Case Study: Building and validating a composite biomarker model to predict emphysema progression
KerusCloud is a ground-breaking new clinical study design and analytics software platform which delivers smarter real-time studies for today’s clinical research challenges.
Using powerful cloud-based processing, KerusCloud can handle the diverse and complex data now collected routinely, to deliver advanced analytics which simplify the study planning and decision-making process.
With unique second-generation study simulation capabilities, KerusCloud provides exceptional support in developing robust evidence packages for drug approval.
- The ability to identify patients with worsening of emphysema is important for selecting the right population in COPD studies.
- Working in collaboration with GSK COPD Clinical Discovery, the GSK Eclipse study was used as a basis for building a predictive model involving multiple baseline biomarkers.
- There is no standard approach for designing a validation study for a multivariate model.
- The ability to build and validate a model was limited by the number of available samples.
- Could the simulation capabilities of KerusCloud be used to design the study given the constraints?
Building a predictive biomarker model with KerusCloud
The biomarker and clinical characteristics as well as the correlation structure was extracted from Eclipse data.
Validation comprised multiple study objectives relating to the performance of individual markers as well as the overall model.
A range of study designs was evaluated for their ability to achieve multiple objectives and a full model (8 biomarkers) and a reduced model (5 biomarkers) were assessed.
Various metrics were derived from 1000 simulated studies including AUC (of ROC curve) for model performance and p-values for individual biomarkers.
Kerus Cloud generated realistic data which could be used to evaluate the overall performance of the model and individual markers. The full model required > 200 samples whilst the reduced model required 60 samples.
- The study demonstrated that KerusCloud could generate realistic data in silico and that it is feasible to use KerusCloud to design validation studies for multivariate predictive models.
- It showed that we could optimize the model building and validation strategy given the available samples.
- The strategy involving the reduced model required 70% fewer samples for the same probability of success.
- Avoided the risk of attempting a study with a low chance of success.
“We were extremely pleased with the results of the work. Until Exploristics became involved we were struggling to come to conclusions about the utility of the biomarkers being evaluated. The results with KerusCloud provided our team with a solid understanding of the data allowing clear conclusions to be drawn.” Senior Scientific Director, Respiratory TAU, GSK