Lithium carbonate as add-on therapy to radioiodine in the treatment on hyperthyroidism: a systematic review and meta-analysis

We followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement [23] to conduct this systematic review (SR) and meta-analysis (MA) for clinical trials whether randomized (RCT) or non-randomized. The study has no registered online protocol.

We included studies that included patients with hyperthyroidism, Graves’ disease, or toxic goiter, which uses LiCO3 with I131 as an intervention and RAI only as a comparative. The study design included clinical trials, whether randomized or non-randomized. Only English-written human-based studies that provide published accessible full text were included.

The primary outcome of the study was the cure rate of the patients, which is the number of euthyroid or hypothyroid patients after the treatment period.

The secondary outcome was the change in serum levels of total tetra-iodothyronine (T4) (ng/ml).

We searched the PubMed, Scopus, Cochrane Central Register of Controlled Trials (CENTRAL) and Web of Science Core Collection databases from database inception through 27 July 2019. The search employed all relevant index terms and keywords and did not utilize any filters. The complete search strategy for all databases is available in the supplementary file.

The study selection process was done in two phases: title and abstract screening and full-text screening. Two independent investigators screened each item, and another investigator solved the conflict.

We extracted the following data: the general features of the included studies, baseline characteristics, and outcomes of interest. General features included study design, country, timing, description of the intervention group, description of the control group, inclusion criteria, antithyroid drugs, duration of lithium administration, the reason for hyperthyroidism, and description of the intervention. Baseline characteristics included age (years), the onset of hyperthyroidism (months), gender (frequency), mean thyroid volume (ml), and mean serum TSH (mIU/I). The extracted outcomes were serum’s total T4 (ng/ml), and the number of euthyroid and hypothyroid patients (frequency).

Some data were reported as mean and SE, so we used Revman 5.3 software to convert SE to SD. Besides, we calculated the mean difference for Serum total T4 outcome by subtracting the mean of baseline from that of Post-treatment value. In addition, we obtained the SD of the change from baseline using the method described by Foolman 1992 [24] and Abrams 2005 [25] for calculating the SD of the change from baseline using a decided on correlation coefficient. The correlation coefficient was decided to be zero as a conservative approach to yield the highest possible SD to avoid significant false results.

The extracted data regarding the cure rate was pooled as risk ratios, and the data about serum total T4 was pooled as mean differences. Both with the corresponding 95% confidence intervals in a random-effects meta-analysis model using the Mantel-Haenszel equation. Student T-test and Chi-square test were used to test the significant differences in the pooled data between the intervention and the control groups. We used Revman 5.3 software to perform the statistical analysis. We interpreted the results in forest plots for better visualization of data. We identified heterogeneity using the Chi-square heterogeneity test. We solved heterogeneity among studies by performing the leave-one-out sensitivity analysis. If the heterogeneity persisted, we used subgroup analysis. We used the random-effect model for heterogeneous data.

For the randomized controlled trials, we used the Risk of Bias version 2 (ROB2) tool. We followed the Cochrane Handbook for Systematic Reviews of Interventions 5.1.0. in using the tool [26]. The tool consists of five domains that assess quality by answering signaling questions that help the authors determine the overall quality in each domain. The five domains include assessing the risk of bias in the randomization process, the risk of deviation from intended interventions, the risk of missing outcome data, the risk of bias in the measurement of the outcomes, and the risk in selecting the reported results. For the non-randomized trials, we used the ROBINS-1 tool [27]. The tool assessed the studies through signaling questions to determine the risk of bias in seven different domains. The tool included the same four domains as ROB-2 except for the randomization bias, in addition to three other domains regarding the confounding bias, selection bias, and bias in the classification of interventions. Funnel plot were used to assess the risk of bias across the studies.

Data collection retrieved 2003 results; 446 of them were duplicates and removed. The remaining 1557 studies were involved in the title and abstract screening phase. Only 108 studies were eligible for the inclusion criteria and entered the full text-screening phase. Nine studies were included in qualitative synthesis and quantitative synthesis (meta-analysis) [14, 21, 22, 28‐33]. Six were randomized controlled trials [14, 21, 22, 28, 29, 31] and three were non-randomized controlled trials [30, 32, 33]. The study selection process was illustrated in Fig. 1.

The nine studies included 477 patients in the intervention group and 451 patients in the control group, with a total of 928 patients. The general features and the baseline characteristics of the nine included studies were organized in Tables 1 and 2, respectively.

Regarding random sequence generation, only two studies—Lingudu et al. [21] and Thamcharoenvipas et al. [31]—reported an appropriate method of randomization. However, the rest of the studies were judged as unclear due to insufficient information. Allocation concealment risk was unclear for all of the included studies due to insufficient information except for the Hammond et al. study [29]. It was judged as high-risk potential because the investigators could intervene with the patients’ assignment to the groups. The study of Thamcharoenvipas et al. [31] was judged as low-risk potential due to proper allocation concealment.

Regarding the blinding risk of bias, the Thamcharoenvipas et al. [31] study was a high-risk study concerning participants and personnel blinding because the physicians were not blinded. The study of Boggazzi et al., [14] was considered low-risk of bias regarding outcome assessors blinding because the examiners were sufficiently blinded. The rest of the included studies were considered as an unclear risk due to insufficient data.

Regarding the Incomplete outcome data bias, the studies of Bal et al. [22], Hammond et al. [29], and Boggazzi et al. 1998 [14] were considered as sources of high-risk bias due to their high patient attrition percentages (10, 12, and 9%, respectively), without any use of ITT method in the analysis of their data. The rest of the studies were judged as low risk.

For selective reporting, all studies were judged low risk except for the Hammond et al. [29] study. It was judged as high-risk potential due to incomplete reporting of the baseline characters of the patients’ biochemical data. Moreover, all included studies were considered as high-risk potential in the domain of other bias due to the unavailability of the study protocols. All the included studies ranged from unclear to moderate concerning overall quality (Figs. 2 and 3).

On the other hand, of the nonrandomized trials, two (Oszukowska et al. 2010 and Sekulić et al. 2017 studies) were judged as low-risk, while one (Brownlie et al. 1979 study) was judged as high-risk bias [30, 34, 35].

Both Oszukowska et al. 2010 and Sekulić et al. 2017 had low-risk in most the ROBINS-I domains; however, Oszukowska et al. 2010 study had some variability in measuring outcomes among all included patients (bias in the measurement of outcomes). The study of Brownlie et al. 1979 was considered high-risk because of lacking information in multiple domains including confounding and in the selection of participants into the study; moderate to high-risk were found among the domains of missing data, measurement of outcomes, and selection of the reported results (Table 3).

To assess the risk of bias across studies, a funnel plot was created for the main outcome. This analysis showed a significant risk of publication bias regarding the RCTs. On the other hand, the non-randomized controlled trials were nearly symmetrical in both sides of the middle line (Fig. 4).

The total number of patients being treated in the Intervention group is 477 with a cure rate of 84.7% (404 patients), while the number being treated in the control group is 451 patients with a cure rate of 78.5% (354 patients), (RR = 1.11, 95% CI, .96–1.28; P = .17), not favoring any of the two compared groups. Substantial heterogeneity was inspected among the pooled studies (P = .0001, I² = 75%) (Fig. 5).

The pooled data from the RCTs alone (6 studies) as a sub-group did not favor any of the two groups with a cure rate of 85.4% in the intervention group and 80% in the control group (RR = 1.11, 95% CI, .89–1.39; P = .34). Considerable heterogeneity was detected among the pooled studies (P = .0001; I² = 84%) (Fig. 5).

Pooling the data of the non-RCTs (3 studies) as a separate sub-group also did not favor any of the compared groups with a cure rate of 81.6% in the intervention group and 74.6% in the control group (RR = 1.09, 95% CI, .94–1.26; P = .25). The pooled studies were homogenous (Fig. 5).

The overall cure rate of LiCO3 (optimized dose: 5000 to 6500 mg) significantly favored the Intervention (LiCO3) group over the control group—with the exclusion of extremely low [31] or extremely high [22, 33] doses of the study groups. The pooled results showed a cure rate of 83.7% in the intervention (221 patients) versus 66.3% in the control group (199 patients) (RR = 1.27, 95% CI, 1.13–1.42; P = .0001). The pooled studies were homogenous (P = .76, I² = 0%) (Fig. 6).

The effect of the intervention on serum total T4 of the patients was reported in three studies (120 patients). The overall pooled result did not favor either of the two groups (SMD = -24.26, 95% CI, − 0.6 – 0.12; P = .18). The pooled studies were homogenous (P = .82, I² = 0%) (Fig. 7).