Prediction of ICU In-Hospital Mortality Using Artificial Neural Networks

This work aims to predict in-hospital mortality in the open-source Physionet ICU database from features extracted from the time series of physiological variables using neural network models and other machine learning techniques. We developed an effective and efficient greedy algorithm for feature selection, reducing the number of potential features from 205 to a best subset of only 47. The average of five trials of 10-fold cross validation shows an accuracy of (86.23±0.14)%, a sensitivity of (50.29±0.22)%, a specificity of (92.01 ± 0.21)%, a positive prediction value of (50.29±0.50)%, a negative prediction value of (92.01±0.00)%, and a Lemeshow score of 119.55±9.87. By calibrating the predicted mortality probability using an optimization approach, we can improve the Lemeshow score to 27.51±4.38. The developed model has the potential for application in ICU machines to improve the quality of care and to evaluate the effect of treatment or drugs.