Validation of Data Exploration Apps
Discussion summary from the inaugural R/Pharma EU Summit at Novartis in Basel on October 5, 2026, alongside BioTechX.
See also: 2023 Discussion, 2024 Discussion
Context and Intended Use
The discussion focused on Shiny apps for clinical data exploration in regulated settings. The two sessions differed sharply: the first assumed validation was essential, while the second questioned why it mattered at all.
Validation should follow risk and intended use. Exploratory apps may not need validation, but requirements change when outputs support decisions, investigations, or regulated analyses.
Lifecycle and Reproducibility
- Long-running apps need governance: one company described an app running for seven years. Copying files into folders is not adequate version control.
- Control the complete analytical unit: the app, R packages, data, and preprocessing all matter for reproducibility.
- Containers: containerisation could provide a scalable foundation. Posit’s support for containerised Shiny apps could help standardise infrastructure and its qualification.
- Reusable validated frameworks: companies described validated R packages for outputs, likely tables, listings, and graphs (TLGs), built on ADaM datasets that could support Shiny apps.
Testing and Delivery Challenges
Quality control remains manual and labour-intensive: review, visual inspection, double programming, and screenshot comparisons against exported results.
An alternative was to treat the entire app as an R package, apparently using a golem-style approach. Define expected functionality, attach tests to features, and systematically check outputs. Participants suggested this could substantially reduce validation time.
Interactive outputs bring additional difficulties. Safety-signal views, spaghetti plots, and 3D visualisations may not reproduce like static reports. Conventional software validation can take months, while business or clinical decisions may be needed within weeks.
Changing Data and AI
A validated app may need controlled data snapshots. Each data update must be shown not to break the app, effectively requiring validation of the data as well as the software.
AI might eventually help with validation, but its models and guardrails would need validation too. Using one model to validate another raises questions about trust, skills, and quality.
Proposed Way Forward
Fully validating interactive Shiny apps remains difficult and can be impractical. Participants proposed combining:
- A risk-based approach and clearly defined intended use.
- Controlled containers and validated packages.
- Automated testing.
- Controlled version management.
- Reusable data snapshots.