Supporting Reliable Clinical Prediction Through Interactive Open-Source Tools
Clinical researchers increasingly rely on AI coding tools to develop prediction models, as these tools can substantially accelerate data analysis. However, they also introduce new challenges. While AI can generate thousands of lines of code within seconds, thoroughly reviewing and validating such code is often infeasible for humans within a reasonable timeframe. As a result, prediction workflows may become overwhelming and unnecessarily complex, making it difficult to understand the original analytical intent.
In this talk, we present an R Shiny application developed in collaboration with clinicians. The platform provides an interactive and user-friendly environment for clinical prediction analyses without requiring advanced programming skills. Users upload cleaned datasets, explore visualizations, adjust analysis parameters, and download results directly through the interface.
The application includes key analysis steps such as missing value handling, imputation, model fitting, prediction, and visualization. Users can choose models from the (parsnip) R ecosystem . The interface guides them toward suitable methodological choices depending on the data structure and study objectives. Predictive outputs come with uncertainty assessment through conformal prediction methods. Those provide prediction intervals with statistical coverage guarantees, not only point predictions.
The talk includes a live demonstration on a clinical use case. The presentation is practical and application-oriented, with limited mathematical content.
Attendees will discover how open-source and interactive tools can support transparent, reproducible, and uncertainty-aware prediction workflows in healthcare research. The project also encourages responsible AI use and helps clinicians adopt more reliable statistical practices.
Clinical researchers increasingly use AI-generated code, but these workflows are often difficult to humanly validate. We developed an R Shiny application with clinicians to support transparent and guided clinical prediction analyses.
The platform integrates missing data handling, model fitting, prediction, visualization, and uncertainty quantification with conformal prediction. It helps users perform reliable analyses through an intuitive interface without requiring advanced coding skills.
The project promotes safer, reproducible, and uncertainty-aware AI-assisted statistical workflows in healthcare research.
I am a PhD student in Epidemiology and Biostatistics at the University of Zürich and the University Hospital Zürich within the (Reliable AI Group). My research focuses on reliable and uncertainty-aware AI methods for biomedicine.
I hold an MSc in Statistics from ETH Zürich, where my thesis focused on linear regression with known error distributions, and dual bachelor’s degrees in Applied Mathematics from Paris 1 Panthéon-Sorbonne University and in Social Sciences from Sciences Po Paris.
When I’m not coding in R, I’m usually embroidering or wandering through Swiss pastures looking for cows and statistical inspiration.