Clinical-Trial-Data-Curation-Bayesian-Network-Meta-Analysis-HbA1c-Week-40-52-

Conducted structured extraction of Phase 3 RCT data to evaluate comparative HbA1c reduction of GLP-1 therapies at ~40 weeks. Harmonized endpoints, converted CIs to standard errors, and built analysis-ready datasets. Implemented a Bayesian network meta-analysis in R and documented estimand and population assumptions for transparency.

#Overview

This project demonstrates a structured pharmaceutical evidence synthesis workflow using the SURPASS-2 randomized clinical trial.

The objective was to extract, curate, and structure clinical trial data in a format suitable for Bayesian network meta-analysis (NMA), consistent with industry practices used in competitive intelligence and statistical evidence synthesis teams.

The project includes:

• Systematic data extraction from published RCT sources
• Estimand-aware endpoint harmonization
• Creation of meta-analysis-ready datasets
• Bayesian fixed and random-effects NMA using gemtc
• Transparent audit trail documentation

#Research Question

What is the comparative efficacy of tirzepatide versus semaglutide (and ultimately liraglutide) in reducing HbA1c at approximately 40–52 weeks in adults with Type 2 Diabetes?

#PICO Framework

Population
Adults (≥18 years) with Type 2 Diabetes inadequately controlled on metformin.

Intervention
Tirzepatide 5 mg, 10 mg, 15 mg (once weekly).

Comparator
Semaglutide 1 mg (once weekly).

Outcome
Change in HbA1c from baseline (%-points).

Timepoint Rule
Closest to 40–52 weeks (SURPASS-2 primary endpoint = Week 40).

#Datasets arms.csv Arm-level dataset containing:

• Study and trial identifiers
• Treatment arms and doses
• Analysis population
• HbA1c mean change at Week 40
• Estimand classification
• Audit source location

relative_effects.csv Contrast-level dataset containing:

• Mean differences vs semaglutide
• 95% confidence intervals
• Derived standard errors
• Direction consistency check

Standard errors were calculated using: SE = (CI_high − CI_low) / (2 × 1.96)

#Analysis Approach

Bayesian network meta-analysis was implemented using the gemtc package.

Models run:

• Fixed-effect consistency model
• Random-effects model (demonstration only; heterogeneity not identifiable with a single study)

Convergence was evaluated using:

• Gelman-Rubin diagnostics
• Effective sample size
• Traceplots
• Only one multi-arm RCT (SURPASS-2) was available.
• Multi-arm covariance was not explicitly modeled due to unavailable variance structure.
• Arm-level SD values were not available in the extracted excerpt.
• Random-effects model is prior-driven due to single-study network.

This project demonstrates workflow capability rather than definitive pooled estimates and core competencies required for evidence synthesis and statistical co-op roles:

• Clinical trial data mining
• Endpoint and estimand interpretation
• Structured dataset harmonization
• Bayesian modeling
• Transparent documentation
• Reproducible analytical workflow

The structure is scalable to multi-trial network meta-analysis including semaglutide and liraglutide trials at 40–52 weeks.

#Technologies Used

• R (statistical computing)

• tidyverse (data wrangling)

• Bayesian modelling framework in R (for NMA)

• Reproducible workflow (scripts / R Markdown outputs) #What I Learned

• How to convert real-world trial reporting into standardized, analysis-ready structures

• Practical challenges in endpoint parsing and cross-trial comparability

• Building a reproducible pipeline in R for evidence synthesis

• Core concepts behind Bayesian network meta-analysis and interpreting outputs

• Importance of documentation and assumptions in regulated/clinical contexts******

#How I Built It

• Designed a curation template to capture trial identifiers, arms, and HbA1c endpoints
• Implemented endpoint harmonization logic for a consistent Week 40–52 window
• Wrote R scripts to transform curated inputs into analysis-ready datasets
• Implemented the Bayesian NMA workflow and generated reproducible summaries