ICU monitoring support has largely been benefited from the implementation of machine learning approaches such as deep learning algorithms, applicable in online settings and wearable technologies. Notwithstanding, deep learning algorithms require large amounts of data in the training phase, which compromise the computational cost and time. Thus, interest in time series prediction and classification using simpler machine learning models has fostered the development of alternative approaches, such as Echo State Networks (ESNs). Furthermore, already well-known algorithms like Random Forests or Support Vector Machines, are also applicable in settings where datasets are not extensive enough for a reliable deep neural network training.
In this work we propose the use of ESNs for the classification of Sepsis based on ICU monitoring records. Our approach, leveraging time dependencies, builds upon a simple ring topology that manages to substantially reduce computational costs while presenting a relevant performance tradeoff, simplifying the training procedure used in recurrent neural networks. In order to assess the performance of our approach, a 10-Fold stratified procedure has been carried out on the dataset from 2 hospitals (40336 patients), obtaining an area under the receiver operating characteristic curve (AUC) = 0.8296, accuracy (ACC) = 0.9643 and score F1 = 0.2037.
Moreover, our implementation allows the training and algorithm optimization in one of the hospitals, tested against the other. The performance obtained when training on hospital A and testing with hospital B is AUC = 0.8006, ACC = 0.9651, F1 = 0.1875. The reciprocal, training with hospital B and testing with hospital A, yields a AUC = 0.7121, ACC = 0.8921, F1 = 0.0972 performance. The results achieved, suggest that ESNs for Sepsis classification is suitable for generalization to other similarly structured ICU databases.