ICU mortality prediction using Cox-Bayesian models
No Thumbnail Available
At Intensive Care Units (ICUs), quick assessment of patient risk is extremely important in terms of patient survival. Predictive models can be used to assess risk of mortality of individual patients, thereby aiding intensivists in their decision making. Statistics show that the mortality risk of ICU patients is largest directly after admission, but a prolonged risk is present even after one year. Despite the availability of large amounts of data, current state of the art predictive models are static models which are limited to predicting in-hospital survival. In this research, two Bayesian models are proposed based on the Cox Proportional Hazards model for addressing 1) in-hospital mortality and 2) the chance of survival at time intervals after ICU admission. The model parameters are sampled using Hamiltonian Monte Carlo sampling with a No-U-Turn Sampler. Results show that the static model achieves an AUC of 0.763 on validation data (N = 1303). The AUC shows that the static model is not as discriminative as the state of the art static models. For the time-series model, AUCs are reported for multiple time windows. The time-series model provides insight into how risk of mortality develops over time and how prolonged risk can be modelled for individual patients. Intensivists should be aware that current state of the art models do not paint a complete picture of risk of mortality over time. Researchers aiming to improve the current state of the art models should extend their models to allow for prediction of mortality both during the ICU stay as well as for an extended period of time after discharge.
Faculteit der Sociale Wetenschappen