Our initial database search revealed a total number of 6243 citations. We removed 2495 duplicate citations, and 3653 additional citations were excluded through title and abstract screening. We retrieved 95 articles for full-text screening, from which 29 studies were excluded due to the following reasons: animal sample size, review article, duplicated articles from one same study, in vitro study, and reporting one single measure in one single study group. An additional number of 5 articles were removed during the data extraction because of insufficient data despite author contact (29, 30), reporting BNGF instead of BDNF (31), unclear data from a big cohort study (Framingham (3)), reporting BDNF in association with other biomarker instead of pure measures (32). Finally, a total number of 62 articles were entered into our systematic review. We could not enter 16 articles to any meta-analysis group due to the following reasons; providing delta BDNF instead of exact measures (33), reporting BDNF as a proportion to the control group (34), full text was not found, although the abstract was sufficient to extract the descriptive data (35), unsatisfactory data and irresponsive author (missing SD) (36), single study group with no comparison or specific grouping which did not fall into our meta-analysis categorizations (37–48).
Serum BDNF levels in PwS vs. HC
Seventeen studies assessed the differences between serum BDNF levels in PwS (n= 1191) and HC (n= 1856) (16, 19, 49–63). PwS had significantly lower serum BDNF levels than HC (SMD [95%CI] = -1.02 [-1.47 to -0.57], p-value< 0.001, I2= 96%, p-value< 0.001) (Figure 3). No publication bias was detected (Figure 4). Sensitivity analysis showed that Align et al. 2019 and Chan et al. 2015 were influential (Supplementary Figure 1). After excluding these studies, the meta-analysis of the remaining fifteen studies showed significantly lower serum BDNF levels in PwS than HC as well (-0.92 [-1.35 to -0.50], p-value< 0.001, I2= 96%, p-value< 0.001) (Results are not shown). To further assess sources of heterogeneity, we performed meta-regression. The age partially explained the high heterogeneity (correlation coefficient= -0.11, R2= 62.81%, p-value= 0.000).
Longitudinal investigation within stroke population
To address the pattern of serum BDNF level through the time of stroke onset, 23 studies longitudinally assessed their samples. A total number of 3678 patients were included in 39 observations. Time points for sample collection varied extremely between the studies. After the precise data extraction for the time of sample collection in regards to the stroke onset, we grouped the observation in 5 general groups. 1) Observations which compared baseline measures to the first day of stroke, baseline vs. day 1 or acute that covered 4 observations and 202 participants (37, 45, 64, 65); 2) Observations comparing the baseline measures to the level of BDNF within the first week of stroke onset (mostly 3-7 days) baseline vs. week1 or sub-acute, consisting of 14 observations and 1366 patients (37, 45, 48, 60, 62, 64, 66–73); 3) Observations investigating the difference between baseline levels of BDNF and day 30 or above, baseline vs. over 1 month or chronic, that included 11 observations and 1093 participants (19, 45, 60, 62, 70–72, 74–76); 4) Comparison of the first day of stroke to the first week after the onset, Day 1 vs. Week 1 consisting 3 observations and 162 participants (37, 45, 64); 5) Lastly the 7 observations investigating the BDNF pattern, in 855 participants, within the first week of stroke and over the one months of affecting the patients (45, 60, 62, 70–72, 77).
Based on the high heterogeneity level detected with I2 in every group, over 74% value, we conducted a random effect model analysis separately for each above mentioned observational group. No significant difference were addressed in the SMD in any of the included groups. No publication bias was found either (Supplementary figures 2-11).
Effect of depression on serum BDNF levels in PwS
Six trials with a total of 411 patients with and 245 patients without depression compared levels of serum BDNF between these two groups (49, 73, 78–81). All the studies followed Diagnostic and Statistical Manual of Mental Disorders 4th version (DSM-IV) criteria to assess their sample for depression. Hamilton Depression Rating scale (HDRS) was used by four studies (49, 73, 80, 81). While every included article assessed PSD, only two of them addressed the depression background in their patients prior to the stroke (79, 80). Patients with depression had significantly lower levels of BDNF than the participant in the non-depressed group (SMD [95%CI] = -0.60 [-1.10 to -0.10], p-value< 0.001, I2= 88%, p-value< 0.001) (Figure 5). No publication bias was detected. Due to the low number of studies, we were not able to conduct meta-regression or sensitivity analysis. (Figure 6).
Effect of physical training on serum BDNF levels in PwS
Sixteen studies with 738 patients provided original data on BDNF concentration before and after applying a specific physical training protocol. The training modality included regular physiotherapy sessions (82–84), aerobic exercises (85–87), high-intensity interval training (HIIT) (88), and proprioceptive neuromuscular facilitation exercise (PNF) (89) (Table 3). In addition, the number of sessions varied among the included studies starting from only one session (85) to as many as 56 sessions (19). Also the duration of sessions ranged from 12 minutes (90) to 120 minutes (82). 13 observational groups were included in the immediate analysis and eleven observational groups were in the delayed analysis.
Sample collection immediately after physical training
Analysis for the immediate group showed a positive effect of physical training in general on BDNF level immediately after the intervention with (SMD [95%CI] = 0.49 [0.09 to 0.88]), p-value = 0.02, I2= 85%, p-value <0.001) (Figure 7). Within the immediate group, we did a subgroup analysis on different training modalities. Eight observations that performed any sort of exercise fell into our exercise subgroup, and five observations applying regular rehabilitation or physiotherapy were categorized as rehabilitation subgroup. The subgroup analysis showed that only in the exercise group BDNF levels significantly increased immediately after physical training (SMD [95%CI] = 0.75 [0.25 to 1.25], p-value = 0.003). No subgroup differences were detected. Sensitivity analysis showed that the study of Anjum et al., 2020 was influential (Supplementary Figure 12). After omitting this record, the overall effect size did not remain significant (SMD [95%CI] = 0.34 [-0.03 to 0.71], I2= 82%) (Supplementary Figure 13). Publication bias was not observed between the included studies (Figure 8).
Sample collection with a delay after physical training
The analysis of the delayed group consisting of eleven observations showed no significant effect of the intervention at this delayed phase of sample collection (SMD [95%CI] = 0.02 [-0.43 to 0.47], I2= 83%) (Figure 9). Similarly, BDNF levels did not change significantly in none of the subgroups. Leave-one-out analysis showed that after omission of Wang et al., 2021 study (19), the I2 index reduced to 68% while the overall effect size remained not significant (Supplementary Figure 14 and 15). No publication bias in neither groups was noted (Figure 10).
Repeated transcranial magnetic stimulation (rTMS)
Four studies measured the BDNF levels in 105 PwS after receiving rTMS; Lu 2015 used plasma samples while the other three used serum (62, 82, 84, 91). The overall meta-analysis of the studies revealed no difference in the BDNF levels between the rTMS and the sham stimulation group (SMD [95%CI] = 0.00 [-0.27 to 0.27]). No significant heterogeneity was found (I2= 0%, p-value= 0.48) (Figure 11). Meta bias in the included studies is presented visually in the funnel plot (Figure 12), with Begg’s and Egger’s test of 0.0603 and 0.081, respectively.