Early Sepsis Prediction Using Ensemble Learning With Deep Features and Artificial Features Extracted From Clinical...

Sepsis is caused by infection and subsequent overreaction of immune system and will severely threaten human life. The early prediction is important for the treatment of sepsis. This report aims to develop an early prediction method for sepsis 6 hours ahead on the basis of clinical electronic health records.

Challenge data are released by PhysioNet/Computing in Cardiology Challenge 2019 and obtained from ICU patients in three separate hospital systems. Part of the data from two datasets, including 40,336 subjects, are publicly available, and the remaining are used as hidden test set. A normalized utility score defined by the organizing committee is used for model performance evaluation.

The supervised machine learning is applied to tackle this challenge. Specifically, we establish the prediction model under the framework of ensemble learning by integrating the artificial features based on clinical prior knowledge of sepsis with deep features automatically extracted by long short-term memory neural network.

Forty clinical variables, including eight vital signs, 26 laboratory values, and six demographics, were measured and recorded once an hour for each individual, and the binary label (0 or 1) was simultaneously provided for each item.

The proposed model was evaluated by 30-fold cross-validation. The sensitivity, specificity, and normalized utility score were 0.641 ± 0.022, 0.844 ± 0.007, and 0.401 ± 0.019 on publicly available datasets, respectively. The final normalized utility score our team (UCAS_DataMiner) has obtained was 0.313 on full hidden test set (0.406, 0.373, and -0.215 on test set A, B, and C, respectively).

We realized a 6-hour ahead early-onset prediction of sepsis on the basis of clinical electronic health record by ensemble learning. The results indicated the proposed model functioned well in the early prediction of sepsis. In particular, ensemble learning had a significant (p < 0.01) improvement than any single model in performance.