Identifying pattern and linking prominent factors in complex infection data
Although the overall survival rate among pediatric oncology cases has risen, difficult cases remain. As patients undergo significant treatment for complex cases, the chance for complications increases. Among the most concerning complications that can arise is infection. To address and assess infection risk among pediatric oncology patients, our group works to identify prominent patterns and factors that can affect the course of disease and link those with patient outcomes.
Infection or pathogen-related complications pose complex issues for pediatric oncology patients. Because of the immunocompromised state of these patients, there is an increased risk of infection that affects care in two dimensions: immediate treatment and long-term outcomes. We analyze infection data in a variety of contexts to help investigators and clinicians best understand how pathogens plays a role in disease progress and ultimate patient outcomes.
As we continually assess factors that impact patient outcomes, we participate in the human microbiome project. Information from the gut microbiome offers value in relation to the whole picture of a patient, which was missed before. To fill in these gaps in understanding, we analyze sequenced samples of oncology patients collected at different stages during and after treatment, to generate a spectrum that details information about bacteria in the gut. Because we now know the gut microbiome plays a significant role in the health of a person, we are working to understand its relation to disease outcomes, treatment response, post-treatment complications, and even some long-term response (3-5 years) in pediatric oncology patients.
Because these samples are collected at different times in a typical longitudinal setting, the challenge comes when we try to view all the data from every patient at every timepoint. This is a large, high-dimension map constructed in a dynamic process, so we are working to develop new methods that help us process the data in a statistically rigorous and computationally efficient manner. Our goal is to make this map useful for predicting treatment response, disease complication, or long-term outcomes.
Another area in which we focus on methodology development is in the creation of a machine learning method for bone marrow transplant patients. Our aim with this project is to develop a computational machine learning pipeline to predict post-transplant immediate mortality (post-100 days) and the long-term survival rate (1-3 years) after transplant. Our guiding question is what factors play a role in the differentiation of mortality outcomes amongst patients?
Because we are using patients treated at St. Jude in the past ten years who had large amounts of lab data collected, this is an immense dataset that current machine learning methods don’t have the capability to process. To develop an adequate machine learning method, we build upon existing methods that can track the longitudinal pattern of change within patients. We are also working to figure out how to address missing clinical data within large sets of longitudinal data with this new machine learning method.
Outside of our oncology focus, we provide biostatistical support to study the short- and long-term immune response associated with COVID within the SJTRC study. Through this study, we hope to answer key unanswered questions regarding vaccine response and immunology profiles of St. Jude employees who volunteered to be part of this study. While this work does not directly relate to oncology, it is important to understand how they body reacts to an invading pathogen, what different types of cells do in that process, and how this can help us understand future response to vaccines among normal and immunocompromised populations.
In this area of work, we employ existing methods and develop new methods to evaluate real-world effectiveness and immunology profiles. The challenge in evaluating real-world effectiveness originates from the fact that we merely observe a multitude of factors, and we cannot control subject actions or changes. There are many factors that we must account for in the model, including the temporally changing baseline risk that occurs due to real-world factors (such as Omicron). This means we must constantly adjust our model and take the dynamic temporal trends into account, which is something not currently done in existing methods or literature. Our goal is to develop dynamic effectiveness prediction methods that can estimate and project real-world situations in a more reliable way.
Our collaborations with the Departments of Infectious Diseases, Pathology, Bone Marrow Transplantation and Cellular Therapy, and Hematology are defining partnerships in our group’s work. These collaborations allow us to provide biostatistical support for ongoing laboratory and clinical trial efforts, and the questions that arise in these collaborations drive our methodological research and development. At the core of all our work is our desire to illuminate infection-related outcomes in pediatric oncology patients.