Computerized prediction of intensive care unit discharge after cardiac surgery: development and validation of a Gaussian processes model

BACKGROUND: The intensive care unit (ICU) length of stay (LOS) of patients undergoing cardiac surgery may vary considerably, and is often difficult to predict within the first hours after admission. The early clinical evolution of a cardiac surgery patient might be predictive for his LOS. The purpose of the present study was to develop a predictive model for ICU discharge after non-emergency cardiac surgery, by analyzing the first 4 hours of data in the computerized medical record of these patients with Gaussian processes (GP), a machine learning technique. METHODS: Non-interventional study. Predictive modeling, separate development (n = 461) and validation (n = 499) cohort. GP models were developed to predict the probability of ICU discharge the day after surgery (classification task), and to predict the day of ICU discharge as a discrete variable (regression task). GP predictions were compared with predictions by EuroSCORE, nurses and physicians. The classification task was evaluated using aROC for discrimination, and Brier Score, Brier Score Scaled, and Hosmer-Lemeshow test for calibration. The regression task was evaluated by comparing median actual and predicted discharge, loss penalty function (LPF) ((actual-predicted)/actual) and calculating root mean squared relative errors (RMSRE). RESULTS: Median (P25-P75) ICU length of stay was 3 (2-5) days. For classification, the GP model showed an aROC of 0.758 which was significantly higher than the predictions by nurses, but not better than EuroSCORE and physicians. The GP had the best calibration, with a Brier Score of 0.179 and Hosmer-Lemeshow p-value of 0.382. For regression, GP had the highest proportion of patients with a correctly predicted day of discharge (40%), which was significantly better than the EuroSCORE (p < 0.001) and nurses (p = 0.044) but equivalent to physicians. GP had the lowest RMSRE (0.408) of all predictive models. CONCLUSIONS: A GP model that uses PDMS data of the first 4 hours after admission in the ICU of scheduled adult cardiac surgery patients was able to predict discharge from the ICU as a classification as well as a regression task. The GP model demonstrated a significantly better discriminative power than the EuroSCORE and the ICU nurses, and at least as good as predictions done by ICU physicians. The GP model was the only well calibrated model.