Author: James Brugarolas, Meixu Chen, Raquibul Hannan, Payal Kapur, Jing Wang, Kai Wang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, Department of Radiation Oncology, University of Maryland Medical Center 🌍
Purpose: Accurate prognosis of clear cell renal cell carcinoma (ccRCC) is essential for guiding personalized treatment planning. In this study, we present a multi-modal ensemble model (MMEM) that integrates clinical data, multi-omics data, and histopathology whole slide image (WSI) features to predict overall survival (OS) and disease-free survival (DFS) of ccRCC patients. The MMEM leverages complementary insights across data types, including WSI features extracted using a pre-trained computational pathology foundation model, to improve prediction accuracy.
Methods: We analyzed 226 ccRCC patients from The Cancer Genome Atlas (TCGA) Kidney Renal Clear Cell Carcinoma (KIRC) dataset, each with OS and DFS follow-up data. The data included clinical information, three multi-omics types (mRNA expression, miRNA expression, and DNA methylation), and WSI data. Cox proportional hazards (CPH) models with iterative forward feature selection were constructed for clinical and multi-omics data. WSI features were extracted using a pre-trained computational pathology foundation model (UNI) and three other pre-trained encoders. Deep learning-based CPH models were trained using WSI features. The MMEM integrated risk scores from the five data modalities, assigning weights based on training performance. Five-fold cross-validation was performed for model evaluation.
Results: The MMEM outperformed single-modality models across all prediction labels. It achieved concordance indices (C-index) of 0.82 for OS and 0.83 for DFS. Binary prediction at a 3-year follow-up yielded AUROC values of 0.83 for patient death and 0.86 for cancer recurrence. Among WSI-based predictions, features extracted using the UNI foundation model demonstrated the best performance.
Conclusion: We developed the first multi-modal prediction model MMEM for ccRCC patients that integrates features across five different data modalities. Our model demonstrated better prognostic ability compared with corresponding single-modality models for both prediction targets. Independent validation is ongoing and will be presented. If findings are independently reproduced, it has the potential to assist in management of ccRCC patients.