Predicting post-stroke pneumonia using deep neural network approaches

BACKGROUND AND PURPOSE: Pneumonia is a common complication after stroke, causing an increased length of hospital stay and death. Therefore, the timely and accurate prediction of post-stroke pneumonia would be highly valuable in clinical practice. Previous pneumonia risk score models were often built on simple statistical methods such as logistic regression. This study aims to investigate post-stroke pneumonia prediction models using more advanced machine learning algorithms, specifically deep learning approaches. METHODS: Using a hospital's electronic health record(EHR) data from 2007-2017, 13,930 eligible patients with acute ischaemic stroke (AIS) were identified to build and evaluate the models (85% of the patients were used for training, and 15% were used for testing). In total, 1012 patients (7.23%) contracted pneumonia during hospitalization. A number of machine learning methods were developed and compared to predict pneumonia in the stroke population in China. In addition to the classic methods (i.e., logistic regression (LR), support vector machines (SVMs), extreme gradient boosting (XGBoost)), methods based on multiple layer perceptron (MLP) neural networks and recurrent neural network (RNNs) (i.e., attention-augmented gated recurrent unit (GRU)) are also implemented to make use of the temporal sequence information in electronic health record (EHR) systems. Prediction models for pneumonia were built for two time windows, i.e., within 7 days and within 14 days after stroke onset. In particular, pneumonia occurring within the 7-day window is considered highly associated with stroke (stroke-associated pneumonia, SAP). MAIN FINDINGS: The attention-augmented GRU model achieved the best performance based on an area under the receiver operating characteristic curve (AUC) of 0.928 for pneumonia prediction within 7 days and an AUC of 0.905 for pneumonia prediction within 14 days. This method outperformed the other machine learning-based methods and previously published pneumonia risk score models. Considering that pneumonia prediction after stroke requires a high sensitivity to facilitate its prevention at a relatively low cost (i.e., increasing the nursing level), we also compared the prediction performance using other evaluation criteria by setting the sensitivity to 0.90. The attention-augmented GRU achieved the optimal performance, with a specificity of 0.85, a positive predictive value (PPV) of 0.32 and a negative predictive value (NPV) of 0.99 for pneumonia within 7 days and a specificity of 0.82, a PPV of 0.29 and an NPV of 0.99 for pneumonia within 14 days. CONCLUSIONS: The deep learning-based predictive model is feasible for stroke patient management and achieves the optimal performance compared to many classic machine learning methods.