Evidence engine
Pick a question, narrow the trials to the people and doses you care about, and see how the pooled result changes. The math is the same inverse-variance method published meta-analyses use.
Tool · Kymata Health 11 outcomes · 77 trials Researched and written by Kymata Health Editorial · Sources cited inline · Checked October 23, 2026 Report an error
What do the trials show for a given age, dose and length?
Trials drop out when they don’t match a filter, with the reason shown. Fewer than 3 left and nothing is pooled.
Forest plot · curcumin vs placebo · knee osteoarthritis pain (SMD)
Bannuru 2018 · Figure 3, analysis 1.1.1In placebo-controlled trials, was knee pain lower with curcumin?
Across 5 placebo-controlled trials, knee pain was lower with curcumin than with placebo: a large difference by the usual yardstick (SMD −0.81, 95% CI −1.25 to −0.37). The trials disagree somewhat.
- Curcumin: the forms and doses these trials used →
- Print it, or save it as a PDF from the print dialog.
- A clinician who knows your health decides what fits you.
* Unpublished data, as cited by the meta-analysis.
All 5 trials included. Pooled SMD −0.81 (95% CI −1.25 to −0.37).
Every trial and where its numbers come from 5 trials
| Trial | Who, dose and how long | SMD [95% CI] | Weight |
|---|---|---|---|
| Haroyan 2018As reported in Bannuru 2018, Fig. 3, analysis 1.1.1Own paper (PMC5761198): Mean change and SD match (Curamed 1.86 ± 2.95, placebo 0.69 ± 2.70; the paper reports improvement as positive). Group sizes differ: the paper analysed 58 and 59; the meta-analysis used 60 and 65. | 125 analysed · mean age 55.4 · 93.3% women · 12 wkCuraMed (BCM-95, 500 mg curcuminoids per capsule) three times a day | −0.41 [−0.77, −0.06] | 24% |
| Madhu 2012As reported in Bannuru 2018, Fig. 3, analysis 1.1.1 | 58 analysed · mean age 56.7 · 56.7% women · 6 wkCurcuma longa extract (NR-INF-02) 500 mg twice a day | −1.64 [−2.24, −1.04] | 19% |
| Moharamzad 2011 (unpublished)As reported in Bannuru 2018, Fig. 3, analysis 1.1.1 | 67 analysed · age not reported · sex not reported · 10 wkCurcumin 600 mg/day | −0.99 [−1.50, −0.48] | 21% |
| Nakagawa 2014As reported in Bannuru 2018, Fig. 3, analysis 1.1.1Own paper (PMC4244558): Mean changes match (0.52 to 0.20 and 0.42 to 0.21). The paper prints no SD of the change; the SDs used (0.24, 0.25) equal its baseline SDs, a standard imputation when change SDs are missing. Across all 41 patients, the paper itself reports no significant difference in VAS change (P = 0.10); its significant result is in the subgroup with baseline VAS above 0.15. | 41 analysed · mean age 68.6 · 78.9% women · 8 wkTheracurmin, 180 mg capsule six times a day (as printed) | −0.44 [−1.06, 0.19] | 18% |
| Panahi 2014As reported in Bannuru 2018, Fig. 3, analysis 1.1.1 | 40 analysed · mean age 57.4 · 77% women · 6 wkC3 curcuminoid complex 500 mg (with 5 mg piperine) three times a dayEach curcuminoid dose came with 5 mg of piperine (Bannuru 2018, Table 1). | −0.65 [−1.29, −0.02] | 18% |
Numbers come from the meta-analysis, as printed in Bannuru RR et al. Semin Arthritis Rheum 2018;48(3):416–429, Figure 3, analysis 1.1.1 (PubMed 29622343); trial details from its Table 1. Pooled with a DerSimonian–Laird random-effects model. Where a trial’s own paper is open, its row says how it compares. With every trial in, this engine reproduces the published pooled result (−0.81 [−1.25, −0.37]). Updated 2026-10-22: Checked against the trials’ own papers where open: every mean matches. Haroyan 2018 analysed fewer people than the meta-analysis lists, and most papers print no SD of the change, so those SDs come from the meta-analysis. Numbers unchanged.
How does the engine pool the trials?
Each trial’s result is the difference between the supplement and comparison groups at the end of the trial. When trials used different scales, the difference is standardized (Hedges’ g) so they can be combined. The engine pools them the way the source paper did: usually with the DerSimonian–Laird random-effects method, which assumes the true effect varies between trials and widens the confidence interval when they disagree, and with a fixed-effect model where the paper used one (each outcome says which). I² is the share of the spread between trials that chance alone does not explain.
Every trial’s numbers are as reported in one published meta-analysis per outcome, copied as printed from the figure named on each row; they have not been re-extracted from each trial’s own paper. For the knee osteoarthritis outcomes (Bannuru et al. 2018), five trials have open full text, and each of those rows says how its own paper compares: the means match, one trial analysed fewer people than the meta-analysis lists, and most papers print no standard deviation of the change. Where a meta-analysis included a trial with no placebo group, a second supplement or another form of the product, its row says so and “Placebo, product alone” drops it. With every trial included, each outcome reproduces its paper’s published pooled result to within 0.02, and our tests check that on each build.
Pooling cannot fix weak trials. Many of these were small, short and funded by the product’s maker, and the products differ from one trial to the next. Read a large pooled effect from a handful of small trials with care.
Questions about the Evidence Engine
What does the Evidence Engine do?
It takes the trials behind a published meta-analysis, lets you keep only those that match an age range, a minimum dose or duration, or who took part, and pools what is left the way the paper did. It covers 11 outcomes and says in plain words what the remaining trials found. Below 3 trials it shows no pooled number.
Does curcumin help knee osteoarthritis pain? What do pooled trials show?
Across 5 placebo-controlled trials, knee pain was lower with curcumin than with placebo: a large difference by the usual yardstick (SMD −0.81, 95% CI −1.25 to −0.37). The trials disagree somewhat. Source: Bannuru et al., 2018, Figure 3, analysis 1.1.1 (PubMed 29622343).
Does melatonin help you fall asleep faster? What do pooled trials show?
Across 15 placebo-controlled trials, the drop in time to fall asleep was larger with melatonin than with placebo: about 7.06 min on average (MD 7.06, 95% CI 4.38 to 9.74). The trials disagree somewhat. Source: Ferracioli-Oda et al., 2013, Figure 1 (PubMed 23691095).
Does creatine add lean mass in older adults who strength-train? What do pooled trials show?
Across 21 controlled comparisons (some trials enter more than once), 3 without a placebo group, the gain in lean tissue mass was larger with creatine than with the comparison groups: about 1.37 kg on average (MD 1.37, 95% CI 0.97 to 1.76). Source: Chilibeck et al., 2017, Figure 1A (PubMed 29138605).