Dynamic Re-admission Prediction of Heart Patients Using Scalable Models

Abstract

Heart Failure is one of the most common reasons for hospitalization of people aged over 65, causing hospitals' economic burden. A large reason for the high costs of these patients is both the high readmission costs and rate. Predicting the readmission risk at discharge helps the hospital keep patients who are not ready yet to be discharged. Traditional readmission prediction methods do not use patients' full admission window to predict readmission because of model or data restrictions. These restrictions remove many useful datapoints of patients to predict their readmission risk. Using the full admission window does significantly increase the data, hence the need for a scalable model.

This thesis aims to adapt the LSTM model to adjust to new high dimensional data efficiently and proposes extending the LSTM with a Convolutional Neural Network, extracting prominent features, and reducing the data dimensions low dimensional sequences. The method ensures the LSTM can scale efficiently with new data and increase its predictive performance.

The research approach is based on the CRISP-DM method. The literature review creates an overview of current readmission prediction methods, resulting in an optimal model for this thesis to improve. The features and data selection are made based on literature and the medical supervisor's input, after which exploratory data analysis is performed to exclude irrelevant data records. The CNN-LSTM model is created with the help of the technical supervisors. Finally, the CNN-LSTM model is evaluated using 5-fold cross-validation measuring the average AUC, training time and number of epochs. The model is compared to baseline models and the standalone parts of the CNN-LSTM regarding their predictive performance and scalability.

The temporal models utilizing the full admission window show the best predictive performance compared to the baseline models, showing the power of including the temporal dimension in predicting readmission risk. The proposed CNN-LSTM model shows to be the most scalable temporal model with slightly lower predictive power than the standalone LSTM option, which is the least scalable model option.