A multimodal deepsurv approach: integrating radiomics and clinical factors for brain metastasis survival prediction.
Authors
Affiliations (5)
Affiliations (5)
- Department of Neurosurgery, Mount Sinai Health System, New York, NY, United States of America.
- Sidney Kimmel Medical College, Thomas Jefferson University, Philadelphia, PA, United States of America.
- Department of Radiology, Hospital of the University of Pennsylvania, Philadelphia, United States of America.
- Department of Neurological Surgery, University of Miami Miller School of Medicine, Miami, United States of America.
- Department of Neurosurgery, Mount Sinai Health System, New York, NY, United States of America. Electronic address: [email protected].
Abstract
Predicting survival outcomes for brain metastasis (BM) patients is crucial for tailoring treatment strategies and improving patient management. Radiomics and deep learning (DL) approaches offer potential for non-invasive survival predictions. In this retrospective cohort study, we aimed to develop and evaluate a radiomics-based DL model for predicting survival outcomes in patients with BM using pre-treatment MRI and clinical data. This study analyzed 199 patients from the Pretreat-MetsToBrain-Masks dataset. Radiomic features were extracted from four MRI sequences using PyRadiomics, yielding 4872 features per patient. Dimensionality reduction was performed using Principal Component Analysis, reducing the features to 71 principal components. In addition to radiomic features, demographic and clinical features were included. The data was split into training (70%) and test (30%) sets. A DL-based survival model, DeepSurv, was trained and evaluated. Overall model performance was assessed using C-index, integrated Brier score, and integrated negative binomial log-likelihood (NBLL) scores. Time-dependent performance metrics for predicting 1-year, 3-year, and 5-year survival were also calculated, including the area under the receiver operating characteristic curve (AUROC). The DeepSurv model demonstrated robust predictive capability. The overall model performance showed a test set C-index of 0.725, an integrated Brier score of 0.212, and an integrated NBLL score of 1.112. Time-dependent AUROC metrics were 0.793 for 1-year, 0.770 for 3-year, and 0.807 for 5-year survival predictions. This study presents a radiomics-based DL approach for predicting survival outcomes in patients with BM using pre-treatment MRI and clinical data. The promising results highlight the model's potential as a non-invasive tool for personalized survival predictions. Key limitations include the relatively small, single-institution sample and class imbalance across primary tumor origins; external multi-institutional validation is needed before clinical deployment.