bayesian-nma-framework

Custom Python implementation for the Bayesian network meta-analysis presented in:

Comparative Efficacy and Safety of Perioperative Enfortumab Vedotin-Pembrolizumab, Dose-Dense MVAC, and Gemcitabine-Cisplatin-Durvalumab in Muscle-Invasive Bladder Cancer: A Bayesian Network Meta-Analysis

A lightweight framework for Bayesian Network Meta-Analysis (NMA) using Metropolis-Hastings MCMC sampling, designed for trial-level aggregate data in star network topologies (common comparator design). No dependency on external Bayesian packages (PyMC, Stan).

Features

• Bayesian NMA with fixed-effects and random-effects models
• Bucher indirect treatment comparison (ITC) for HR, OR, and RR
• SUCRA treatment rankings from posterior samples
• Pseudo-IPD reconstruction from published Kaplan-Meier curves (Guyot et al. 2012)
• Weibull parametric fitting for trials without digitizable KM data
• Cox proportional hazards on reconstructed IPD
• Logistic regression for binary endpoint IPD-derived ORs
• Convergence diagnostics: Gelman-Rubin R-hat, effective sample size (ESS)

Installation

pip install -r requirements.txt

Quick start

import numpy as np
from src.nma_engine import prepare_nma_data, run_nma_mcmc, nma_summary, compute_sucra

# Define trials (each experimental arm vs common comparator)
trials = {
    'Trial_A': {
        'hr_efs': 0.60, 'hr_efs_lo': 0.45, 'hr_efs_hi': 0.80,
        'exp_label': 'Drug X',
    },
    'Trial_B': {
        'hr_efs': 0.75, 'hr_efs_lo': 0.58, 'hr_efs_hi': 0.97,
        'exp_label': 'Drug Y',
    },
}

# Prepare and run NMA
treatments = ['Control', 'Drug X', 'Drug Y']
nma_data = prepare_nma_data(trials, endpoint='efs', treatments=treatments)

np.random.seed(42)
d_samples, tau, trts = run_nma_mcmc(nma_data, n_iter=50000, burnin=10000)

# Results
print(nma_summary(d_samples, trts, metric='HR'))
print(compute_sucra(d_samples, trts, higher_is_better=False))

See examples/example_star_nma.py for a complete working example with 3 trials.

Modules

Module	Description
src/nma_engine.py	MCMC sampler (FE/RE), NMA summary, SUCRA, R-hat, ESS
src/itc.py	Bucher ITC for HR, OR, RR; risk difference; NNH
src/ipd_reconstruction.py	Guyot algorithm, Weibull fitting, Cox PH, logistic regression

Statistical models

Fixed-effects NMA

y_k ~ N(d[trt_k], se_k^2)
d ~ N(0, 100)          # vague prior

Random-effects NMA

y_k ~ N(delta_k, se_k^2)
delta_k ~ N(d[trt_k], tau^2)
d ~ N(0, 100)          # vague prior
tau ~ HalfNormal(0, 0.5)

MCMC configuration

• Metropolis-Hastings with Gaussian proposals (SD = 0.05)
• 4 independent chains
• 50,000-80,000 iterations, 10,000-20,000 burn-in
• Convergence: R-hat < 1.05, ESS > 1,000

Registration

• OSF Preregistration: https://osf.io/7un2d/
• PROSPERO: [Registration ID pending]

References

• Dias S et al. Stat Med. 2010;29(7-8):932-944.
• Salanti G et al. J Clin Epidemiol. 2011;64(2):163-171.
• Bucher HC et al. J Clin Epidemiol. 1997;50(6):683-691.
• Guyot P et al. BMC Med Res Methodol. 2012;12:9.

Citation

If you use this framework, please cite:

Subiela JD, Gonzalez-Padilla DA, Villacampa F, et al. Comparative Efficacy and Safety of Perioperative Enfortumab Vedotin-Pembrolizumab, Dose-Dense MVAC, and Gemcitabine-Cisplatin-Durvalumab in Muscle-Invasive Bladder Cancer: A Bayesian Network Meta-Analysis. [Journal pending]. 2026.

Related

• Study-specific analysis (data + results): Available upon publication at github.com/jdsubiela/EV-P-vs-MVAC-vs-GC-Durva

License

MIT License. See LICENSE.