Predicting mortality of ICU patients using statistics of physiological variables and support vector machines

We began using the same variables as SAPS-1 score, adding the rest of variables one by one as recommended by physicians, to observe whether the SVM classification improves. These variables include: Age, HR, SysABP, NISysABP, Temp, RespRate, MechVent, Urine, BUN, HCT, WBC, Glucose, K, Na, HCO3, GCS, and other variables that were added for phase 1: DiasABP, NIDiasABP, Cholesterol, Creatinine, and SaO2. We found a 6.1% error in the Set-A files due to the absence of measures such as: RespRate, Temp, and age. To solve for these errors on phase 1 we chose to input values within the normal range for these physiological variables. We calculated: mean, standard deviation, and range of variation (max and min) for each one of the physiological variables. These values were placed in nodes corresponding to an index and a value of the variable, which were escalated between 0 and 1. We created a matrix where the columns corresponded to: means and standard deviations of the input variables, and rows corresponded to the individual patient´s records. We decided to use SVM. Five SVM machines were tested and scored. To conclude, we demonstrate the applicability of SVM for predicting mortality of ICU patients with a final score using set-B of 0.350352 for event 1.