Prediction of response to immune checkpoint inhibitors in solid tumors using ct-based biomarkers

Abstract

While the number of diagnosed cancer patients continues to rise, so does the availability of treatments. The latest breakthrough has been immunotherapy, particularly immune checkpoint inhibitors (ICIs), which aims to enhance the body’s natural defenses to eliminate malignant cells. However, cancer cells have developed multiple mechanisms to evade immune recognition, often resulting in unconventional tumor response patterns that should be addressed to improve patient care. Current FDA-approved biomarkers for predicting immunotherapy response (i.e, PD-L1 expression, microsatellite instability [MSI], mismatch repair system deficiency [dMMR] and tumor mutational burden [TMB]) have limited efficacy and require invasive tumor biopsies. An alternative promising approach is radiomics, which involves extracting quantitative image features from medical scans such as Computed Tomography (CT). Radiomics is particularly promising because it is non-invasive and takes advantage of widely used and established imaging methods. In this study, the aim was to predict clinical benefit in patients treated with immunotherapy using CT-based radiomic features. For this purpose, two sub-objectives were defined: to find the optimal machine learning (ML) pipeline for predicting immunotherapy response, and to determine the additional predictive power that radiomics adds to clinical data alone. The cohort data analyzed consisted of 185 patients treated with ICIs at Hospital de la Vall d’Hebron with different primary tumor types. The pipeline included four different feature aggregation methods to combine features from lesion level to patient level, three feature selectors to find the most predictive features and potentially identify biomarkers, and three classifiers. The best results were obtained using the weighted average of the three largest lesions (feature aggregation), LASSO (selector), and logistic regression (classifier). Furthermore, better results were obtained when implementing the same pipeline on both clinical and radiomics data combined (AUC: 0.8 ± 0.05), confirming the potential of radiomics for predicting immunotherapy response. In this case, the features with the most predictive power, whose SHAP values were carefully analyzed, were flatness, number of baseline affected organs, surface volume ratio, size zone non-uniformity normalized, and short run emphasis. The results indicate the promising approach that radiomics holds. However, several limitations must be addressed before it can be routinely incorporated into clinical practice. These include improving biomarker accuracy, automating the image analysis process, and integrating these pipelines into radiology workflows.