New University of Cincinnati Cancer Center research published in the Journal of Statistical Theory and Practice describes the potential for a more robust and accurate breast cancer survival model using a combination of statistical methods and a deep learning framework that could give physicians a better tool to make personalized care decisions for each patient.
How does the new survival model differ from current approaches?
Each case of breast cancer is unique and depends upon many variables, including clinical, biological and genetic factors. Traditional survival models are useful to estimate each patient's risk and expected outcomes, but they cannot fully account for how the complex relationships between these factors affect survival prediction for a given patient.
"We wanted to explore whether combining statistical methods with deep learning could capture these relationships and improve the overall survival prediction," said corresponding author Wedad Alatebi, a doctoral student in the Department of Biostatistics, Health Information and Data Science in UC's College of Medicine. "Deep learning can be described as teaching a computer to recognize complex patterns in patient data."
Coauthor Anand Seth, PhD, said the "fusion" strategy combining multiple statistical approaches allows a much more precise prediction outcome. The deep learning framework allows patients who are seemingly similar to be stratified into more specific risk groups.
"That way, you can individualize therapy for a specific group of patients," said Seth, research manager in the Department of Biostatistics, Health Information and Data Science in UC's College of Medicine.
How does the model account for genetic information?
Tumor genetic information can improve survival models, but genetic testing is not always feasible or cost-effective in hospitals worldwide. To account for this, Alatebi and her team created simulated gene expression data that is built into the new model.
"We first transform the clinical information into a better representation of patients' risk," she said. "We then simulate biologically plausible molecular features to explore how integrating molecular and clinical information may improve survival prediction. The model also considers similarities between patients and includes uncertainty in its predictions."
How did the new model perform on a real-world dataset?
The team tested the new model on a publicly available dataset of 1,980 breast cancer patients' clinical and survival data. In this dataset, the model achieved higher accuracy at predicting survival outcomes than compared models in the study.
"It was exciting to see how much information we could gain by combining different approaches rather than relying on one method alone," Alatebi said.
While more research is needed to determine the model's precise accuracy, Seth said even being a few percentage points more accurate could provide physicians with more confidence in the treatment strategies and prognoses for their patients.
"This is a personalized medicine approach," added coauthor Shesh Rai, PhD, director of the Cancer Center's Biostatistics and Informatics Shared Resource and professor in the Department of Biostatistics, Health Information and Data Science in UC's College of Medicine. "We're either using or simulating the patients' gene signatures in addition to other features, other demographics, combining all of this together, and from there predicting what could actually happen."
What are the next steps for the research?
Alatebi and the team will further evaluate the model using larger and more diverse patient datasets. Their long-term goal is to develop a tool that could help physicians to better identify patient risk and support more personalized treatment and follow-up decisions.
"These deep learning-based models are not meant to replace physicians," she said. "They are tools that could help physicians understand complex patient information and support their decisions. This combination of statistics, machine learning and cancer research has been exciting for me to help us improve how we understand patient outcomes."
Other Cancer Center collaborators on this research include Syed Ahmad, MD, and Marepalli B. Rao, PhD.
