A comparative study of survival prediction and feature selection/extraction methods for quantitative radiomics risk modelling

This is the authors' abstract. We don't add key points for this paper.

Radiomics extracts quantitative imaging features from medical images to develop risk prediction models that support clinical decision-making. Owing to the high dimensionality and potential instability of radiomic features, accurate survival prediction depends critically on the selection of appropriate feature engineering and survival modelling methods. However, comprehensive benchmark studies comparing these methods remain limited, particularly for recently developed deep learning models. Here, we systematically benchmarked 16 survival prediction methods and 12 feature selection/extraction methods using three groups of publicly available radiomics datasets from non-small cell lung cancer and ovarian cancer cohorts, comprising 1160 patients and representing different levels of feature stability. Model performance was evaluated using repeated random training/testing splits and independent external validation with the C-index, integrated area under the curve (IAUC), and integrated Brier score (IBS). Ridge-Cox, glmboost, CoxBoost, and sslasso were consistently ranked among the top-performing methods according to the C-index and IAUC, whereas deep learning models, including CoxTime, DeepHit, and PCHazard, performed better when evaluated by IBS, highlighting that different evaluation metrics capture complementary aspects of survival prediction performance. Among feature engineering methods, supervised principal component analysis achieved the best overall performance, particularly for datasets with unstable radiomic features, whereas the Cox proportional hazards p -value-based method may be preferred when discrimination performance, as measured by the C-index or IAUC, is the primary objective. Furthermore, we investigated factors affecting the performance of deep learning models and assessed the influence of the proportional hazards and linear assumptions on survival prediction. These benchmark results provide practical guidance for selecting feature engineering and survival modelling strategies according to the characteristics of radiomics datasets, thereby facilitating the development of more robust radiomics-based survival prediction models.