Sepsis disease prediction by deep learning

Sepsis is a disease that can lead to death of the patient if not diagnosed early. There are previous studies for the early diagnosis of the disease. This study is a first in Turkey in terms of diagnosis of Sepsis disease using deep learning method. It was tried to predict whether they had Sepsis by using the data of the first 12 hours of the patients who came to the Sadi Konuk Training and Research Hospital Intensive Care Unit with the suspicion of Sepsis. Sepsis was diagnosed in 216 patients out of 640 patients between the ages of 18-60, while 424 patients were healthy. In this study, a new hybrid metaheuristic algorithm is proposed in the model used for the prediction of Sepsis disease with deep neural networks. The gradient-based algorithms used in the optimization of the parameters of deep neural networks did not show proper results. The reason for this is that it is stuck at the local minimum value in the studies in the literature. Gradient-based algorithms have failed to reach the global minimum. In recent years, meta-heuristic algorithms, population-based solutions, swarm intelligence and evolutionary approach have shown good results. The proposed algorithm aims to reach the global minimum of particles in particle swarm optimization (PSO) by using the mental search operator of the Human mental search algorithm (HMS) with the local search strategy. In this study, a new hybrid algorithm named HMS-PSO is presented, taking into account the advantage of human mental search (HMS) and particle swarm optimization (PSO) algorithms. The proposed algorithm tries to balance the exploration and exploitation process, which leads to escaping the local minimum and provides a chance to obtain the global minimum. Then, to measure the performance of HMS-PSO, it was compared with HMS and PSO by running on benchmark functions. It has been shown that the proposed algorithm is more reliable, more durable and more flexible than other algorithms. HMS-PSO is integrated with deep neural networks (HMS-PSO-DNN). A sampled Sepsis dataset was used to select the appropriate neural network architecture to be used in the prediction of sepsis disease and to determine the hyperparameters of the neural network. To measure the predictive performance of the proposed architecture, real datasets, diabetes, heart failure, cryotherapy and immunotherapy medical datasets were used. The results demonstrated that the proposed integration model outperforms other existing gradient-based and aforementioned meta-heuristic algorithms in terms of accuracy, robustness, and performance. After the reliability of the integrated algorithm was proven, Sepsis disease was tried to be predicted using the Sepsis data set consisting of 640 patients between the ages of 18-60. Providing a balanced dataset simplifies the training of the model as it helps prevent the model from learning a single class in the Sepsis dataset. In our study, the data set for the solution of this learning problem was balanced. The performances of the proposed algorithm and other algorithms for balanced data set estimation in 2000 iterations and 30 independent run times are compared and the results are presented. In addition, in order to compare the performance of the proposed algorithm with the studies in the literature, the first 6-hour Sepsis data set was estimated. The results obtained gave better results in terms of accuracy than the studies in the literature.