Unified Deep Learning Survival Analysis for Competing Risk Modeling with Functional Covariates

From R/Medicine 2025 Full Title: Unified Deep Learning Survival Analysis for Competing Risk Modeling with Functional Covariates and Missing Data Imputation Discharging patients from the intensive care unit (ICU) marks an important moment in their recovery—it’s a transition from acute care to a lower level of dependency. However, even after leaving the ICU, many patients still face serious risks for adverse outcomes, such as ICU readmission due to complications or subsequent in-hospital death. Such events can slow recovery, increase healthcare costs, and substantially impact patients’ quality of life. Accurate prediction of these outcomes could transform the way we care for ICU survivors. With improved predictive capabilities, clinicians can intervene sooner, personalize post-ICU support, and ultimately improve recovery while reducing the chances of returning to critical care. That said, making these predictions is anything but simple. ICU data is challenging—there are multiple competing risks (such as septic shock, ICU readmission, or death during or after hospitalization), and the data is both complex and multidimensional, spanning demographic factors, clinical information, and continuous waveform covariates (e.g., blood pressure and respiratory rate). Additionally, some data may be frequently missing, and time-dependent treatments make it even more difficult. Traditional competing risk models, such as cause-specific hazard models and the Fine & Gray sub-distribution hazard models, often fall short in managing these complexities. To address these challenges, our research team has developed an innovative deep-learning model specifically designed for competing risk analysis in ICU settings. This framework integrates discrete-time competing risk analysis techniques, adaptive data-driven basis layers with micro neural network for functional variables, and advanced Imputation-Regularized Optimization (IRO) method to manage missing data effectively. Unlike traditional basis representation methods—such as those using Fourier or spline bases—which do not leverage outcome information during dimension reduction, our approach fully integrates this information. This comprehensive approach greatly enhances our ability to understand and predict complex patient outcomes based on vital signs and other critical data. We validated our model through simulations and real-world ICU data from both a large public ICU database MIMIC-4 (over 12,000 patients) and the Cleveland Clinic’s ICU records (over 25,000 patients). Across different clinical scenarios, our deep-learning framework outperformed traditional methods on most occasions, showing better predictive accuracy based on the metrics, integrated brier score and concordance index. Ultimately, this research represents an important step forward in improving ICU patient care, enabling more precise, timely, and effective clinical decisions. Resources R/Medicine: https://rconsortium.github.io/RMedici... R Consortium: https://www.r-consortium.org/