Hemorrhage diagnosis using sequential deep learning models of waveform vital sign data

Abstract

Introduction: Hemorrhage is the most common cause of trauma deaths and the most frequent complication of major surgery. It is difficult to identify until profound blood loss has already occurred. We aim at detecting hemorrhage early and reliably using waveform vital sign data routinely collected before, during, and after surgery.

Methods: We use waveform vital sign data collected at 250 Hz during a controlled transition from a stable (non-bleeding) to a fixed bleeding state of 93 pigs. These vital signs include airway, arterial, central venous and pulmonary arterial pressures, venous oxygen saturation (SvO2), pulse oximetry pleth and ECG heartrate, continuous CO, and stroke volume variation (LiDCO). We used Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM) and dilated, causal, one-dimensional convolutional neural networks presenting time windows drawn from the raw vital signs as inputs during training.

Results: Our GRU model reaches Area Under ROC (AUC) at a clinically relevant False Positive Rate (FPR) <1% of 0.7469 and True Positive Rate of 0.5639, thus achieving an operationally relevant performance for a real-world clinical application (Table 1). However, outside of the very low FPR range (cf. ROCs in Fig. 1 and 2), our models appear inferior to a referenced Random Forest (RF) classifier. Conclusions: Our work demonstrates the applicability of deep learning models to diagnose hemorrhage based on raw, waveform vital signs. Future work will address why the RF classifier can address the greater homogeneity of subjects when they bleed compared to an apparently wide dispersion of their statuses when being stable.