This repository contains the Bayesian-ECD Shiny application, a local graphical interface for applying the Bayesian-ECD historical-borrowing workflow to a current randomized clinical trial (RCT). The application provides posterior survival and treatment-effect estimation, restricted mean survival time (RMST) summaries, prespecified subpopulation analyses, prediction in new patient populations, and an ECD-compatibility test for trials containing an RCT-control arm.
The application runs locally in R and uses stored posterior information from the historical Bayesian-ECD model. Current RCT data therefore do not need to be uploaded to a separately hosted Bayesian-ECD service.
For complete installation instructions, platform-specific setup, data requirements, analysis workflows, interpretation guidance, output descriptions, reproducibility recommendations, and troubleshooting, see:
A rendered PDF version of the user guide is provided separately with the journal supplementary materials.
The user guide is the primary reference for using the application. The example below is intended only as a quick orientation to the workflow.
The example datasets included with the repository can be used to walk through the main Bayesian-ECD workflow. The screenshots below illustrate one example analysis using Control as the reference treatment and Drug A as the comparison treatment.
After completing the installation steps in the user guide, open:
app/app.R
in RStudio and select:
Run App > Run in Window
Alternatively, from the repository root, run:
shiny::runApp("app")Optionally, set a random-number seed before launching the application:
set.seed(12345)
shiny::runApp("app")The Uploaded Data tab is the starting point for the primary Bayesian-ECD workflow.
Use:
example_data/example_model_data.csv
In the Uploaded Data tab, enter the column names corresponding to the patient ID, overall survival, censoring indicator, treatment, cohort, age, sex, KPS, and extent of resection (EOR), and then select the example CSV file.
For the censoring indicator, TRUE denotes an observed event/death and FALSE
denotes a censored patient/alive at last follow-up. The application supports
one current trial dataset at a time, so the cohort column should contain a
single cohort value. The categorical variables must use the
labels expected by the stored model: Female and Male for sex; > 80,
(60, 80], and <= 60 for KPS; and GTR, STR, and biopsy for EOR.
Click Submit. The uploaded data are displayed in the application so that the selected variables and their values can be reviewed before proceeding.
Detailed variable definitions, coding requirements, and input-format guidance are provided in the user guide.
When the uploaded trial contains an RCT-control arm labeled exactly Control,
the application evaluates the ECD-compatibility test using the current
compatibility settings. With the example dataset, the compatibility analysis
begins automatically after the uploaded data become available.
The test compares the observed RCT-control survival experience with covariate-standardized Historical-Control predictions from a Stage 2 Bayesian-ECD fit. Historical-Control potential outcomes are generated across 1,000 retained posterior draws, producing a posterior distribution of log-rank chi-square statistics. The mean posterior log-rank statistic is then calibrated against a nonparametric bootstrap distribution generated from the observed RCT-control patients.
The default settings use a plotting range of 150 to 1,200 days, 500 bootstrap samples, and a significance level of 0.05. These settings can be changed in the sidebar and the test rerun by clicking Submit.
The displayed summary reports the RCT-control sample size, mean posterior log-rank statistic, bootstrap p-value, number of bootstrap samples, significance level, and resulting compatibility assessment. The plot can be downloaded as a PNG, and detailed compatibility results can be downloaded as a ZIP archive.
If the uploaded dataset does not contain an RCT-control arm labeled Control,
the compatibility test is unavailable, but the remaining Bayesian-ECD analyses
can still be performed.
The Plot Output tab provides posterior survival-probability and time-varying hazard-ratio summaries. Select a reference treatment, comparison treatment, plot type, and time range, and click Submit.
For example, the analysis below compares Drug A with Control using posterior survival probabilities from 150 to 1,200 days.
The displayed plot can be downloaded as a PNG. The underlying Bayesian-ECD results can also be downloaded as a ZIP archive for additional analysis.
The RMST Output tab summarizes restricted mean survival time for a selected treatment comparison and time horizon. The example below compares Drug A with Control at a horizon of 730.5 days (2 years).
The output displays treatment-specific RMST curves together with posterior summaries of the RMST difference and RMST ratio. The combined RMST figure can be downloaded as a PNG.
The Subpopulation Analysis tab repeats the survival or hazard-ratio analysis within a user-defined subset of the uploaded RCT population. Subpopulations can be defined using sex, EOR, KPS, and age.
For categorical variables, leaving a selection empty leaves that variable unrestricted. Selecting all available categories likewise does not restrict that variable. For age, the lower bound is included and the upper bound is excluded.
The example below compares Drug A with Control among Female patients aged
48 to <81 years, with EOR and KPS unrestricted. In interval notation, the age
selection is [48, 81). The resulting plot also reports the number of patients
meeting the selected criteria.
The Subpopulation RMST Output tab uses the treatment comparison and subpopulation defined in the Subpopulation Analysis tab. Specify the RMST horizon and click Submit. The example below uses a 730.5-day horizon.
The resulting RMST table can be downloaded as a CSV file.
Bayesian-ECD can also generate treatment-specific predictions for a new patient population with a different covariate distribution. Use:
example_data/example_prediction_data.csv
in the Prediction Data tab.
Unlike the model-fitting dataset, the prediction dataset requires only a patient ID and the baseline covariates used by the model. It does not require observed survival outcomes, censoring indicators, treatment assignments, or cohort labels.
After entering the prediction patient-ID column and selecting the CSV file, click Submit to make the prediction population available to the prediction tabs.
The Prediction Output tab generates posterior survival-probability or time-varying hazard-ratio summaries standardized to the covariate distribution of the uploaded prediction population.
By default, all patients in the prediction dataset are used. Selecting Apply subpopulation filters to prediction dataset exposes optional sex, EOR, KPS, and age filters. As in the RCT subpopulation analysis, the minimum age is included and the maximum age is excluded.
The example below compares Drug A with Control for the complete example prediction dataset over 150 to 1,200 days.
The plot reports the number of prediction patients included in the analysis and can be downloaded as a PNG.
The Prediction RMST Output tab uses the same treatment comparison and prediction cohort or subpopulation selected in the Prediction Output tab. Specify an RMST horizon and click Submit.
The example below uses the complete example prediction population and a 730.5-day horizon.
The table reports posterior RMST summaries for the two treatments together with their difference and ratio and can be downloaded as a CSV file.
Additional posterior-probability utilities are available outside the Shiny interface for analyses such as:
- posterior probabilities for prespecified hazard-ratio thresholds;
- posterior probabilities for RMST-difference thresholds; and
- posterior probabilities for RMST-ratio thresholds.
The utility functions are located in:
analysis_utils/posterior-probability-utils.R
A worked example is provided in:
examples/run_posterior_probabilities.R
From the repository root, the example can be run with:
source("examples/run_posterior_probabilities.R")Generated posterior-probability results are written to the Git-ignored:
outputs/
directory.
The main repository directories are:
app/: Shiny application files.app_core/: Core modeling, plotting, RMST, subpopulation, and C++ routines used by the application.app/data/storedMCMCiter/: Stored Stage 1 Bayesian-ECD posterior components used by the application.analysis_utils/: Optional analysis utilities that are not part of the Shiny interface.examples/: Worked R examples for optional analyses.example_data/: Example model-fitting and prediction datasets.docs/: Bayesian-ECD Shiny User Guide documentation.