Can lowering blood pressure protect cognition?

The defensible conclusion is that blood pressure management has cognitive protective potential, with the size and certainty of benefit depending on the population, comparison and measurement. SPRINT provides more support for mild cognitive impairment than for dementia alone; a rural Chinese trial reports a lower four-year proportion assessed as having dementia; the 2026 ESPRIT report finds no clear improvement in the average cognitive test score. Each result needs its own design context.

01What did the trials actually measure?

A dementia diagnosis generally requires cognitive decline that interferes with daily life. Mild cognitive impairment can leave substantial independence intact. Cognitive tests measure performance in selected abilities; brain imaging measures structure or blood flow. These outcomes are not interchangeable.

These are trial targets, not personal targets proposed for readers. Achieved pressure, follow-up and medication adjustment are parts of the interventions. CRHCP's CIND definition also differs from SPRINT's requirement for two consecutive MCI assessments. SPRINT Protocol and supplement CRHCP ESPRIT

02SPRINT: favourable direction, continuing uncertainty for dementia

SPRINT randomized 9,361 adults aged at least 50 years with elevated cardiovascular risk. People with diabetes or previous stroke, among others, were excluded. The blood pressure intervention ended early because of cardiovascular benefit, after approximately 3.3 years; cognitive follow-up continued. Nearly seven years of cognitive observation therefore does not mean seven years of continuous intensive treatment. Original report Extended follow-up

The 2025 report combined the original trial and extended follow-up, with median observation of 6.9 years:

For dementia, p=0.08 establishes neither a definite reduction nor the absence of an effect: the interval still accommodates a hazard reduction of approximately 28% and a small increase. For MCI, the reported p=0.04 can coexist with a confidence limit rounded to 1.00. Secondary outcomes were not adjusted for multiple comparisons, which limits the strength of inference. Table 3

We recalculated crude rates from events and person-years. Dementia rates were approximately 8.51 versus 10.15 per 1,000 person-years. These are not fixed seven-year risks and do not reproduce the clinic-stratified Cox model. The 2019 and 2025 reports contain overlapping follow-up of one trial and cannot be pooled as independent trials.

SPRINT uses cause-specific hazard ratios; CRHCP uses ratios accounting for village clustering and stratification. The figure preserves those different estimands and supplies no combined “anti-dementia effect.”

03The rural trial: lower dementia proportions, with consequential denominators

CRHCP randomized 326 rural Chinese villages, enrolling 33,995 adults aged at least 40 years with poorly controlled blood pressure. Trained nonphysician community providers initiated and adjusted medication with primary-care physician supervision. The comparator was local usual care. This tests a care programme and does not identify an effect of one particular antihypertensive drug. Abstract, main tables and supplementary materials

The main table reports 668 versus 734 dementia cases, with a primary RR of 0.85 (95% CI 0.76–0.95). Further baseline covariate adjustment gives 0.88 (0.79–0.98). The primary model already accounts for village clustering and stratification: the absence of additional covariate adjustment does not mean treating all participants as independent.

The denominators matter:

The last row comes from supplementary and figure source data and agrees with the main table's 4.59% versus 5.40%. Using those denominators gives a crude difference of approximately 8.06 fewer cases per 1,000 cognitively assessed participants. This is not the causal benefit per 1,000 randomized participants at a common time point. Its reciprocal is not a defensible population number needed to treat: death and incomplete assessment affect who enters the comparison.

The authors also report imputation analyses and a dementia-or-death composite, with results in the same direction. These do not automatically resolve every missingness mechanism. One supplementary cross-trial summary uses the larger clinical follow-up denominators. Supplementary Table 4 also reports an adjusted RR of 0.89 (0.80–0.98) alongside p<0.0001; the corresponding test details remain unresolved. We retain these source differences and use the mutually cross-checkable main tables and figure data, without reconstructing a new meta-analysis from them. Table 3, supplementary Tables 3–8 and source data

Pressure separation also changes with time. At 48 months, mean systolic pressure was 127.6 versus 147.7 mm Hg, an endpoint difference of approximately 20.1. Because baseline values differed, the difference in changes from baseline was approximately 22.0. Both numbers can be correct, but they are not interchangeable.

Trapezoidal integration between the labelled group means gives an average systolic separation of approximately 15.01 mm Hg over four years. This descriptive calculation shows that the larger final difference was not sustained throughout follow-up. It neither identifies a dementia effect per mm Hg nor recovers individual pressure trajectories.

04Mean test scores and diagnostic outcomes answer different questions

The 2026 ESPRIT cognitive report includes 11,255 Chinese adults followed for a median of approximately 3.4 years, including people with previous stroke and diabetes. The intensive-minus-standard difference in mean MMSE change was +0.05 points (95% CI −0.07 to +0.17), p=0.40. This does not show a clear average improvement. The interval also constrains the average harm compatible with this measurement and analysis, without proving that all cognitive risks are absent. ESPRIT

MMSE was measured twice, and cognitive assessors knew the assigned group. Dementia relied on investigator reporting, with too few cases to assess prevention reliably. A mean score result cannot negate all dementia prevention evidence; conversely, fewer dementia diagnoses do not establish improved memory in every participant.

Another issue deserves separate treatment. Final MMSE was missing for 815 participants, including 363 who had died. The primary analysis imputed missing scores, including scores for deceased participants. A real cognitive score after death does not exist, so this analysis needs a clearly stated hypothetical interpretation. The authors also excluded deaths, examined complete cases and analysed death or a decline of at least three MMSE points; these analyses likewise showed no clear group advantage. Actual supplementary tables

The figure presents author-reported sensitivity analyses. We did not obtain participant data or rerun their imputation. Comparing survivors can itself introduce selection, so similar results across analyses do not establish the absence of all relevant bias.

Other tests reinforce the limits. In ACCORD participants with diabetes, intensive blood pressure treatment did not clearly improve the primary cognitive test. In SPRINT's detailed testing substudy of approximately 2,900 participants, memory change did not clearly differ, while processing speed showed a very small adverse annual difference of approximately −0.010 standardized score units (95% CI −0.017 to −0.002). Scores were standardized using baseline medians and interquartile ranges; these are not simply standard deviation units. The findings do not support improvement across every cognitive domain. Clinical importance, substudy selection and multiple testing all matter. SPRINT domain analysis ACCORD

05Vascular evidence does not make imaging a substitute for clinical benefit

Longstanding hypertension may affect brain tissue through small-vessel injury, vascular remodelling and impaired regulation of perfusion. Lower pressure does not necessarily cause a proportional fall in cerebral blood flow: flow depends on both the perfusion pressure difference and vascular resistance.

A simple relation is: relative flow = relative pressure difference ÷ relative resistance. If pressure difference falls 15% and resistance stays constant, modelled flow falls 15%. If both fall 15%, flow is unchanged; if resistance falls 20%, flow increases approximately 6.25%. These are conditional illustrations, not measured human autoregulatory thresholds.

Randomized imaging data provide empirical constraints. The SPRINT perfusion substudy had baseline data for 547 participants and follow-up scans for 315. The between-group difference in whole-brain perfusion change was +2.30 mL/100 g/min (95% CI 0.30–4.30). That conflicts with a simple claim that sustained intensive treatment necessarily reduces resting cerebral perfusion. But scans measured resting supine perfusion and cannot exclude hypoperfusion during standing or acute changes. Perfusion study

Structural imaging was mixed. White-matter lesion volume increased less, with a robust model estimate of −0.54 cm³ (−0.87 to −0.20) for the between-group difference in change. Total brain volume decreased more, with a difference of −3.7 cm³ (−6.3 to −1.1). The former supports an effect on small-vessel-related injury; the tissue and functional meaning of the latter remains uncertain. Selecting only the favourable marker cannot establish “brain rejuvenation,” and volume loss alone does not establish additional neuronal injury. Structural imaging study

These imaging and cognitive reports belong to the same SPRINT study programme. They do not establish how much cognitive benefit is mediated by blood flow, white-matter lesions or a particular molecular pathway. Similar findings after statistical adjustment for medication classes cannot replace a randomized drug-class comparison or identify mediation.

06Competing death: longer survival can create more opportunity for diagnosis

For a first dementia diagnosis, death is not ordinary missing measurement: a person cannot receive their first diagnosis after death. Cause-specific hazard ratios, cumulative diagnosis probabilities accounting for death, and dementia-or-death composites therefore answer different questions.

We constructed an explicitly hypothetical model with annual diagnosis and prediagnosis death hazards of 0.01 and 0.02. Treatment lasts 3.3 years and observation seven years. Scenarios reduce the diagnosis hazard by 20%, the death hazard by 25%, or both, with effects either ending or partly persisting after treatment. The model describes two competing first events; it does not model death after diagnosis or fit trial participants.

Reducing only the death hazard slightly increases cumulative diagnosis probability in this model, because more people remain alive to receive a diagnosis. Even when the diagnosis hazard falls, cumulative benefit depends on death risk and treatment duration. A hazard ratio is therefore not a fixed-time risk ratio and cannot by itself tell us how many years dementia is delayed.

SPRINT's extended report presents both cause-specific Cox and competing-death analyses, with similar dementia estimates. Nondeath loss to follow-up remains a separate issue: cognitive status was obtained for approximately 59% of those eligible for extended follow-up, and some baseline characteristics differed among nonparticipants. Similar response fractions in the two groups do not establish similar missing cognitive outcomes. Extended follow-up

The original SPRINT missingness supplement also reveals a definitional distinction. Its imputation analyses addressed dementia and first MCI, rather than MCI requiring two consecutive confirmations. Imputed dementia estimates still did not exclude no effect; most imputed first-MCI intervals also crossed one. These analyses cannot be presented as equivalent robustness verification for the two-consecutive-MCI result. Original statistical supplement

07Present cognitive benefit alongside specific harms

SPRINT's 2021 final report completed event ascertainment and corrected safety coding. Updated intervention-period data show serious hypotension in 99/4,678 versus 58/4,683 participants, HR=1.71 (1.24–2.38). Serious acute kidney injury or acute renal failure occurred in 193/4,678 versus 115/4,683, HR=1.69 (1.34–2.13). Our crude proportions correspond to approximately 8.78 and 16.70 additional affected participants per 1,000 randomized, respectively. Final safety report

These are intervention-period proportions. Subtracting them from four-year cognitive proportions among assessed participants in another trial would not produce a valid net-benefit score. One participant may also appear in several adverse-event categories, so rows must not be added together.

CRHCP reported fewer overall serious adverse events, defined to include death and hospitalization. Intervals for specific outcomes such as hypotension and syncope still accommodate increased risk. The trials differ in baseline pressure, usual-care comparison and safety definitions; their findings cannot cancel one another out. Cardiovascular benefit and increased specific adverse events can coexist. CRHCP Table 3

08Scope of the conclusion

In the hypertensive populations studied, some more intensive management strategies can affect selected cognitive outcomes, with supporting evidence from small-vessel imaging and resting perfusion. The evidence does not establish benefit for everyone, every cognitive domain or every lower pressure target. Cognitive findings alone also do not establish general slowing of aging.

We recalculated public event counts, person-years, proportions, blood pressure figure data and safety tables. Two explicitly assumed models explain perfusion and competing death, with computations checked through separate implementations. We did not obtain participant data or refit Cox models, village-level models, imputation or brain imaging. The complete CRHCP main text and original protocol were unavailable; verified materials include the official abstract, main tables, flow diagram, public figure data and supplementary materials. Baseline cognitive exclusion and the full missing-data procedures remain reading limitations, not details this article can invent.

Download calculation code, inputs, figures and verification results