A soy-cluster facet under Soy Products and Health. The decision question: for a peri/postmenopausal adult (the trial-dominant stratum), does taking soy isoflavones (ISFs) do anything measurable for cognition — and does the earlier tofu-harm scare survive scrutiny?
Bottom line (low confidence): supplemental ISFs produce a small, statistically significant improvement on pooled neuropsychological test scores (SMD 0.19), driven almost entirely by memory; every non-memory domain is null. The effect is on a surrogate (test performance over ≤2 y), not on dementia or cognitive-decline events, and the trials are small and short. The older claim that soy/tofu harms cognition does not survive confounding scrutiny. Net: a low-rank candidate lever, not a big rock.
The RCT evidence — small benefit, memory-carried [@cui2020isoflavone]
Cui 2020 pooled 16 RCTs of ISF supplementation (doses ~80–160 mg/d; genistein/daidzein mixes, soy protein, soy milk, or extract; durations 6 weeks–2 y) in adults, mostly postmenopausal women with a few male trials.
- All domains: SMD 0.19 (95% CI 0.07–0.32), P<0.01, I2 18.1% — small, homogeneous, significant.
- Memory: SMD 0.15 (0.03–0.26), P=0.01, I2 0.0% — the domain carrying the pooled signal.
- Global cognition, executive function, psychomotor speed, attention/working memory, language, visuospatial reasoning: all NS (every CI crosses 0).
- «In adults, ISFs improve cognitive function. In the RCTs included in this review, ISF groups were associated with better performance in neuropsychological batteries than the control groups and no severe adverse effects were reported.» (Cui et al., 2019)
Read this with its two load-bearing caveats. The outcome is a continuous test score, a surrogate whose transmission to a patient-important endpoint (incident dementia, rate of decline, preserved function) is not itself evidenced here — no trial measured an event. And the studied range is short: the longest trial ran 2 y, and Cui concedes «it is possible that the durations were too short to show a difference» (Cui et al., 2019). A short trial can under-read a slow-latency benefit or over-read a practice/placebo effect on repeated testing; the direction is not fixed.
The tofu-harm question — answered, and it does not hold [@cui2020isoflavone]
A standing worry (traceable to the Honolulu-Asia Aging Study’s midlife-tofu arm plus two Asian cross-sectional studies) held that soy harms cognition. Cui rebuts each on confounding: HAAS’s cognition arm was secondary to a CVD study with no baseline cognition measured, and its high-tofu group «had a less preferable profile regarding age, education, and ApoE4»; the cross-sectional studies’ better-cognition groups were «much younger and more educated» (Cui et al., 2019). Verdict: «current evidence does not support the proposition that ISFs negatively affect cognition.» (Cui et al., 2019) This is a type-D-adjacent resolution — an apparent harm signal dissolved as confounded observational artifact, matching the The U-Shaped Association Artifact pattern (a directional claim that must survive a confounding/referent check before it is believed).
Reconciling with the observational null — a distinction, not a tension
Flavonoid Intake and Cognitive Function (Peng 2025, observational) reports an isoflavone slice of OR 0.96 (0.81–1.13), I2 56.4% — null (Peng et al., 2026). That looks like it clashes with Cui’s positive RCT pool. It does not join as a tension: the two estimate different quantities. The parameter table (built before this prose, per the cross-source rule):
| Parameter | Cui 2020 (RCT MA) | Peng 2025 isoflavone slice (obs MA) | Same quantity? |
|---|---|---|---|
| Design | RCT (randomized supplementation) | Prospective observational (habitual intake) | NO |
| Exposure | Supplemental / extract ISFs, dosed 80–160 mg/d | Habitual dietary isoflavone intake (high vs low) | NO |
| Outcome scale | Continuous SMD on test batteries | Dichotomous OR on cognitive-impairment/decline events | NO |
| Effect | 0.19 (0.07–0.32), significant | 0.96 (0.81–1.13), null | — (not comparable) |
| Horizon | 6 wk – 2 y | Cohort follow-up (years–decades) | NO |
| Population | Mostly postmenopausal women | Mixed adult cohorts | partial |
With no cell answering same quantity? yes, there is no joined issue. This is recorded as a type-F refinement + type-B disambiguation: isoflavones and cognition is not one object — a dosed short-term supplement effect on a test score and a habitual-diet effect on a decline event are distinct exposures on distinct outcomes, and the evidence can legitimately be positive on the first while null on the second. Neither refutes the other.
The surrogate-transmission gap (type-G). The one thing that would join them — does the test-score bump convert into fewer decline events? — is unmeasured on both sides: Cui has no event endpoint, Peng has no supplement arm. Cui himself calls for «more observational studies on ISFs and dementia». Held as a named gap. — an RCT with an incident-cognitive-outcome endpoint, or a cohort of supplement users, would close it.
Effect modification — equol producers (route b, unestablished) [@cui2020isoflavone]
Equol (a gut-microbial daidzein metabolite; higher ERβ affinity, produced by only ~20–30% of westerners vs 40–70% of Asians) is a plausible effect-modifier: two western RCTs’ post-hoc analyses found equol-producers trended better but «did not reach statistical significance … mainly due to the small percentages of equol-producers in these trials (ie, <25%)» (Cui et al., 2019). A genuine route-(b) effect-modification candidate, underpowered and not established — do not personalize on it yet.
Mechanism (directional) [@cui2020isoflavone]
ISFs selectively bind estrogen-receptor-beta (ERβ), densely expressed in hippocampus and prefrontal cortex — «the effect of ISFs was pronounced in the domain of memory. Memory is highly related to the hippocampus and prefrontal cortex.» (Cui et al., 2019). This is offered to explain why ISFs might differ from HRT (ERα+ERβ), whose trials failed to show cognitive benefit in postmenopausal women — see Hormone Therapy After Menopause, Menopause and the Shifting Levers. The mechanism is corroborative and domain-specific (memory), not itself outcome evidence.
Where this sits in the hierarchy
Low-rank lever. The pooled effect is small and on a surrogate; the population is narrow (peri/postmenopausal); the substitutable-and-mature-drug landscape for cognition is thin, but so is the evidence that the test-score bump matters for lived outcomes. For Dementia Prevention and Modifiable Risk Factors this is a candidate lever, not one of the established levers. The decision-change it does support is a negative one: it licenses dropping the tofu/soy-harm worry.