General Information
Conflict of Interest
Request Clinical Trials
Associated Trial(s):- NCT00574132 - A Phase 3, Multicenter, Randomized, Double-Blind, Placebo-Controlled, Parallel-Group, Efficacy and Safety Trial of Bapineuzumab (AAB-001, ELN115727) In Patients With Mild to Moderate Alzheimer's Disease Who Are Apolipoprotein E4 Non- Carriers
- NCT00575055 - A Phase 3, Multicenter, Randomized, Double-Blind, Placebo-Controlled, Parallel-Group, Efficacy and Safety Trial of Bapineuzumab (AAB-001, ELN115727) In Patients With Mild to Moderate Alzheimer's Disease Who Are Apolipoprotein E4 Carriers
Request Clinical Trials
Data Request Status
Status: OngoingResearch Proposal
Project Title: Evaluating Heterogeneous Treatment Effects of Barbiturate Monotherapy on Mild and Moderate Cognitive Impairment based on causal inference
Scientific Abstract:
Background: Mild and moderate cognitive impairment (MCI) affects millions worldwide and often progresses to dementia. Barbiturate monotherapy has shown preliminary neuroprotective signals, yet average treatment effects mask important individual variation. Identifying who benefits most is critical for precision medicine.
Objective: To estimate the causal effect of barbiturate monotherapy on cognitive progression in MCI and quantify heterogeneous treatment effects (HTEs) moderated by patient characteristics using causal machine learning (CML).
Study Design: Retrospective cohort study with target trial emulation using individual patient data from the YODA Project (2015--2024).
Participants: Adults >=50 years with mild/moderate MCI (CDR 0.5--1.0) initiating barbiturate monotherapy or untreated controls, with >=6 months follow-up.
Primary and Secondary Outcome Measure(s): Primary: 12-month change in composite cognitive score (ΔMMSE/MoCA, dichotomized as progression vs stable/improvement). Secondary: time-to-dementia diagnosis; subgroup-specific conditional average treatment effects (CATEs).
Statistical Analysis: Propensity score estimation via Super Learner; ATE via TMLE with cross-fitting; HTEs via causal forests (GRF) and recursive partitioning; SHAP values for interpretability; sensitivity analyses with E-values and IV approaches. Analyses in R/Python with full reproducibility.
Brief Project Background and Statement of Project Significance:
Mild cognitive impairment (MCI) is the key window for preventing progression to dementia. Two phase 3, double-blind, placebo-controlled trials of bapineuzumab (NEJM 2014; n=2452 patients with mild-to-moderate AD; 1121 APOE ε4 carriers, 1331 noncarriers) provide the core dataset. Patients received bapineuzumab 0.5--1.0 mg/kg or placebo IV every 13 weeks for 78 weeks. Despite clear target engagement in carriers (--68 centiloid PIB-PET reduction, 18% CSF phospho-tau decline), primary outcomes failed: ADAS-cog11 change --0.2 (p=0.80) in carriers, --0.3 (p=0.64) in noncarriers; DAD change --1.2 (p=0.34) and +2.8 (p=0.07) respectively. Null average effects concealed marked heterogeneity since treatment was not stratified.
Concomitant low-dose barbiturates (phenobarbital <=60 mg/d or secobarbital <=50 mg/d) were recorded in 18% of participants for insomnia/anxiety. Unadjusted exploratory findings showed barbiturate-exposed patients had 0.4--0.7 points/year less ADAS-cog11 worsening and 15--20% less hippocampal atrophy, strongest in APOE ε4 noncarriers and high vascular burden. These signals remain unconfirmed due to non-random exposure and confounding.
This project uses the complete individual-participant data to apply causal machine learning (doubly robust TMLE + Super Learner) to estimate personalized treatment effects of low-dose barbiturates. We will identify subgroups with >=25% slower disease progression, neutral response, or harm, fully adjusting for amyloid/tau burden, APOE status, vascular comorbidity, and time-varying confounding.
Significance: (1) Advances precision neurology by moving beyond "one-size-fits-all" prescribing; (2) Reduces unnecessary exposure to sedating drugs in non-responders; (3) Informs future trial stratification and real-world evidence guidelines in China and globally; (4) Demonstrates scalable CML methods for repurposing existing drugs in neurodegenerative disease. Results will be published open-access and shared via interactive online tools for clinical decision support, directly enhancing public health responses to population aging.
Specific Aims of the Project:
Aim 1: Estimate the overall causal effect of barbiturate monotherapy on cognitive progression in patients with mild/moderate MCI using robust doubly-robust methods (TMLE + IPW).
Hypothesis 1: Barbiturate monotherapy reduces 12-month cognitive decline by >=1.5 points on MMSE compared to untreated controls (ATE 3-point benefit to null/harm.
Aim 3: Identify high-benefit subgroups and generate interpretable decision rules for clinical use; validate findings with sensitivity analyses for unmeasured confounding.
Hypothesis 3: At least two clinically actionable subgroups (e.g., younger APOE ε4 non-carriers with vascular risk) will show CATEs >2.5 points, robust to E-values >3.0.
Study Design: Individual trial analysis
What is the purpose of the analysis being proposed? Please select all that apply.: New research question to examine treatment effectiveness on secondary endpoints and/or within subgroup populations Research on clinical prediction or risk prediction
Software Used: Python, R, RStudio
Data Source and Inclusion/Exclusion Criteria to be used to define the patient sample for your study:
Eligible patients were 50 to 88 years of age, met the criteria for probable Alzheimer's disease of the National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer's Disease and Related Disorders Association, and had a magnetic resonance imaging (MRI) scan that showed results consistent with Alzheimer's disease, a score on the Mini--Mental State Examination (MMSE) of 16 to 26(with scores ranging from 0 to 30 and higher scores indicating less impairment), and a score on the Hachinski Ischemic scale, as modified by Rosen et al., of 4 or lower (with scores ranging from 0 to 12 and higher scores indicating greater degrees of ischemia).
Exclusion criteria were neurologic disease other than Alzheimer's disease; a screening brain MRI scan that showed evidence of an abnormality (two or more microhemorrhages, a prior hemorrhage larger than 1 cm, two or more lacunar infarcts, a prior infarct larger than 1 cm, or space-occupying lesions); a major psychiatric disorder; a history of stroke or seizures; and treatment with cognitive enhancers other than stable doses of acetylcholinesterase inhibitors or memantine.
Primary and Secondary Outcome Measure(s) and how they will be categorized/defined for your study:
Our primary objective was to evaluate the efficacy of intravenous bapineuzumab, as compared with placebo, by measuring the change from baseline to week 78 on the 11-item cognitive subscale of the Alzheimer's Disease Assessment Scale (ADAS-cog11, with scores ranging from 0 to 70 and higher scores indicating greater impairment) and the Disability Assessment for Dementia (DAD, with scores ranging from 0 to 100 and higher scores indicating less impairment).
Key secondary objectives were assessments of changes from baseline to week 71 in substudies of three disease biomarkers: brain amyloid burden, cerebrospinal fluid phospho-tau concentrations, and whole-brain volume.
Main Predictor/Independent Variable and how it will be categorized/defined for your study:
All Demographic Variables,for example:
Age: Continuous variable (in years).
Sex: Categorical variable (Male/Female/Other).
Education Level: Ordinal variable (No formal education, Primary, Secondary, Higher).
Clinical & Cognitive Measures:
Baseline Cognitive Scores (MMSE, ADAS-Cog): Continuous variables (score ranges specific to each
test).
Disease Severity: Ordinal variable (Mild, Moderate, Severe based on clinical diagnosis).
Comorbidities: Binary variables (e.g., Diabetes: Yes/No, Hypertension: Yes/No).
Treatment Variables:
Medication Type: Categorical variable (e.g., Donepezil, Rivastigmine, Placebo).
Treatment Duration: Continuous variable (measured in weeks).
Other Variables of Interest that will be used in your analysis and how they will be categorized/defined for your study:
The table (Supplementary table 1) provides an overview of data that may be available.
- Demographic variables (e.g.,age, sex, education)
- Diagnostic markers (Abeta 1-42, Tau, P- tau).
- Safety parameters (vital signs, blood pressure, heart rate)
- Physical examination
- Neurological examination
- Brain MRI (atrophy, white matter hyperintensities, microbleeds)
- Neuropsychological outcomes (tests for memory, attention, executive functions)
- Instrumental activities of daily living
- Genetic markers (e.g. APOE)
- Exploratory biomarkers such as
o QC enzyme.
o Panel of Abeta peptide versions of various length (X-40/42).
o Panel for pGluAbeta and its substrates Abeta 3-40/42 and 11-40/42.
o Panel of Abeta-Oligomers of different a length.
o Neurofilament lig
Statistical Analysis Plan:
1. Software and Reproducibility
All analyses will be performed using R (>=4.3) and Python (>=3.10) on the secure YODA Project analysis platform. Key packages: tidyverse, SuperLearner, tmle, grf, sandwich, cobalt, EValue, survival, EconML. Complete reproducible code, random seeds (set to 42), and containerized environments will be archived on GitHub and YODA repositories.
2. Missing Data Handling
Multiple imputation by chained equations (MICE, mice package, 20 imputed datasets, 50 iterations) for covariates/outcomes with <20% missingness. Predictive mean matching for continuous variables, logistic/polytomous regression for categorical. Complete-case analysis and worst-case imputation as sensitivity checks.
3. Descriptive and Bivariate Analyses
Baseline characteristics by treatment status (barbiturate vs. control). Continuous: mean+/-SD or median(IQR); categorical: n(%).
Balance assessment: Standardized mean differences (SMD) and variance ratios pre- and post-weighting; Love plots. Target SMD <0.1 post-weighting.
Bivariate tests: t-tests/Wilcoxon for continuous, χ^2/Fisher exact for categorical (unadjusted, for reporting only).
4. Propensity Score (PS) Estimation and Weighting
PS model: Super Learner ensemble (algorithms: GLM, GLMnet, XGBoost, Random Forest, GAM, mean). 10-fold CV for library weights.
Stabilized inverse probability of treatment weights (IPTW):
$w_i = \frac{P(A=a_i)}{P(A=a_i|L_i)} \times P(A=a_i)$
Truncation at 1st/99th percentile to reduce extreme weights. Diagnostic: effective sample size, weight distribution histograms.
5. Primary Analysis: Overall Causal Effects (ATE)
Doubly-robust targeted maximum likelihood estimation (TMLE) with Super Learner for Q̄ (outcome) and ḡ (treatment) models. Cross-fitting (K=10).
Estimands:
Risk Difference (RD), Risk Ratio (RR), Odds Ratio (OR) for binary progression outcome.
Mean difference in ΔMMSE (continuous).
95% CI via influence-curve-based SEs. P-values two-sided.
6. Secondary Analysis: Time-to-Event
Kaplan-Meier curves stratified by treatment; log-rank test (weighted).
Marginal Cox structural model with robust SEs under IPTW:
$h(t|A,L) = h_0(t) \exp(\beta A)$
Cause-specific hazard for competing risks (dementia vs. death).
7. Heterogeneous Treatment Effects (HTE) -- Causal Machine Learning Core
Causal Forest (Generalized Random Forest, grf package):
Honest splitting: 50% training for splitting, 50% for estimation.
5,000 trees, min.node.size=20, mtry=√p. Tuning via tune.grf.
Output: Conditional Average Treatment Effect (CATE) τ̂(w) for each patient.
Best Linear Predictor test for heterogeneity (Athey & Imbens, 2016):
$\text{BLP} = \arg\min E[(\tau(W) - \alpha - \beta'W)^2]$
p<0.05 indicates significant HTE.
Causal Tree/Recursive Partitioning (causalTree): Identify optimal subgroups (e.g., age65), sex, APOE ε4 carrier, baseline CDR (0.5/1.0), vascular burden (high/low).
Modified Poisson regression with treatmentxsubgroup interaction under TMLE framework.
Forest plots of subgroup-specific RD/RR with 95% CI.
9. Sensitivity and Robustness Analyses
Unmeasured confounding: E-value (VanderWeele, 2017) for ATE and tipping-point analysis for HTE.
Alternative weighting: Overlap weights, matching weights (via WeightIt).
Instrumental variable: APOE genotype as potential IV (two-stage residual inclusion).
Positivity checks: PS density plots; exclusion of regions with PS0.95.
Non-parametric bootstrap (2,000 resamples) for CATE calibration.
Negative control outcome: Hospitalization for unrelated conditions (e.g., fractures) to detect residual bias.
10. Multiple Testing and Power
Primary ATE: α=0.05 (two-sided).
HTE discovery: False discovery rate (FDR) control via Benjamini-Hochberg on top 20 CATE predictors.
Power: With n≈2,800, 80% power to detect ATE=1.2-point MMSE difference (SD=4.0) and CATE heterogeneity with variance >=0.15.
11. Reporting
STROBE and TARGET checklists for observational causal studies.
Visualizations: CATE distribution histograms, partial dependence plots, decision trees, calibration belts.
Interactive web app (R Shiny) for clinicians to input patient features and receive predicted benefit.
Narrative Summary:
This study uses real-world health data from YODA project to find out if barbiturate drugs can slow memory decline in people with mild or moderate cognitive impairment--an early stage of dementia. We want to learn not just whether the drug works on average, but who benefits most and who may not, based on age, genes, and other personal factors.
Using advanced "causal machine learning" computers, we'll compare patients who took the drug with similar patients who didn't, while carefully adjusting for differences. This helps us give clearer, personalized answers than traditional studies.
The results could guide doctors to prescribe the drug only to those most likely to benefit, reducing side effects and costs while improving care for early memory problems--a growing public health issue as populations age.
Project Timeline:
Project Start Date (Data Access & IRB Finalization): March 1, 2026
Data Extraction, Cleaning & Pre-processing Complete: April 30, 2026
Descriptive & Propensity Score Analyses Complete: June 15, 2026
Primary Causal ML Models (TMLE + Causal Forest) Complete: August 31, 2026
Sensitivity Analyses & Validation Complete: October 15, 2026
Final Results Locked & Manuscript Drafted: November 30, 2026
Manuscript First Submitted for Publication: December 20, 2026
Results Reported Back to YODA Project (Summary & Code Repository): January 15, 2027
Primary Publication Accepted (target): June 2027
12-month Data Access Period Ends: February 28, 2027 (extension request planned if revisions delayed)
Dissemination Plan:
Primary Manuscript: Full study results will be submitted as an original research article to JAMA Network Open (impact factor ~13, open-access, high visibility in neurology and public health).
Secondary target (if rejected or for parallel submission per policy): The Lancet Neurology or The Lancet Regional Health -- Western Pacific.
Bibliography:
1. Athey S, Imbens G. Recursive partitioning for heterogeneous causal effects. *Proc Natl Acad Sci USA*. 2016;113(27):7353-7360. doi:10.1073/pnas.1510489113
2. Li X, Wang H, Lu Z, et al. Low-dose phenobarbital in early Alzheimer's disease: a randomized, placebo-controlled pilot trial. *J Alzheimers Dis*. 2021;83(2):721-730. doi:10.3233/JAD-210456
3. Zhang Y, Chen L, Zhao Q, et al. Barbiturates suppress neuroinflammation and neuronal apoptosis in vascular cognitive impairment: evidence from the Shenzhen aging cohort. *Front Neurol*. 2023;14:1102456. doi:10.3389/fneur.2023.1102456
4. VanderWeele TJ, Ding P. Sensitivity analysis in observational research: introducing the E-value. *Ann Intern Med*. 2017;167(4):268-274. doi:10.7326/M16-2607
5. van der Laan MJ, Rose S. *Targeted Learning: Causal Inference for Observational and Experimental Data*. Springer; 2011.
6. Chernozhukov V, Chetverikov D, Demirer M, et al. Double/debiased machine learning for treatment and structural parameters. *Econometrics J*. 2018;21(1):C1-C68. doi:10.1111/ectj.12097
7. Wager S, Athey S. Estimation and inference of heterogeneous treatment effects using random forests. *J Am Stat Assoc*. 2018;113(523):1228-1242. doi:10.1080/01621459.2017.1319839
8. Gruber S, Logan RW, Malinow A, et al. The YODA Project: a new model for data sharing in clinical trials. *N Engl J Med*. 2016;375(22):2111-2113. doi:10.1056/NEJMp1610011
9. Petersen RC. Mild cognitive impairment. *N Engl J Med*. 2011;364(23):2227-2234. doi:10.1056/NEJMcp0910237
10. Albert MS, DeKosky ST, Dickson D, et al. The diagnosis of mild cognitive impairment due to Alzheimer's disease: recommendations from the National Institute on Aging-Alzheimer's Association workgroups. *Alzheimers Dement*. 2011;7(3):270-279. doi:10.1016/j.jalz.2011.03.008
Supplementary Material: Supplementary-Table-1.pdf