Biostatisticians make clinical trial data more dynamic to gain deeper insights

Creating a new treatment can be a grueling process. It can take decades for a discovery to make its way from the lab to a phase 3 clinical trial, the final hurdle before reaching patients. Phase 3 trials are the largest and compare an experimental treatment to the current standard, providing the final evidence of whether the drug is superior before gaining widespread acceptance. In rare diseases, such as pediatric cancer, getting enough patients on a clinical trial to perform statistical analysis is challenging. Therefore, scientists seek ways to get more insights out of the data they do have.

Recently, St. Jude researchers created a statistical package that gives deeper insights into clinical trial data, which they published in JCO Precision Oncology. The software is designed for a standard statistical program, R, which has a special framework for building interactive applications, R Shiny, and uses Cox regression models (a statistical tool used in survival analysis). This led the team to name the software ShinyCox.

“We’re always striving to make sure we get the most out of what we learn from our patients in clinical trials,” said first and corresponding author Subodh Selukar, PhD, Department of Biostatistics. “ShinyCox takes something that already falls out of our normal trial analysis, but is usually not published, and presents it in a dynamic and easy-to-understand format.”

Looking beyond traditional hazard ratios

Traditional clinical trial reporting frequently includes something called a hazard ratio. When scientists want to know about the relative impact of a new treatment, they may estimate different ratio statistics between the treatment and control groups. The hazard ratio summarizes the relationship between two groups in the context of a time-to-event outcome, such as survival. A hazard ratio number further away from one indicates a larger treatment effect between the treatment group compared to the control group, capturing the most essential information in a single number. In most cases, a table full of hazard ratios is also segmented by specific patient groupings, such as age, sex and race.

While the hazard ratio is useful, it is also static. Contextualizing the numbers can be difficult, especially in understanding the complex interactions for patients belonging to multiple groupings. However, a common statistical model for the hazard ratio, a Cox regression model, also contains that complex information, though it has been too difficult historically to include in physical copies of a published study or share between investigators.

“We realized we could use more information from these models than what we typically present,” said senior author Stan Pounds, PhD, Department of Biostatistics. “We could use Cox models, which use the same data as hazard ratios, to predict survival functions for different patient groups. Then we could make a package that does so automatically and presents the data in a dynamic format that is more clinically useful.”

With ShinyCox, scientists can put their clinical trial data into the application to see more complex information about participants. They can select which segmentations to look at, such as a specific age range and gender, then see predicted survival curves for that particular cross-grouping. While not intended for clinical practice, the package promises to provide investigators new hypotheses from this more complex data, which may give novel insights into treatment. That knowledge could galvanize new or retrospective trials that explicitly test those insights, ultimately leading to more personalized care and better patient outcomes.

Finding more data in acute myeloid leukemia trials as a proof of principle

Before creating ShinyCox, Pounds analyzed pediatric acute myeloid leukemia (AML) clinical trial data to find a way to predict disease risk and treatment outcomes. Children with AML have a five-year survival rate of 60-70%, indicating a need to improve their odds. One way to do so is to proactively identify predictors of treatment efficacy, enabling physicians to choose the best possible treatment for an individual. Pounds’s group created ACS10, a pharmacogenomic score that uses a patient’s DNA sequence to determine their likely response to different treatments.

In the new study, the St. Jude researchers verified that ShinyCox could help provide new insights by looking at several historical AML clinical trials. While ACS10 alone performs well, ShinyCox identified age and sex as important predictors that can add to the pharmacogenomic score, improving projections and altering some conclusions about the best treatments for certain subgroups.

“While age is well-known to be important in cancer treatments, we found a novel interaction between the ACS10 pharmacogenomics score and age,” Selukar explained. “It demonstrates ShinyCox’s ability to incorporate added information to build a fuller picture of patients, instead of just looking at one factor in isolation.”

Creating a user-friendly way to analyze clinical trial data

The St. Jude group created the ShinyCox application in a unique way. Every year, St. Jude hosts a BioHackathon to provide intense focus on a specific computational problem for a few days. Pounds realized there was more data to pull out of clinical trials, but creating the computational programming to do so would be substantial. Therefore, he led a BioHackathon team to produce a prototype, which was further developed by biostatisticians, such as Selukar, into ShinyCox until it worked well and was easy to use. The highly collaborative team also included a Biostatistics summer intern student, other members of the Biostatistics department, members of the AML team and external collaborators.

“It was really exciting to create something that can enhance what we learn from St. Jude pediatric oncology data and what the greater biomedical community can learn from their data,” Selukar said. “We made it feasible for anyone to implement, potentially something like 10 or fewer lines of code, making it as accessible as possible so people can maximize the learnings from their trial.”

The program uses a simple user interface, where a scientist can select the subgroups they wish to view. It then generates their predicted survival curves based on the clinical trial data provided. The statistical analysis is performed in the background, and the program can be hosted on a website, making the process accessible to even those without a strong coding or statistics background. The St. Jude researchers still recommend having a biostatistician work with a user for further analysis, as there may be subtleties of each clinical trial that the program does not consider. Still, it serves as a starting place to understand trial data better, getting more out of the information the scientists are already collecting.

“ShinyCox is not built to replace hazard ratios or traditional analysis, but rather to augment them,” Pounds concluded. “Instead of only looking at static tables, we have given the biomedical community a tool to better see how patient groups might differ, providing investigators a far more nuanced understanding of their clinical trials and ultimately empowering them to make discoveries that will push forward personalized therapy research that creates better treatment outcomes for patients.”