Mortality prediction on unsupervised and semi-supervised clusters of medical intensive care unit patients based on MIMIC-II database

Abstract

IntroductionThe study aimed to propose a framework for identifying patient clusters in medical intensive care units (MICUs) based on the Medical Information Mart for Intensive Care II (MIMIC-II) database. The suggested framework makes use of the survival outcomes and physiological information available in the dataset and is hence called a semi-supervised approach. Five neural networks were trained on the clusters identified using the proposed approach to determine whether the proposed framework could improve the predictive accuracy of the deep learning models. MethodsThis study utilized data from the MIMIC-II database, which is a publicly available database that contains information on patients admitted to intensive care units. The clusters underlying the MICU patient population were identified using unsupervised and semi-supervised K-means clustering. Mortality in the resulting clusters was predicted using five deep learning-based survival models and the performance of these models was compared using two metrics. ResultsThree clusters (cluster 1, n = 1304; cluster 2, n = 474; cluster 3, n = 1079) were identified using unsupervised K-means, and another three clusters (cluster 1, n = 479; cluster 2, n = 1492; cluster 3, n = 886) were identified using semi-supervised K-means clustering. Experimental results demonstrate that, in general, the performance of deep learning models was better on semi-supervised clusters obtained by combining Cox proportional hazards (Cox-PH) model-based feature selection and K-means compared to unsupervised clusters. ConclusionsIn the present study, it was observed that deep learning-based survival models tend to perform better on clusters that are identified in a semi-supervised fashion. This approach helps to extract more meaningful patterns and associations between different clinical features and patient outcomes.