What the Evidence Actually Shows for Traditional Chinese Medicine in Age-Related Cognitive Decline

Herbal formulas used in traditional East Asian medicine have now been compared against standard dementia drugs in a substantial number of randomized trials. Pooled analyses favour them on the cognitive endpoint most often used as a primary measure in Alzheimer’s registration trials, with adverse event rates indistinguishable from active comparators. The trials carry well-documented methodological weaknesses, and the finding is worth examining rather than dismissing on that basis.

Findings

The larger of two recent syntheses pooled 57 randomized controlled trials of herbal medicine in Alzheimer’s disease (Lee et al., 2022). Relative to conventional treatment, herbal medicine improved both the Mini-Mental State Examination and the Alzheimer’s Disease Assessment Scale cognitive subscale. Among trials comparing herbal medicine directly against donepezil, the agent most widely prescribed for the condition, MMSE scores were equivalent and ADAS-cog scores favoured the herbal arm. Given as an adjunct rather than a substitute, herbal medicine improved both scales beyond conventional treatment alone and lowered plasma amyloid beta.

A second synthesis covering 23 trials and approximately 2,000 patients reported a standardized mean difference of −0.22 on ADAS-cog favouring herbal treatment, together with a benefit on the Hasegawa Dementia Scale (Wang et al., 2025). Adverse event rates did not differ between arms, and heterogeneity on that outcome was zero, the only such value in either analysis.

The ADAS-cog result is the most durable finding in this literature. It appears independently in both syntheses and survives direct comparison against active drug rather than placebo. ADAS-cog is the more sensitive of the two instruments in common use and the one generally preferred as a primary endpoint in Alzheimer’s trials, which lends the convergence some weight.

Two endpoints showed no benefit. MMSE improved in the larger synthesis but not in the smaller one. Behavioural and psychological symptoms, assessed by the Neuropsychiatric Inventory, did not differ from placebo. Cognitive and behavioural domains in dementia respond to different mechanisms and dissociate routinely in trials of approved agents, so a cognitive signal unaccompanied by a behavioural one is an ordinary pattern rather than an internal inconsistency.

Why dementia tests the multi-target hypothesis

The plausibility of these results rests on what kind of disease dementia is. Alzheimer’s disease involves amyloid deposition and tau pathology alongside chronic neuroinflammation, cerebrovascular insufficiency, mitochondrial dysfunction, impaired synaptic maintenance and the accumulation of senescent cells. Several of these processes are also hallmarks of biological aging, which places age-related cognitive decline within aging biology rather than beside it.

Agents directed at single nodes within that network have produced limited clinical benefit. Cholinesterase inhibitors and memantine yield small, temporary symptomatic effects without altering disease course. Anti-amyloid monoclonal antibodies, developed over decades of work on one hypothesis, slow decline by margins that remain contested in clinical significance and carry amyloid-related imaging abnormalities as a recognized risk.

Multi-target interventions have a structural rationale under these conditions. Traditional East Asian medicine has produced such interventions empirically for roughly two millennia, selecting and refining multi-herb formulas by observed response. The trials summarized above represent the first substantial attempt to evaluate that inheritance against modern comparators.

Pharmacological basis

The case for multi-herb formulas is usually made by observing that they contain many compounds acting on many targets. Stated that way the claim carries little information, since the same is true of most foods. Two distinctions make it testable.

Pharmacokinetic against pharmacodynamic synergy. Interactions between constituents of a formula frequently concern absorption and clearance rather than target biology. One compound inhibits the enzyme that metabolizes another, and systemic exposure to the second rises accordingly. The Alzheimer’s formula Bushen-Yizhi provides a well-characterized example. Screening its constituents for CYP3A4 inhibition identified four active compounds, of which a Schisandra lignan proved the most potent, with an IC50 of 0.339 µM. In rats, administration of the whole formula raised the area under the curve for osthole, its principal active constituent, from 228 to 2,196 ng/L·h, an approximately tenfold increase. The isolated lignan achieved 1,053, while ketoconazole, a clinical CYP3A4 inhibitor used as a positive control, reached only 463 (Liu et al., 2022).

The classical framework anticipated this category of interaction. Formula construction in traditional practice assigns constituent herbs to four roles, the fourth of which, conventionally translated as envoy, denotes the herb that directs the others to their site of action. A system with no concept of hepatic metabolism nonetheless reserved a functional position for delivery.

Characterized against uncharacterized polypharmacology. Dasatinib engages dozens of kinases, as does a decoction. The two differ in documentation rather than in promiscuity. Dasatinib’s target profile has been mapped across the kinome at known affinities and known dose, and several of its adverse effects are attributable to specific off-target activity. A decoction engages a comparably broad target range at unknown occupancies from a dose that varies between batches. The limitation of traditional formulas lies in characterization, and that limitation is tractable.

Summary of endpoints

Taken together, the evidence supports a modest cognitive benefit on the more sensitive of the two standard instruments, in Chinese trial populations, with a safety profile comparable to approved agents. It does not support benefit for behavioural symptoms or daily function.

Two methodological cautions

Two observations from our own work on senescence-targeting formulations bear on how evidence of this kind should be read.

The first concerns transcriptional readouts. A published approach proposed identifying natural substitutes for a synthetic senolytic by matching gene-expression signatures. Reproducing that logic, the compounds under evaluation failed to rank, which initially appeared to be a negative result about the compounds. The better interpretation is that the readout was mismatched to the endpoint. Senolysis proceeds through protein-level machinery already present in the cell and does not require a distinctive transcriptional program, while bulk transcriptomics systematically loses the dying subpopulation it is meant to detect. Gene expression remains well suited to measuring suppression of inflammatory signalling, which genuinely is transcriptional. The same caution applies to network-pharmacology analyses of traditional formulas, where a map of converging targets constitutes a mechanistic hypothesis rather than evidence of effect.

The second concerns exposure. Quercetin, fisetin and curcumin all engage relevant targets in vitro at concentrations that human plasma does not reach following oral administration. Target engagement established in cell culture should therefore be filtered by achievable systemic exposure before a compound advances. This constraint applies with particular force to decoctions, which deliver variable doses of multiple poorly absorbed constituents, and it is the reason the osthole pharmacokinetic result described above carries weight: a tenfold increase in exposure represents substantive pharmacological work.

Limitations

The trials supporting these findings are methodologically weak, and the weaknesses are consistent across both syntheses. In the 2025 analysis, 15 of 23 trials scored 3 or below on a 7-point quality scale. Ten described their randomization procedure adequately. Nineteen reported a random-number method without establishing that allocation was concealed from enrolling investigators. Blinding was seldom reported and only two trials employed a placebo. Heterogeneity reached 81 percent on MMSE and 94 percent on the Hasegawa scale, levels at which pooled estimates become difficult to interpret. All participants were recruited in China, so generalization to other populations is untested. Publication bias cannot be excluded: the authors report none overall, though their test on the headline effectiveness measure falls close to the conventional threshold, and a literature composed largely of small trials carries that risk independent of any single statistic. The 2022 authors state directly that concrete conclusions cannot be drawn from their inputs. These limitations constrain the confidence attachable to the effect size; they do not explain away a result that replicates across two independent pools of trials against an active comparator, and they are addressable by better trial conduct rather than inherent to the subject.

What would settle the question

Four changes in trial design would convert this from a promising signal into a decided question.

1. Blinded, placebo-controlled trials. Two of 23 trials used a placebo. Matching the appearance and taste of a decoction is genuinely difficult, which partly explains the omission, though encapsulated standardized extracts remove most of that difficulty.
2. Concealed allocation, documented. The least expensive correction available and the most consistently omitted. Nineteen trials described randomization without establishing that allocation was hidden from enrolling investigators.
3. One standardized formula tested repeatedly. Much of the heterogeneity that makes pooling unreliable arises because each trial evaluates a different preparation. A single characterized formula with quantified marker compounds, examined across several independent centres, would yield evidence that can legitimately be combined.
4. Replication outside China. An effect observed only within one country’s trial system remains a finding about that trial system until demonstrated elsewhere.

None of this requires methodological innovation. It requires the standard apparatus of clinical research applied to a subject that has largely been studied without it. The obstacle is economic rather than scientific, since an unpatentable formula seldom attracts funding sufficient for a definitive trial. That explains why the evidence is thin without making thin evidence stronger than it is.

Sources

• Lee J, Kwon S, Jin C, et al. Traditional East Asian Herbal Medicine Treatment for Alzheimer’s Disease: A Systematic Review and Meta-Analysis. Pharmaceuticals 2022;15(2):174.
• Wang Q, Wang D-L, Zhang X-C, et al. Efficacy and safety of traditional Chinese medicine and Western medicine in Alzheimer’s disease: a systematic review and meta-analysis. Frontiers in Neurology 2025;16:1607945.
• Liu JM, Chen JM, Lin MJ, et al. Screening and verification of CYP3A4 inhibitors from Bushen-Yizhi formula to enhance the bioavailability of osthole in rat plasma. Journal of Ethnopharmacology 2022;282:114643.
• Jiang W, Tang M, Yang L, et al. Analgesic Alkaloids Derived From Traditional Chinese Medicine in Pain Management. Frontiers in Pharmacology 2022;13:851508.
• Zhu Y, Tchkonia T, Pirtskhalava T, et al. The Achilles’ heel of senescent cells: from transcriptome to senolytic drugs. Aging Cell 2015;14(4):644–658.
• Meiners F, Hinz B, Boeckmann L, et al. Computational identification of natural senotherapeutic compounds that mimic dasatinib based on gene expression data. Scientific Reports 2024;14:6286.

These statements have not been evaluated by the Food and Drug Administration. Nothing here is intended to diagnose, treat, cure or prevent any disease, and nothing here is medical advice. Dementia is a serious condition that warrants care from a qualified clinician.