NEW YORK – Using clinical, molecular, and radiomic markers, City of Hope researchers have built a clinically relevant model to help predict if CAR T-cell therapies are likely to extend the lives of high-grade glioma patients.
At the American Society of Clinical Oncology's virtual annual meeting on Sunday, Alec Wong, lead data scientist at City of Hope, described how he and his colleagues at the institution's National Medical Center and Comprehensive Cancer Center have been "using data science approaches to tackle scientific problems in translational cancer research and precision oncology."
Specifically, Wong presented data from a Phase I clinical trial on 60 HGG patients, who underwent surgical resection and CAR T-cell administration in the resected cavity. The researchers obtained baseline data points from patients before surgery and CAR T-cell treatment, aiming to identify the biological features that could be incorporated into a model predictive of their survival.
HGG is an aggressive, heterogeneous primary neoplasm of the central nervous system, with high recurrence rates and poor survival. Within multiple ongoing clinical trials, researchers are trying to develop targeted molecular and immunotherapeutic agents such as pembrolizumab (Merck's Keytruda) and CAR T-cell therapy that can improve survival.
Certain predictive models have shown value in identifying biomarkers predictive of treatment response and prognosis. The researchers were aiming to develop a model that could have an impact in the clinic, and which could potentially optimize clinical trial enrollment through more precise patient screening and treatment planning.
They performed brain tumor segmentation on post-contrast MRIs and calculated the tumor volume, surface area, and sphericity. They then created a 12-factor model, combining these radiomic features with clinical data taken from electronic medical records, including age, gender, race, and ethnicity, and molecular features from pathology reports, such as tumor histology, grade, and location, the presence of unifocal or multifocal lesions, and the level of IL-13Rα2 expression.
With this predictive modeling, the researchers were hoping to solve a binary classification problem by sorting patients into two cohorts — those who had survived for more than 180 days and those who had not, Wong said. The model was designed to generate a numeric survival score of 0 to 1, with a higher score indicating a better probability of survival. They also designed the model to stratify patients around the median risk score, such that it became a split point for risk stratification.
After analyzing the data, they found patients with higher tumor surface area, tumor volume, and age had lower survival scores, while patients with higher IL-13Rα2 expression and tumor sphericity had higher survival scores. After further narrowing that list down to the top three features related to overall patient survival, they found that higher tumor surface area, and patient age were negatively associated with survival, while higher expression of IL-13Rα2 was associated with a better survival probability, Wong said.
"We have developed an explainable machine learning model to predict overall survival for high-grade glioma patients treated with CAR T-cell therapy," he explained. "We were able to identify the [factors] contributing to model predictions."
Since in this study, Wong and colleagues looked specifically for variables that were clinically relevant, their findings have the potential to be used in clinical settings, he added. Wong cautioned, however, that the study was limited by a small sampling of patients from a single institution, and that the data would have to be validated in a larger and more ethnically diverse cohort.