Hazel Tang A science writer with data background and an interest in the current affair, culture, and arts; a no-med from an (almost) all-med family. Follow on Twitter.
Abdelaziz Farhat MD, Neel Shah MD, Jefferson Tweed MS, Ziheng Wang PhD, Jeon Lee PhD, Rafe McBeth PhD, Ravi Thiagarajan MD, & Lakshmi Raman MD
Background & Problem: Extracorporeal membrane oxygenation (ECMO) is one of the most complex therapies used in the intensive care unit, and its use continues to increase globally. Overall survival is approximately 60%, but neurologic injury is a significant morbidity. Patients who suffer an ECMO-related neurological injury have near 80% mortality. Despite plentiful data and experience, it remains a significant challenge to predict which patients are at most significant risk for neurological injury.
Solution: We developed a hybrid deep learning neural network integrating a convolutional neural network (CNN) and recurrent neural network (RNN) to analyze 38 spatial and temporal variables across 1.4 million unique datapoints.
Methods: We retrospectively collected hourly data of 174 ECMO patients with post EMCO neurological imaging. Inputs include physiological data, markers of end organ perfusion, markers of coagulation and hemolysis, acid-base homeostasis, vasoactive data, and ECMO mechanical data. We collected data for the duration of the ECMO run, as well as 24 hours prior to cannulation. We utilized a CNN and RNN in parallel. This approach was implemented using Keras with Tensorflow as backend, run over 300 learning epochs and utilizing stochastic gradient descent. Validation was by using k-fold cross validation. We further performed drop-out testing to identify which category of variables held the greatest influence on outcomes.
Results: 174 patients were analyzed over 40,000 clinical hours of ECMO. The model achieved a 75% accuracy with 70% sensitivity and 80% specificity. Positive predictive value was 78% and negative predictive value was 73%. The positive likelihood ratio with a 95% CI was 3.50 (2.32 – 5.29). The negative likelihood ratio with a 95% CI was 0.38 (0.27 – 0.51). Physiological data when withheld had the highest impact on model accuracy, while markers of coagulation and hemolysis had the lowest impact on model accuracy.
