Background: The success of sepsis treatment depends on early predictionsrecognition. Many sepsis prediction programs are being developed, but no completely reliable means exist.

Methods: Clinical data were retrospectively collected from the electronic medical records (EMR), and utilized data from 2010 to 2019 as development data, while those from 2020 to 2021 were validation data. The collected EMR consists of eight vital signs, 13 laboratory data and three demographic information (gender, age, and oxygen delivery types). We developed deep-learning-based sepsis and septic shock early prediction system (DeepSEPS) and compared our system with other traditional early warning score systems, such as NEWS, SOFA, and qSOFA.

Results: DeepSEPS showed fair prediction power with AUROC, 0.7888, and 0.8494 for sepsis and septic shock. Compared to NEWS, SOFA, and qSOFA, DeepSEPS score system significantly outperforms all other conventional score systems. Furthermore, at onset time prediction, DeepSEPS shows the highest AUROC which is 0.9346. For DeepSEPS, the model considered the Lactate features the most in the septic shock prediction. In addition, we also discover that the heart rate and Albumin features were important for sepsis and septic shock predictions.

Conclusion: DeepSEPS reasonably predicted the onset of sepsis and septic shock early. This novel sepsis and septic shock prediction system exhibited better performance when compared with other conventional score systems of sepsis.