Improving Recurrent Neural Network Responsiveness to Acute Clinical Events

Abstract

Predictive models in acute care settings must immediately recognize precipitous changes in a patient’s status when presented with data reflecting such changes. Recurrent neural networks (RNN) have become popular for clinical decision support models but exhibit a delayed response to acute events. New information must propagate through the RNN’s cell state before the total impact is reflected in the model’s predictions. Input data perseveration is a method to train more responsive RNN-based models. Input data is replicated $k$ times during training and deployment. Each replication propagates through the cell state and output of the RNN, but only the output at the final replication is maintained and broadcast as the prediction for evaluation. De-identified Electronic Medical Records (EMR) of 12, 826 patients admitted to a tertiary care pediatric academic center between $01/2009{\textrm -}02/2019$ were analyzed. A baseline Long Short-Term Memory (LSTM) model ( $k=1$ ), four LSTMs with increasing amounts of input data perseveration ( $k=2$ to $k=5$ ), and an LSTM with an attention mechanism were trained to predict ICU-mortality. Performance of models was compared using Area Under the Receiver Operating Characteristic Curve (AUROC) after increasing periods of observation from one to 12 hours. The average variation of the change in predicted mortality immediately following defined acute events measured responsiveness. The AUROC gains due to input perseveration were larger at the earlier times of prediction ( $\leq 6$ hours), increasing at the first hour from 0.77 with no input data perseveration to 0.83 when $k=5$ . An LSTM with $k=5$ was $2-3$ times more responsive to acute events than a baseline LSTM.

Keywords

• Electronic medical records
• health and safety
• Kalman filters
• long short-term memory
• machine learning
• performance evaluation