Patient deterioration predictive model using long short-term memory recurrent neural network with genetic algorithm optimization

Abstract

The clinical investigation found that early recognition and intervention are crucial for preventing clinical deterioration in patients in Intensive Care units (ICUs) as well as in general wards. Deterioration of patients is predictable and can be avoided if early risk factors are recognized and developed in the clinical setting. Existing patient deterioration prediction methods generally have some disadvantages such as limited to specific patient groups or diseases that lead to lack of generalization, low prediction performance, and less optimized model parameter setting. This thesis proposes a patient deterioration predictive model based on Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) with Genetic Algorithm (GA) optimization. The LSTM-RNN predictive model able to accept multiple input and data types in both static and dynamic parameters to predict patient deterioration, in terms of mortality and sudden transfer of patients from general wards to ICU with good accuracy. Another main strength of this predictive model is the input dataset is based on minuteby- minute time-series data obtained from open-source MIMIC-III research database for both model training and testing, hence also contribute to good prediction performance. To identify the baseline reference model with optimal performance, the setting of LSTM-RNN predictive model is explored using heuristically approach in terms of number of hidden layers, number of neurons in the first hidden layer, number of epochs, feature selection approach, as well as the impact of data cleaning in data pre-processing. On the other hand, the GA acts as an optimization model to further enhance the prediction performance of the baseline reference LSTM-RNN predictive model by exploration and identification of the optimum parameter settings, which include observation window size, prediction window size, and number of neurons in the first hidden layer. In this study, the proposed predictive model is benchmarked with other related work in terms of various prediction model, data sequence type, patient’s age involved, number and types of features, dataset splitting ratios, prediction and observation window size and data source. For standard benchmarking result comparison, the selected performance metrics includes accuracy, area under receiver operating curve (AUROC), and test loss. The benchmarking results show that the proposed model outperforms other related models in general as it is capable to predict patient deterioration up to six hours before the onset with minimum prediction accuracy above 0.80 as recommended in the clinical setting. In specific, the best optimum LSTM-RNN predictive model after GA optimization able to achieve AUROC of 0.933, prediction accuracy of 0.921, test loss of 0.435, longer prediction window of 4.77 hours while reducing the observation window from 24 hours to 9.6 hours (60%) at the same time. The proposed patient deterioration prediction model based on LSTM-RNN, and GA will be very useful to clinical team as they have more sufficient time window to take prompt medical action before the onset of deterioration. As a result, this will help to reduce the mortality rate of patients or sudden transfer of patients from general wards to ICU.