Forecasting Hemorrhagic Shock Using Patterns of Physiologic Response to Routine Pre-operative Blood Draws

Abstract

Introduction: Irreversible hemorrhagic shock (IHS), a critical condition associated with significant blood loss and poor response to fluid resuscitation, can induce multiple organ failures and rapid death [1]. Determining the patients who are likely to develop IHS in surgeries could greatly help pre-operative assessment of patient outcomes and allocation of clinical resources.

Methods: machine learning model of IHS is developed and validated via porcine induced bleed experiment. 36 healthy sedated Yorkshire pigs first had one 20mL rapid blood draw during a stable period, and then were bled at 20mL/min to mean arterial pressure (MAP) of 30 mmHg. 10 subjects had IHS defined as MAP<20mmHg. Arterial, central venous and airway pressures collected at 250 Hz during the blood draw [Fig 1] were used to extract characteristic sequential patterns using Graphs of Temporal Constraints (GTC) methodology [2], and a decision forest (DF) model was trained on these patterns to determine subjects at high risk of impending IHS. Results: In a leave-one-subject-out cross-validation, our method confidently identifies 30% (95% CI [15.6%, 44.4%]) of the subjects who are likely to experience IHS when subject to substantial bleeding, while only giving on average 1 false alarm in 10,000 such predictions. This method outperforms logistic regression and random forest models trained on statistically featurized data [Tab 1, Fig 2].

Conclusions: Our results suggest that by leveraging sequential patterns in hemodynamic waveform data observed in pre-operative blood draws, it is possible to predict who are prone to develop IHS resulting from blood loss in the course of surgery. Future work includes validating the proposed method on data collected from human subjects, and developing a clinically useful screening tool with our investigations.