We conducted a search of databases such as PubMed, EMBASE, and the Cochrane Library using a specific search strategy for each database, covering literature from inception up to December 31, 2022. Additionally, we conducted a review of the references of each included study to ensure that no relevant papers were missed during the search. The full search strategy is available in Supplementary Table S1.

(1) Participants: older adult people aged ≥60 years, as defined by the World Health Organization (13, 14).

(2) Inventions: any form of Tai Chi exercise.

(3) Controls: receiving either exercise that is stretching or other low-level exercises, usual care, or wellness education.

(4) Outcomes: the number of fallers (participants who experienced at least one fall), average number of falls per person, and balance assessment indicators including the timed up and go (TUG) test, functional reach test (FRT), single leg balance test (SLB), Berg balance scale (BBS), short physical performance battery (SPPB) score, fall efficacy scale (FES), and gait speed assessment. Among these, FRT is a simple, practical, and widely used clinical tool to assess an individual’s balance capability and risk of falling. In this test, the maximum distance an individual can reach forward while maintaining a fixed support base in a standing position is measured.

(5) Study: only include RCTs in English.

After eliminating duplicate studies using Endnote software, two reviewers (W.C. and M.L.) independently selected eligible studies and extracted relevant data. In case of any discrepancies, the reviewers reached a consensus through discussion and negotiation. The collected data included the author’s name, year of publication, sample size, older adult population, health status, age, Tai Chi style, frequency, total exercise time, and follow-up time and outcomes.

The assessment parameters for the evaluation of study quality included random sequence generation, allocation concealment, blinding of participants and assessors, handling of incomplete outcome data, selective reporting, and other biases. Each item was assessed as having “low risk,” “unclear risk,” or “high risk.” Two reviewers (W.C. and M.L.) independently assessed the data using the Cochrane Risk of Bias tool (15) and resolved any discrepancies through mutual discussion based on justifications.

The effect size of binary variables was represented by the risk ratio (RR) with a 95% confidence interval (CI), and the effect size of continuous variables was represented by the mean difference (MD) with a 95% CI, evaluated using a random-effects model. The I² statistic was used to evaluate the heterogeneity among studies, where I² < 50% indicated low heterogeneity and I² ≥ 50% indicated significant heterogeneity. Subgroup analysis was performed on exposure time, Tai Chi style, weekly frequency, risk of fall, and follow-up time.

Sensitivity analysis was performed by using the one-by-one exclusion method and excluding studies with a small number of participants to test the stability of the outcomes. For each result that included 10 or more original trials, we used funnel plots to test for publication bias, and assessed the symmetry of the funnel plot using Egger’s test and Begg’s test (16, 17). If the funnel plot was asymmetric, we used the trim-and-fill method for adjustment (18). A two-tailed p < 0.05 was considered statistically significant. All analyses were conducted using RevMan 5.3 and Stata 14.0 software.

A total of 1948 articles were initially searched, of which 837 duplicates and 25 articles without titles were excluded automatically. An additional 828 articles were excluded by reading titles and abstracts. From the remaining 258 articles, 88 were conference papers, other reports, and reviews. Further exclusions were made for 33 non-RCT, 85 articles with incomplete data or unrelated research content, and 28 articles for other reasons. In the end, 24 articles (with one or more outcomes) were included for analysis (7, 8, 19–40) (Figure 1).

Among the 24 articles, 6 reported participants with a history of falls (7, 24, 27, 29, 30, 36), and 5 reported patients with a history of neurological diseases such as stroke or Parkinson’s disease (8, 21, 35, 37, 38). The Tai Chi styles used in the interventions were mainly Sun-style (19, 25, 33, 34, 36, 39) or Yang-style (7, 21, 23, 24, 26, 32, 35, 37, 40), with one article using Chen-style (30) and others not specifying. Exercise frequency ranged from one to three times a week, with one article reporting five times a week (23). Total exercise time was mainly within 72 h, with one article reporting 96 h (19) and another reporting 144 h (40). Follow-up periods were ≤ 3 months in 6 articles (25, 30, 32–35), 3–6 months in 9 articles (8, 20, 23, 26–28, 36, 37, 39), 6–12 months in 7 articles (7, 19, 21, 22, 29, 38, 40), and > 12 months in 2 articles (24, 31) (Table 1).

Due to the appropriate methods used in random sequence generation and allocation concealment, 22 (8, 19–31) or 17 (8, 19, 20, 23–28, 31, 33, 35–40) articles were considered low risk, respectively. In addition, only 7 articles (7, 8, 20, 26, 29, 36, 38) were considered low risk due to the difficulty in blinding participants, while 19 articles used blinding well in assessing outcomes (7, 8, 19–21, 23–29, 32, 34–37, 39, 40). In terms of incomplete outcome data and selective reporting, 21 (8, 19–21, 23–35, 37–40) and 22 (8, 19–27, 29–32) articles were considered low bias, respectively (Figure 2 and Supplementary Figure S1).

A total of 18 studies provided data on the number of fallers (7, 8, 19–24, 26–29, 31, 36–40), and meta-analysis found that Tai Chi exercise could significantly reduce the number of fallers (1,041/2766 VS. 1,321/2703; RR: 0.76, 95% CI: 0.71–0.82, I²: 25%, p < 0.001) (Figure 3A). Sensitivity analysis was performed by sequentially excluding each trial or removing studies with fewer than 100 participants, and the results showed that there was no significant change in statistical significance and heterogeneity. Subgroup analysis showed that Tai Chi could prevent falls in any population at risk, regardless of the overall duration of Tai Chi exercise, follow-up time, or Tai Chi style (Sun or Yang) (all p < 0.01). However, in terms of exercise frequency, exercising twice (7, 8, 19, 20, 27, 29, 36, 38) or ≥ 3 times (21–23, 26, 28, 31, 37, 40) per week showed more significant benefits than exercising once a week (RR [95% CI] were 0.78 [0.73, 0.84] and 0.67 [0.58, 0.79], respectively) (Table 2).

According to 15 articles providing the total number of falls, the Tai Chi group had a lower incidence of falls compared to the control group (1816/2539 VS. 2,681/2475, IRR, 0.66) (see Supplementary Table S2) (7, 8, 19–21, 23, 24, 26–29, 36–39). Among them, 9 articles provided mean and standard deviation information (7, 20, 21, 23, 24, 27, 28, 38, 39). Tai Chi significantly reduced the number of falls per person (MD [95%CI]: −0.26 [−0.39, −0.13], I² = 61%, p < 0.001) (Figure 3B). Excluding the Li et al. study effectively reduced heterogeneity (I² = 49%), but did not significantly affect the results (−0.12 [−0.17, −0.06], p < 0.001) (7, 20, 21, 23, 24, 28, 38, 39). Moreover, studies excluding fewer participants did not significantly change the results (−0.12 [−0.18, −0.07], p < 0.001) (7, 20, 23, 24, 38, 39). Subgroup analysis showed that Tai Chi is more effective in reducing falls per person with longer exercise time (p = 0.15, 0.02, and < 0.01, respectively), and that Yang-style Tai Chi is more effective than Sun-style Tai Chi (p < 0.01 and = 0.09, respectively). Furthermore, regardless of the exercise follow-up duration (≤6 months, 6 ~ 12 months, or > 12 months), Tai Chi exercise was found to significantly reduce the number of falls per person (p = 0.01, <0.01, and < 0.01, respectively) (Supplementary Table S3). These findings can be used to design better Tai Chi interventions for fall prevention.

Meta-analysis of 12 studies showed that Tai Chi can significantly reduce TUG time (MD [95%CI]: −0.69 [−1.09, −0.29], p < 0.001) (Figure 4A) (8, 21, 22, 25–28, 30, 32–34, 36). Sensitivity analysis was conducted by sequentially deleting each trial and excluding studies with smaller sample sizes, and the significance and heterogeneity of the results remained unchanged. With increasing exercise time and frequency, Tai Chi was more effective in reducing TUG time (p = 0.93, 0.01, and <0.001; p = 0.76, 0.28, and <0.001, respectively). Yang-style Tai Chi was more effective than Sun-style Tai Chi (p < 0.01 and = 0.66, respectively). In addition, Tai Chi can shorten TUG time in older adult people with falls related diseases (p = 0.02) (Supplementary Table S4) (21, 38).

Meta-analysis of 9 studies providing FRT data has shown that the functional reach distance of practicing Tai Chi was significantly greater than that of the control group (MD [95%CI]: 2.69 [1.14, 4.24], p < 0.001) (Figure 4B) (8, 22, 25–27, 30, 32–34). Sensitivity analysis, conducted by sequentially excluding each study or studies with smaller sample sizes, did not significantly alter the results (all p < 0.01). Subgroup analysis revealed that Yang-style Tai Chi was more effective than Sun-style Tai Chi (p < 0.001 and 0.82, respectively), and practicing Tai Chi twice a week had a better effect (p < 0.001). Furthermore, Tai Chi was found to be particularly effective in improving FRT test scores in older adult individuals with a history of falls or fall-related diseases (p = 0.003 and < 0.001, respectively) (Supplementary Table S5).

Through a meta-analysis of other fall or balance ability tests, it was found that Tai Chi can effectively lengthen SLB test times (including with eyes open and closed) (MD [95% CI]: 9.63 [5.87, 13.40], p < 0.001) (25, 26, 30, 33, 34), increase BBS scores (MD [95% CI]: 1.80 [0.09, 3.51], p = 0.04) (21, 26, 29, 32), and improve gait speed (MD [95% CI]: 9.26 [1.00, 17.52], p = 0.03) (8, 25, 26, 34). In addition, Tai Chi did not have a significant effect on overall SPPB scores (27, 35, 37) or FES scores (24, 29, 34) (MD [95% CI]: −0.07 [−1.07, 0.93], 0.17 [−0.49, 0.83]; p = 0.89, 0.61; respectively) (Table 3).

Publication bias analyses were conducted on the results of more than 10 studies on fallers and TUG. It was found that there was a significant publication bias in the Fallers results (Begg’s test: 0.014 and Egger’s test: 0.003) (Figure 5A). The trim-and-fill method was used to adjust for publication bias, and the outcome still showed statistical differences (p < 0.001). In addition, there was no significant publication bias in the TUG results (Begg’s test: 0.092, Egger’s test: 0.315) (Figure 5B).