고차원 생존데이터에 대한 딥러닝 분석 및 해석력에 관한 연구

Abstract

High-dimensional survival data are characterized by having a far greater number of input variables (p) than sample size (n), along with the presence of censored observations. Traditional approaches to analyz ing such data have primarily relied on penalized survival analysis based on the proportional hazards model proposed by Cox (1972). However, these penalized methods assume linear risk function of covariates, limiting their ability to capture complex nonlinear relationships or interaction patterns among input variables. To address this limitation, a variety of deep learning models based on the Cox proportional hazards framework have been developed in recent years. Among them, Kim et al. (2025) pro posed an attention-based multi-modal convolutional neural network (MCNN), which showed enhanced predictive performance for high-dimen- sional survival data. Nevertheless, the interpretability of the MCNN model has not been investigated. In this study, we explore the interpretability of the MCNN model applied to high-dimensional survival data by incorporating explainable artificial intelligence (XAI) techniques widely used in the AI field. Specifically, we utilize Shapley Additive Explanations (SHAP), Feature Ablation, and Permutation Feature Importance (PFI) to identify the main input variables that contribute to the model’s predictions. Experiments were conducted using three real-world high-dimensional survival datasets with distinct characteristics, and the predictive performance of MCNN was compared with various conventional machine learning and deep learning-based sur- vival analysis models. This study demonstrates not only the strong pre- dictive power of the MCNN model but also its interpretability, thereby highlighting the potential of explainable deep learning models for analyz- ing high-dimensional survival data.