Abstract

In this paper, the possibility of an early prediction for sepsis using deep learning was investigated. The sepsis is a disease that infects microorganisms, resulting in systemic inflammatory reactions such as fever, rapid pulse rate, respiratory increase rate, increase or decrease of white blood cell count. Sepsis also currently incur the highest medical costs of all diseases, affecting about 30 million people a year worldwide. Sepsis, an infectious disease, is essential for early detection because it can lower mortality rates with treatment during the initial three hours of infection and faster antibiotic administration. However, it takes a lot of tests to detect sepsis. White Blood Cells (WBCs) are particularly important in diagnosing sepsis but require blood tests. This acts as an obstacle to early detection by having additional medical expenses and time spent on the examination. Thus, this paper studied a deep learning model that can initially predict sepsis by calibrating the white blood cell count values acting as an important factor in sepsis detection. The data set in this paper utilizes PhysioNet's 'Early Prediction of Sepsis from Clinical Data the PhysioNet Computing in Cardiology Challenge 2019. Data are inpatient data from the intensive care unit released by two hospitals, including biometric signals (1 to 8), body component test results (9 to 34), other information (35 to 40) and annotations (41). This study used nine of the above 41 items and annotations, especially WBC is an important factor in diagnosing sepsis. However, because WBC require blood tests, they are harder to measure than biometric signals. This omitted many values (91 percent). Thus, this paper studied algorithms to calibrate the WBC and, furthermore, predict sepsis early, using the GAN (Generative Adversarial Network) technique. In this paper, the indicators of objectification focus on the ROC (Receiver Operating Characteristics) Curve and AUC (Area Under the Curve). The study for sepsis prediction showed excellent performance by the LSTM (Long Short-Term Memory) method, which is advantageous for sequential pattern learning, because it predicts the patient's condition by time. The LSTM-style study was constructed using approximately 6,000 data and showed a performance of 0.929 AUC. Traditional studies have also used WBC as a factor to predict sepsis, but many of the missing data have not been reflected in the study. However, in this study, missing data values are generated and corrected through a deep learning model, which can be a strength for data-based deep learning algorithms. In this paper, mixed data and pre-calibration data were divided into separate cases and tested. Mixed data is composed of calibrated data and original data mixed at half-and-half ratio, and the pre-calibration data is original data. The test results showed that the two cases were 0.705 (original) and 0.98 (mixed) AUC, respectively. Thus, this paper has improved the performance of sepsis predictive model by calibrating the WBC among blood test factor. The future works will verify this model in a real hospital and first it will complement the WBC calibration model according to the patient's Vital Signal. Since then, cross-verification is also planned with data from various agencies.