Sustainable Development Goals

Abstract/Objectives

The development of electronic health record databases has led to the use of machine learning for medical decision-making. Identifying patients at risk of cardiac arrest or death in the intensive care unit is crucial, and multimodal machine learning has been proposed as a useful tool for performing this task. In this study, various combinations of multimodal machine learning were used including data-level fusion, feature-level fusion, and decision-level fusion to predict death and cardiac arrest. The study utilized time-independent data, time-related data, and imaging data to process the predictions. Logistic regression, random forest, support vector machine, k-nearest neighbor, long short-term memory, and convolutional neural network were used as machine learning models. Results indicate that the combination of data types for prediction is beneficial, and the sensitivity and AUROC for predicting in-hospital death and cardiac arrest increased when using multimodal machine learning. The study concluded that this approach can help clinicians better identify patients at high risk of death or cardiac arrest and reduce healthcare costs.

Results/Contributions

We propose multiple combinations of multimodal machine learning using data-level fusion, feature-level fusion, and decision-level fusion to predict in-hospital death and in-hospital cardiac arrest of patients in the intensive care unit. In our study, we use time-independent (static) data (such as age, gender, and race), time-related (dynamic) data (such as heart rate, breathing rate, and blood oxygen concentration), and imaging data (chest X-ray images). We use logistic regression (LR), random forest (RF), eXtreme Gradient Boosting tree (XgBoost), support vector machine (SVM), and K-nearest neighbor (KNN) to process static data, use long short-term memory (LSTM) to process dynamic data, and use convolutional neural network (CNN) to process chest X-ray images. In our research, we use publicly available MIMIC-IV data for model training and internal validation, and further use publicly available eICU data for external validation. According to the experimental results, the combination of different types of data for prediction is beneficial. The area under the receiver operating characteristic (AUROC) for predicting in-hospital death within one hour reached 0.94, with a sensitivity of 0.88. The AUROC for predicting in-hospital cardiac arrest within one hour reached 0.91 with a sensitivity of 0.88. Compared to results obtained using only static data, the AUROC for predicting death rate increased by 0.09, and sensitivity increased by 0.12. The AUROC for predicting cardiac arrest increased by 0.03, and sensitivity increased by 0.09. Therefore, we believe that this multimodal framework can help clinicians more accurately identify patients at high risk of death or impending cardiac arrest and reduce healthcare costs.

Keywords

Electronic health recordsmachine learningintensive care unitcardiac arrestdeathdata-level fusion feature-level fusiondecision-level fusion