In patients undergoing orthotopic liver transplantation (OLT), the incidence of acute kidney injury (AKI) is up to 60%. Prediction of the individual patient's risk of AKI during OLT might allow to adapt therapeutic interventions aimed at preventing postoperative AKI.

Deep learning is a type of machine learning that uses a multi-layered neural network with hidden layers to increase model capacity and accuracy. A recurrent neural network (RNN) is a deep learning technique in which the output of components of the neural network are fed back into the network. Long short term memory (LSTM) is a RNN unit that allows for “memory” of earlier inputs when training model parameters and has been used on medical time series data in other settings. The aim of this project was to evaluate the use of RNNs to accurately predict AKI after OLT.

We included 515 patients undergoing OLT at UCSF between June 2012 and December 2015 in this study. For each transplant, 89 variables, including perioperative flowsheet data such as continuous vital signs, ventilator settings, anesthetic concentrations, and all available laboratory values, were used. The data were normalized to reduce the range of magnitudes of inputs. AKI was defined as a serum creatinine increase of 50% within the first seven days after surgery. The patient input matrices and corresponding output values were separated into training, validation, and test sets. The training set was trained on the deep learning model using the python package Keras with TensorFlow as the backend package that was composed of: 64 LSTM's with a dropout of 0.7, a dense (fully connected) layer of 100 units, a dropout of 0.6, a dense layer of 20 units, a dropout of 0.4, and a dense layer of 2 with an activation layer of Softmax to turn the output into a binary representation. The loss function used was categorical cross entropy and the optimizer used was “ADAM” with an initial learning rate of 1E-7.

We achieved an accuracy of greater than 77% on both the validation and test sets with the above model.

A RNN can be used on perioperative data to make accurate predictions about postoperative outcomes. Such predictions could be updated in real-time during the intraoperative period to guide and adapt patient management. A better understanding of the individual patient's risk for AKI can also influence postoperative management to potentially lead to improved post-operative outcomes.

CITATION INFORMATION: Bishara A., Kothari R., Bokoch M., Niemann C., Adelmann D. Predicting Acute Kidney Injury after Orthotopic Liver Transplantation Using a Recurrent Neural Network Am J Transplant. 2017;17 (suppl 3).

« Back to 2018 American Transplant Congress