Older adults often become acutely confused after major surgery. This condition, postoperative delirium, is common, frightening, and linked to later death and dementia. The leading explanation is that surgery causes body-wide inflammation, which then reaches the brain and activates its immune cells, but nobody knows how much of the damage happens in the brain itself. Statins, the cholesterol drugs, look like a cheap way to dampen that inflammation. One hospital record study found statin users had about half the odds of delirium after vascular surgery, while trials in cardiac surgery remain inconsistent. We propose a records-based study that uses a pharmacological fact as a probe. Some statins are lipophilic, meaning fat-soluble, and cross into the brain (atorvastatin, simvastatin). Others are hydrophilic, meaning water-soluble, and mostly stay in the blood (pravastatin, rosuvastatin). Both kinds lower cholesterol and blood inflammation markers. If the two kinds protect equally against delirium, the benefit is probably peripheral. If only brain-penetrant statins protect, that points to an action inside the brain.
Comparing two statins rather than statin users with non-users also fixes the main weakness of earlier work. Patients chosen for any statin differ systematically from those who are not, in comorbidities, in how they engage with healthcare, and in unrecorded frailty. Patients on one statin versus another differ far less. The study is a retrospective cohort (an analysis of historical hospital records) built as a target trial emulation, a protocol that copies a randomized trial's eligibility rules, treatment arms, start of follow-up and outcome definition onto observational data. Its head-to-head design asks a question about disease mechanism that a lumbar puncture study cannot ask at scale.
Core Claim
A head-to-head comparison of brain-penetrant and peripherally restricted statins, run with the same measured blood inflammation response in both arms, can tell us whether perioperative neuroinflammation needs a central target. It can do so in tens of thousands of surgical patients, across surgery types where cerebrospinal fluid (CSF, the fluid around the brain and spinal cord) can never be sampled. The design is also better protected against hidden confounding than the statin-user versus non-user comparisons in the literature. We will test that protection directly with negative-control outcomes rather than assume it.
Why This Gap, and What Prior Work Does
The mechanistic case for a central site of action comes from prospective biofluid cohorts and animal work. Kang et al. sampled preoperative CSF in 186 hip-fracture patients under spinal anesthesia. CSF Lipocalin-2, a neuron-derived signal that activates microglia (the brain's resident immune cells), predicted delirium (OR 2.546, AUC 0.713), while plasma Lipocalin-2 did not separate cases from non-cases. Meng et al. showed in aged mice that surgery triggers stress responses inside hippocampal cells and inflammasome activation, and that blocking these rescues cognition. These studies motivate our hypothesis that the brain compartment matters. They also constrain the design, because the central signal they identify cannot be measured in routine records. The PNDABLE analyses go further. Wang et al., the cardiometabolic multimorbidity study and Liang et al. find that CSF tau mediates only 10–23% of the effect of peripheral risk factors on delirium. Most of the pathway is unexplained, and all of these cohorts are restricted to orthopedic surgery under spinal anesthesia.
The pharmacological evidence is where the two directions meet, and where it is weakest. Lee et al. compared statin users and non-users in 1,132 vascular surgery patients. They found a "sicker patient paradox": no crude association, because users carried more hypertension, diabetes and renal disease, but a protective adjusted OR of 0.54. In 291 propensity-matched pairs the estimate weakened to OR 0.59 (95% CI 0.34–1.02, p = 0.060), alongside lower CRP. They state explicitly that they could not separate statin molecules. Pang et al. review cardiac surgery and describe statin and melatonin trial results as inconsistent. Our proposal treats Lee et al. as the closest existing work and differs from it in three ways. The exposure contrast is within the statin class, chosen to separate central from peripheral action. Measured CRP response serves as a check that both arms are equally anti-inflammatory peripherally. The design spans surgery types, which directly tests whether the vascular/cardiac disagreement reflects real effect modification.
The methods come from the retrospective-cohort literature. Zang et al. ran high-throughput target trial emulation over 4,300 drugs in OneFlorida and MarketScan. They identified atorvastatin, a lipophilic statin, among five candidates for slowing progression to Alzheimer's disease (hazard ratios 0.74–0.92). They also found that regularized logistic regression balanced covariates as well as deep propensity models. We reuse their emulation template and their recommendation to prefer simple propensity models. Pretrained Medical Representations for the Practical Screening of Drug Repositioning Candidates flagged pitavastatin in Japanese claims and warned that rich learned patient representations can leak treatment assignment into baseline covariates. That warning is one reason we avoid learned embeddings as confounder sets. Proximal causal inference, Ying et al. and Proximal Causal Inference for Complex Longitudinal Studies, corrects for an unmeasured confounder using two sets of imperfect proxies: one related to treatment choice and one related to the outcome. We use it as a secondary analysis rather than the main tool, for reasons given below.
Research Program
The emulated trial enrolls adults aged 60 and older admitted for major inpatient surgery (cardiac, vascular, orthopedic, abdominal and thoracic) with at least one overnight stay. Eligible patients must have an outpatient statin prescription covering the day of surgery and present for at least 90 days beforehand. Time zero is the start of surgery. The treatment arms are a lipophilic statin (atorvastatin or simvastatin) versus a hydrophilic statin (rosuvastatin or pravastatin). Pitavastatin and others are analyzed separately because their brain penetrance is less clearly classified. The primary outcome is delirium within seven postoperative days, defined by any positive structured screen (CAM, CAM-ICU or NuDESC) as in Bishara et al..
Two secondary outcome definitions address the known undercounting of delirium in records. The first is an LLM-based note phenotype following Eun et al.. The second is a strict definition (antipsychotic plus documentation). In a head-to-head design, misclassification should be similar across arms. That would bias estimates toward no effect rather than create a spurious one, and agreement across the three definitions is itself a result.
This is a prevalent-user design, because almost no one starts a statin days before surgery. The main confounding threat is not "who gets a statin" but "who gets which statin." Rosuvastatin and atorvastatin are both used at high intensity in high cardiovascular risk patients. Pravastatin and simvastatin skew toward lower intensity, older patients and patients on interacting drugs. We therefore adjust for prescribed intensity on the standard LDL-lowering scale, baseline LDL, kidney function, CYP3A4-interacting medications, prior stroke and dementia, frailty indices built from diagnosis codes, surgery type and urgency, anesthesia type, and prescriber specialty. We use inverse probability of treatment weighting with an L1-regularized logistic propensity model, following Zang et al. Covariate balance must reach standardized mean differences below 0.1. A dose-intensity-matched sensitivity analysis compares atorvastatin 40–80 mg with rosuvastatin 20–40 mg only, the cleanest pair, since both are high-intensity agents prescribed to similar patients.
The peripheral check is what turns a drug comparison into a test of mechanism. Among patients with perioperative CRP measured, we estimate the effect of statin class on postoperative peak CRP. We expect no difference, since both classes have comparable anti-inflammatory effects on blood markers. If that holds and lipophilic statins still reduce delirium, the peripheral-only explanation weakens. If hydrophilic statins lower CRP more and also protect equally or better, the central hypothesis weakens. We prespecify effect modification by surgery type (cardiac vs. vascular vs. orthopedic) and by preoperative cognitive impairment. Blood-brain barrier disruption is plausibly greater after cardiopulmonary bypass, which could change the relevance of brain penetrance.
Data come from a multicenter EHR network of the kind Zang et al. used (OneFlorida). Public replication uses MIMIC-IV, processed with the open Extensive Data Processing Pipeline for MIMIC-IV. In MIMIC-IV, admission medication continuation stands in for home statin use and the outcome is CAM-ICU, so the replication is restricted to patients with an ICU stay. Rough power: with about 10,000 statin users split 70/30 and 10% delirium incidence, the standard error of the log odds ratio is about 0.075. That gives good power to detect an OR near 0.8, which is smaller than Lee et al.'s user-versus-non-user estimate.
Baselines
The naive reference point is the unadjusted delirium rate in lipophilic versus hydrophilic users. Alongside it we report the crude statin-user versus non-user comparison, which in Lee et al. showed no association because of the sicker-patient paradox. These numbers show how much adjustment moves the estimate in each design.
The strongest existing method on this exact question is the Lee et al. design: statin users versus non-users with multivariable adjustment and 1:1 propensity matching. Their reported results are an adjusted OR of 0.54 (95% CI 0.33–0.87) and a matched OR of 0.59 (95% CI 0.34–1.02). We rerun this design on our cohort, with the same covariates, the same outcome definitions, and the same regularized propensity model and tuning budget as our main analysis. "Beating" it here does not mean a smaller odds ratio. It means our design passes falsification tests that the user/non-user design fails. Our estimate should be null on negative-control outcomes where the user/non-user estimate is not, and it should be stable across outcome definitions.
The strongest general-purpose causal competitor is target trial emulation with regularized-logistic IPTW as in Zang et al., applied to user versus non-user. For their Alzheimer's outcome they report hazard ratios of 0.74–0.92; no reported result found on Lacuna for this method on postoperative delirium. We run it with identical eligibility and time zero.
A further competitor is proximal causal inference with the doubly robust estimator of Ying et al. and Proximal Causal Inference for Complex Longitudinal Studies. Treatment-side proxies are prescriber and site prescribing preference. Outcome-side proxies are prior-year inflammatory labs and prior encounter codes for confusion or falls. Ying et al. report that on SUPPORT this approach turned a non-significant standard estimate (p = 0.199) into a significant harmful one (p < 0.001). That shows it can change conclusions, which is why we include it.
The key ablation isolates the active-comparator ingredient. We estimate lipophilic versus no statin and hydrophilic versus no statin separately, then compare their ratio with the direct head-to-head estimate. Agreement means non-user confounding largely cancels. Disagreement shows the head-to-head design is doing real work.
Why not just…?
Why not just run user-versus-non-user propensity matching stratified by statin class? This is the cheapest alternative, and it is our ablation. Its weakness is that the healthy-adherer and sicker-patient biases documented by Lee et al. enter each stratum separately. The class contrast then depends on those biases being equal in size, which nobody can check. The negative-control comparison settles this empirically. If the stratified estimates show "protective" effects on outcomes statins cannot plausibly affect while the head-to-head estimate does not, the simpler design is ruled out. If both pass, the simpler design is adequate and we will say so.
Why not just use proximal causal inference on the user/non-user contrast and skip the drug-class trick? Proximal methods rest on proxy assumptions that cannot be verified from data. With prescriber preference and prior labs as proxies, it is unclear whether they capture subclinical brain inflammation, which Kang et al. found blood markers fail to track. A design that shrinks confounding by construction is more robust than a statistical correction whose inputs are uncertain. We still run the proximal estimator, and agreement between the two approaches strengthens the result.
Why not just wait for or run a prospective CSF cohort or randomized trial? CSF sampling is feasible only under spinal anesthesia in orthopedic surgery, and existing cohorts have 186–875 patients. Even the 788-patient Wang et al. cohort saw lower delirium incidence than its power target. A trial is ultimately needed, but trial designers currently cannot say whether to test atorvastatin or rosuvastatin. This study answers that question at a fraction of the cost.
Why not just treat perioperative CRP as the mediator and run mediation analysis? Kang et al. show that blood markers can fail to reflect central inflammation. Mediation through CRP would therefore underestimate any central pathway and could not separate the compartments. We use CRP as a check that the arms are matched peripherally, not as the explanation.
Decisive Checks
Several prespecified checks guard against residual bias. Negative-control outcomes are postoperative events with no plausible statin-class link: surgical site infection, urinary retention, and postoperative ileus. These should show odds ratios near 1 in the head-to-head design. A negative-control exposure, pre-surgery use of proton pump inhibitors versus H2 blockers, should show no association with delirium after the same adjustment. An E-value analysis reports how strong an unmeasured confounder would need to be to explain any effect.
The prespecified readouts are the head-to-head odds ratio for each outcome definition, the CRP-difference estimate, and their joint pattern. Equal CRP with lower delirium on lipophilic statins supports a central action. Equal CRP and equal delirium supports a peripheral-only action or no effect, as does lower CRP on hydrophilic statins with no delirium disadvantage. A null head-to-head result alongside a protective user/non-user result would suggest that Lee et al.'s association was largely confounding.
The surgery-type interaction directly tests whether the vascular-versus-cardiac disagreement in the literature is real. Replication in MIMIC-IV, with a different outcome instrument (CAM-ICU) and different exposure ascertainment, tests whether the finding depends on one data source.
References
- Lipocalin-2 in preoperative cerebrospinal fluid is a biomarker for postoperative delirium after hip fracture surgery in older adults: a prospective cohort study
- Inhibiting endoplasmic reticulum stress alleviates perioperative neurocognitive disorders by reducing neuroinflammation mediated by <scp>NLRP3</scp> inflammasome activation
- Correlation of β2-microglobulin with postoperative delirium and 3-year mortality undergoing knee or hip replacement surgery: a prospective cohort study
- Association of cardiometabolic multimorbidity with postoperative delirium and three-year mortality in patients undergoing knee/hip arthroplasty: a prospective cohort study
- Tau protein mediates the association between frailty and postoperative delirium: a machine learning model incorporating cerebrospinal fluid biomarkers
- Preoperative statins are associated with a reduced risk of postoperative delirium following vascular surgery
- Effects of inflammation and oxidative stress on postoperative delirium in cardiac surgery
- High-throughput target trial emulation for Alzheimer’s disease drug repurposing with real-world data
- Pretrained Medical Representations for the Practical Screening of Drug Repositioning Candidates
- Proximal Causal Inference for Marginal Counterfactual Survival Curves
- Proximal causal inference for complex longitudinal studies
- Postoperative delirium prediction using machine learning models and preoperative electronic health record data
- Efficacy of large language models in detecting postoperative delirium from unstructured clinical notes: A retrospective cohort study
- An Extensive Data Processing Pipeline for MIMIC-IV