A novel and reliable framework of patient deterioration prediction in Intensive Care Unit based on long short-term memory-recurrent neural network

Abstract

The clinical investigation explored that early recognition and intervention are crucial for preventing clinical deterioration in patients in Intensive Care units (ICUs). Deterioration of patients is predictable and can be preventable if early risk factors are recognized and developed in the clinical setting. Timely detection of deterioration in ICU patients may also lead to better health management. In this paper, a new model was proposed based on Long Short-Term Memory-Recurrent Neural Network (LSTM-RNN) to predict deterioration of ICU patients. An optimisation model based on a modified genetic algorithm (GA) has also been proposed in this study to optimize the observation window, prediction window, and the number of neurons in hidden layers to increase accuracy, AUROC, and minimize test loss. The experimental results demonstrate that the prediction model proposed in this study acquired a significantly better classification performance compared with many other studies that used deep learning models in their works. Our proposed model was evaluated for two tasks: mortality and sudden transfer of patients to ICU. Our results show that the proposed model could predict deterioration before one hour of onset and outperforms other models. In this study, the proposed predictive model is implemented using the state-of-the-art graphical processing unit (GPU) virtual machine provided by Google Colaboratory. Moreover, the study uses a novel time-series approach, which is minute-by-minute. This novel approach enables the proposed model to obtain highly accurate results (i.e., an AUROC of 0.933 and an accuracy of 0.921). This study utilizes the individual and combined effectiveness of different types of variables (i.e., vital signs, laboratory measurements, GCS, and demographic data). In this study, data was extracted from MIMIC-III database. The ad-hoc frameworks proposed by previous studies can be improved by the novel and reliable prediction framework proposed in this research, which will result in predictions of more accurate performance. The proposed predictive model could reduce the required observation window (i.e., a reduction of 83%) for the prediction task while improving the performance. In fact, the proposed significant small size of observation window could obtain higher results which outperformed all previous works that utilize different sizes of observation window (i.e., 48 hours and 24 hours). Moreover, this research demonstrates the ability of the proposed predictive model to achieve accurate results (>80%) on 'raw' data in an experimental work. This shows that the rule-based pre-processing of clinical features is unnecessary for deep learning predictive models.