”Artificial Neural Network Modelling of Intensive Care Mortality"
In order to study and evaluate the care provided at the Intensive Care Unit (ICU), an accurate model to asses severity of illness and predict patient mortality based on admission data is required. Since 1983, the Simplified Acute Physiology Score (SAPS) based on logistic regression has become one of the international standards for this purpose and is used in Sweden for patients of age ≥ 16 years. While being a simple, interpretative model, logistic regression results in a decision boundary in the form of a hyperplane, which puts a limit to the classification abilities of something as complex as vital status. The Artificial Neural Network (ANN) is a popular machine learning model inspired by the neurons and synapses of the human brain, which result in non-linear decision boundaries for classification tasks. In this work, several variants of ANN models are assessed when classifying 30-day mortality using various optimization and regularization techniques. The optimal ANN was trained using batch normalization, dropout, and autoencoder imputation of missing values. The resulting area under the receiver operating curve (AUC) is 0.889 (95%CI: 0.888 − 0.890) on a test set of 20,000 patients, compared to 0.852 of the latest version of SAPS, SAPS3. Even when removing 79% of the lowest ranked predictors (30/38), SAPS3 is still outperformed. The Brier Score (BS), measuring calibration error, is improved from 0.114 to 0.099, whereas no large errors were obtained for high risk patients, in contrast to SAPS3. Major improvements are obtained especially for patients suffering from complex conditions such as cancer, cardiovascular diseases, and unconsciousness. The results pose interesting questions regarding further applications within the ICU beyond modelling of mortality, and the potential of the ANN model to be used as an international standard of risk adjustment.