Machine Learning Model for the Prediction of Hemorrhage in Intensive Care Units

Abstract

Objectives: Early hemorrhage detection in intensive care units (ICUs) enables timely intervention and reduces the risk of irreversible outcomes. In this study, we aimed to develop a machine learning model to predict hemorrhage by learning the patterns of continuously changing, real-world clinical data.

Methods: We used the Medical Information Mart for Intensive Care databases (MIMIC-III and MIMIC-IV). A recurrent neural network was used to predict severe hemorrhage in the ICU. We developed three machine learning models with an increasing number of input features and levels of complexity: model 1 (11 features), model 2 (18 features), and model 3 (27 features). MIMIC-III was used for model training, and MIMIC-IV was split for internal validation. Using the model with the highest performance, external verification was performed using data from a subgroup extracted from the eICU Collaborative Research Database.

Results: We included 5,670 ICU admissions, with 3,150 in the training set and 2,520 in the internal test set. A positive correlation was found between model complexity and performance. As a measure of performance, three models developed with an increasing number of features showed area under the receiver operating characteristic (AUROC) curve values of 0.61-0.94 according to the range of input data. In the subgroup extracted from the eICU database for external validation, an AUROC value of 0.74 was observed.

Conclusions: Machine learning models that rely on real clinical data can be used to predict patients at high risk of bleeding in the ICU.

Keywords: Blood Transfusion; Hemorrhage; Intensive Care Units; Monitoring; Physiological; Prognosis.

Figure 1

Flowchart of the patient selection process. A detailed flow chart of the patient selection process by dataset. We selected 5,670 intensive care admissions including hemorrhage cases (n = 1,134) and hemorrhage controls (n = 4,536). MIMIC: Medical Information Mart for Intensive Care, ICU: intensive care unit, PRBC: packed red blood cells.

Figure 2

Overall architecture of data preprocessing. For patients with hemorrhage, prediction results were obtained 3 hours prior to the point of onset during the period of ICU stays, and patient data for the previous 12 hours were used as input. Control patients’ input data were extracted at random times during ICU stays. All input data were preprocessed through missing-data imputation and a standardization process. ICU: intensive care unit.

Figure 3

Architectural overview of the hemorrhage prediction model. Dynamic features are extracted as time series, whereas static features are replicated over time. These values are integrated as a matrix of all features and labels for each patient. At each time step, the model receives current slice data as input, and features are captured in a truly sequential structure. GRU: gated recurrent unit.

Figure 4

AUROC and AUPRC curves in the MIMIC-IV test and validation sets. (A) ROC curves for the different models depending on the number of input variables. (B) Precision-recall curves for the different models depending on the number of input variables. Model 3, which achieved the highest performance, was evaluated with an external dataset. (C) ROC curves for the eICU validation set. (D) Precision-recall curves for the eICU validation set. AUROC: area under the receiver operating characteristic curve, AUPRC: area under the precision-recall curve, MIMIC: Medical Information Mart for Intensive Care, ROC: receiver operating characteristic.