Using ECG signals for hypotensive episode prediction in trauma patients.

BACKGROUND AND OBJECTIVES

Bleeding is the leading cause of death among trauma patients both in military and civilian scenarios, and it is also the most common cause of preventable death. Identifying a casualty who suffers from an internal bleeding and may deteriorate rapidly and develop hemorrhagic shock and multiorgan failure is a profound challenge. Blood pressure and heart rate are the main vital signs used nowadays for the casualty clinical evaluation in the battlefield and in intensive care unit. However, these vitals tend to deteriorate at a relatively late stage, when the ability to prevent hazardous complications is limited. Identifying, treating, and rapidly evacuating such casualties might mitigate these complications. In this work, we try to improve a state-of-the-art method for early identification of Hypotensive Episode (HE), by adding electrocardiogram signals to several vital signs.

METHODS

In this research, we propose to extend the state-of-the-art HE early detection method, In-Window Segmentation (InWise), by adding new types of features extracted from ECG signals. The new predictive features can be extracted from ECG signals both manually and automatically by a convolutional auto-encoder. In addition to InWise, we are trying to predict HE using a Transformer model. The Transformer is using the encoder output as an embedding of the ECG signal. The proposed approach is evaluated on trauma patients data from the MIMIC III database.

RESULTS

We evaluated the InWise prediction algorithm using four different groups of features. The first feature group contains the 93 original features extracted from vital signs. The second group contains, in addition to the original features, 24 features extracted manually from ECG signal (117 features in total). The third group contains the original features and 20 ECG features extracted by the AE (113 features in total), and the last group is the union of all three previous groups containing 137 features. The results show that each model, which has used ECG data, is outperforming the original InWise model, in terms of AUC and sensitivity with p-value <0.001 (by 0.7% in AUC and up to 3.8% in sensitivity). The model which has used all three feature types (vital signs, manual ECG and AE ECG), outperforms the original model both in terms of accuracy and specificity with p-value <0.001 (by 0.3% and 0.4% respectively).

CONCLUSION

The results show an improvement in the prediction success rates as a result of using ECG-based features. The importance of ECG features was confirmed by the feature importance analysis.