Discrete-time survival analysis in the critically ill: a deep learning approach using heterogeneous data.

Area of Science:

• Critical Care Medicine
• Artificial Intelligence in Healthcare
• Biomedical Informatics

Background:

• Predicting patient survival in intensive care units (ICUs) is crucial but challenging due to complex, heterogeneous data.
• Existing models often struggle to integrate diverse data types like free text, medical history, and high-frequency physiological data.
• The potential of deep learning for automated data integration and survival prediction in ICUs remains largely unexplored.

Purpose of the Study:

• To develop and evaluate a deep learning model for discrete-time survival prediction in individual ICU patients.
• To assess the feasibility of using automated data integration with minimal pre-processing of mixed data domains.
• To compare the performance of deep learning models against traditional survival models.

Main Methods:

• A deep learning model was trained on electronic patient record data from 37,355 admissions across ten ICUs (2011-2018).
• Electronic health record data were mapped to an embedded representation, inspired by natural language processing techniques.
• A recurrent neural network with a multi-label output layer was used for survival prediction, evaluated with time-dependent concordance index and SHAP for interpretability.

Main Results:

• Deep learning models significantly outperformed traditional Cox proportional-hazard models across various time points (0-72 hours).
• Concordance indices for deep learning models ranged from 0.69-0.73, demonstrating superior predictive accuracy.
• The SHAP methodology successfully quantified and visualized the key drivers of survival predictions, highlighting model interpretability.

Conclusions:

• Deep learning models integrating entity embeddings and survival modeling offer a feasible approach for individualized survival estimation in data-rich ICU settings.
• The interpretable nature of these models allows for a better understanding of factors influencing patient survival.
• This approach advances the prediction of patient outcomes in critical care by leveraging complex electronic health record data.