. Subgroup analyses revealed that only patients with TBI presented significant improvement in the level of consciousness compared to patients with CVA or HIE. Moreover, we also found that higher initial level of consciousness (MCS) and shorter duration of prolonged DoC (Subacute phase of DoC) may be associated with better clinical awaking effects.

Regarding the stimulation parameter, almost all studies adopt the same stimulation intensity and stimulation time per session. The results of meta-regression also showed a non-significant “dose-dependent” correlation between total stimulation duration and effectiveness. A possible explanation for this might be that the current result was based only on the short-term effects. Whether the benefit of tDCS in long-term effects improves with an increasing number of sessions remains to be discussed in future studies. As for the stimulation sites, five studies selected DLPFC as the anodic stimulation site and only one study selected PPC. As a result, only the effectiveness of tDCS applied on DLPFC could be confirmed. The effectiveness of tDCS applied on other sites remains to be explored.

Compared to previously published reviews (Zaninotto et al., 2019; Feng et al., 2020), our finding was consistent with that of Feng et al. who reported a positive effect of tDCS in patients with MCS. In addition to specific initial level of consciousness of patients, we found that etiology and phase of DoC could be significant factors for effectiveness of tDCS. This finding was also consistent with the current consensus that patients with MCS or TBI had a better prognosis compared to other diagnostic or aetiologic subtypes (Giacino et al., 2018a). Our results, while preliminary, suggested that the above characteristics of patients could contribute to the effectiveness of tDCS.

The meta-analysis of the effect of TMS on level of consciousness in patients with prolonged DoC indicated a small but significant effect size. Regarding the stimulation parameter, the result of meta-regression showed no linear relationship between stimulation frequency, stimulation duration, or number of stimulation pulses and effectiveness. One reason for this result might be that all included studies utilized high-frequency (5–20 Hz) TMS. Similar excitatory effect on the cortex was produced with the long-term potentiation induced by high-frequency stimulation (Pascual-Leone et al., 1994). On the other hand, the absence of “dose-dependent” correlation might also be attributed to the lack of long-term follow-up data. Whether the benefit of TMS in long-term effects improves with an increasing number of sessions or pulses remains to be discussed in future studies.

Subgroup analysis revealed that only patients who applied TMS on DLPFC presented significant improvement in the level of consciousness compared to patients who applied TMS on M1. Interestingly, this finding was partly consistent with that of Feng et al. A non-significant awaking effect was found by two studies that applied TMS on M1 (Feng et al., 2020). Compared to this previous review, the reason that caused the difference stemmed from three newly included TMS studies in our study. Given the results of subgroup analysis and meta-regression, the main factor for the opposite conclusions might be attributed to the different stimulation sites of TMS. As compared to M1, it seems that TMS has an awaking effect via DLPFC. A possible explanation for this hypothesis could be linked to the function of different cerebral networks. DLPFC, as a critical component of executive control network (ECN), plays a vital role in mediating environmental awareness and repairing the imbalance between the ECN and default mode network (DMN) (Seeley et al., 2007). Therefore, it can be assumed that stimulation of DLPFC could modulate internetwork connectivity between ECN and DMN via salience network and accelerate patients' transition from VS/UWS to MCS. However, this assumption needs to be verified further. Future work is required to determine whether DLPFC is the most optimal stimulation site for TMS.

The meta-analysis of the effect of MNS on level of consciousness in patients with coma indicated a small but significant effect size. Notably, the results should be interpreted cautiously, considering the high risk induced by the limited number of available studies. In contrast to previous studies, one systematic review reported qualitative results and expressed concerns about the effectiveness of MNS (Feller et al., 2021). Currently, the mechanism regarding the awaking effect of MNS remains unclear. One possible mechanism is that MNS enhances ARAS activity by stimulating the locus coeruleus and dorsal raphe nucleus, which represents the origins of the noradrenergic and serotonergic neurotransmitter systems, respectively (Kayama and Koyama, 1998). Whether MNS has an awaking effect in patients with prolonged DoC also lacks evidence from research. Nonetheless, the convenience and economics of MNS allow caregivers to provide beside therapy without the assistance of medical staff. As a result, the effectiveness of MNS on level of consciousness remains to be explored in future studies.

Our finding, while preliminary, further supported the validity of the cellular and circuit mechanism (Edlow et al., 2021). Both central neuromodulation applied on the DLPFC (e.g., tDCS and TMS) and peripheral neuromodulation applied on the median nerve were involved in the reorganization of dynamic interactions between multiple cerebral networks. Notably, as consciousness was dominated by complex cerebral networks, the stimulation of a single neural circuit might not extend to other neural networks. Therefore, compared with the single NINT commonly used in clinical research, whether the combination of multiple neuromodulation therapies can achieve better awaking effects by activating widespread functional connectivity between brain regions remains to be further investigated.

Our review cannot be ruled out with limitations. Firstly, although most studies applied assessments other than neurobehavioral evaluation as outcomes, it is difficult to combine these results into synthesis analysis because of their varying collection and analysis methods. Therefore, our meta-analyses were based only on CRS-R and GCS. Secondly, due to few studies reported follow-up results, our finding was only applied to short-term effects. Future studies need to further explore the effectiveness of NINT in long-term awaking effects. Thirdly, limited by the fact that included tDCS studies and TMS studies had only one common outcome, as well as the lack of direct comparison between tDCS and TMS, it is difficult to conduct a network meta-analysis. As a result, it is hard to draw a definite ranking list of the superiority of the two interventions. Finally, although several new studies of NINT such as taVNS, low-intensity focused stimulations (LIFUS), and focused shock wave therapy (F-SWT) have been published in recent years and both have reported encouraging results in awaking therapy (Hesse et al., 2016; Cain et al., 2021, 2022). Most of them are still case series and need further validation through more high-quality randomized controlled trials. Therefore, only tDCS, TMS, and MNS were included in our review.