Towards the prediction of mortality in Intensive Care Units patients: A Simple Correspondence Analysis approach

In the setting of the PhysioNet/CinC Challenge 2012 Event 1, a new method to predict in hospital mortality in the Intensive Care Units (ICU) is proposed. The predictor, retrieved by Simple Correspondence Analysis (SCA), is based on a combination of clinical and laboratory data with more traditional score systems such as APACHE-II and SAPS-II. Information from records out of 12000 ICU patients was equally divided in three sets: A, B and C. Up to 37 variables were recorded during the first 48h after admission to the ICU. Using Set A, SCA was applied to select the variables most related to patients mortality from their hospitalizations. The proposed predictor combines these variables using the traditional APACHE II and SAPS II scores. SCA results show that variables such as creatinine, urine output, bilirubin and mechanical ventilation support were capable to discriminate between patients who survive or do not survive their ICU stays. Using these variables, the prediction method provides a SCORE1=43.50% using set A, SCORE1=42.25% using set B and SCORE1=42.73% using set C, where SCORE1 is defined as min(sensibility, positive predictivity). These results represent an improvement of 14% in SCORE1 when compared with traditional score SAPS-I (43.50% vs. 29.60%).