Background:

Sepsis worsening to septicemia after general ward admission, as opposed to direct transfer from the emergency room to the intensive care unit, may lead to a relatively poorer prognosis due to decreased early medical attention. Among sepsis cases, sepsis accompanied by pneumonia holds a significant portion, necessitating respiratory research and analysis.

Objective:

This study aims to predict sepsis within 4 hours of admission in general ward patients using machine learning-based artificial intelligence algorithms.

Method:

A retrospective cohort study was conducted on 9,824 adult patients diagnosed with pneumonia from Chest X-ray or Chest CT results out of 43,147 general ward admissions at Hospital between January 2018 and December 2022. The study compared the predicted timing of sepsis using the AI software(AITRICS-VC, SEPS) with the operational definition of Sepsis-3 guidelines and analyzed risk factors for sepsis at admission.

Result:

The AI predictive model demonstrated superior performance with an AUROC value of 0.870 and AUPRC of 0.133, surpassing NEWS (AUROC 0.697, AUPRC 0.133) and MEWS (AUROC 0.661, AUPRC 0.019). Sensitivity and specificity for AI were 76.7% and 84.1%, respectively, with an average prediction time 183 minutes earlier than the diagnosis based on operational diagnosis from Sepsis-3 guidelines.

Conclusion:

In pneumonia-associated sepsis patients, the AI showed a predictive timing that averaged 3 hours earlier than the diagnosis based on operational definitions of sepsis. It also demonstrated superior performance compared to other clinical indicators like NEWS and MEWS.