Infection-Aware Risk ML Model for Early Prognosis and Treatment Contextualization in Sepsis

Sunday, March 22, 2026

2:30 PM - 3:30 PM Central Time

Location: Connections Central - RST 14

LS

Longxiang Su, n/a (he/him/his)

Peking Union Medical College Hospital, Beijing, CHN

Disclosure(s): No relevant financial relationship(s) to disclose.
XM

Xudong Ma, NA

MD
National Health Commission of the People's Republic of China

Disclosure information not submitted.
SG

Sifa Gao, NA

MD
National Health Commission of the People's Republic of China

Disclosure information not submitted.
JS

Jialu Sun, NA

MD
National Health Commission of the People's Republic of China

Disclosure information not submitted.
YW

Yuna Wei, NA

MM
Peking Union Medical College Hospital

Disclosure information not submitted.
YW

Yao Wang, NA

MM
Peking Union Medical College Hospital

Disclosure information not submitted.
WP

Wei Pan, NA

MM
Peking Union Medical College Hospital

Disclosure information not submitted.
YW

Ye Wang, NA

MPH
Peking Union Medical College and Chinese Academy of Medical Sciences

Disclosure information not submitted.
MS

Maximilian Schaefer, NA

MD
Beth Israel Deaconess Medical Center

Disclosure information not submitted.
DT

Daniel Talmor, N/A

Prof
Beth Israel Deaconess Medical Center

Disclosure information not submitted.
TS

Ting Shu, NA

MD
National Health Commission

Disclosure information not submitted.
KW

Kai Wang, NA

MD
National Health Commission

Disclosure information not submitted.
DL

Dawei Liu, NA

MD
Peking Union Medical College Hospital

Disclosure information not submitted.
BD

BIn Du, NA

MD
Peking Union Medical College Hospital

Disclosure information not submitted.
XZ

Xiang Zhou, NA

MD
Peking Union Medical College Hospital

Disclosure information not submitted.

Introduction: To develop and validate an infection-aware Sepsis Risk Stratification (SRS) system for predicting in-hospital mortality and guiding early clinical decision-making.

Methods: This was a multi-cohort prognostic study using a retrospective nationwide Chinese database (n = 116,852) for model development and internal validation. External validation was performed using the MIMIC-IV (n = 18,407), eICU Collaborative Research Database (n = 9,816), and a prospectively collected cohort. A machine learning–based model incorporating physiologic and infection-specific variables, including infection foci, was developed. The primary outcome was in-hospital mortality. Model performance was assessed by area under the receiver operating characteristic curve (AUROC) and calibration. Secondary outcomes included 28-day mortality, ICU length of stay, ventilator days, and hospitalization costs.

Results: The study analyzed 116852, 18,407, and 9,816 sepsis patients in the internal, MIMIC-IV, eICU-CRD, and the prospective cohorts, respectively, with in-hospital mortality rates of 36.1%, 18.2%, 17.1%, and 27.6%. The risk model achieved AUROCs of 0.80 (95% CI: 0.79–0.80; internal), 0.68 (95% CI: 0.67–0.69; MIMIC-IV), 0.73 (95% CI: 0.71–0.74; eICU-CRD), and 0.75 (95% CI: 0.69–0.81; prospective cohort). External validation showed that the risk model significantly outperformed the SOFA (internal validation: 0.80 vs. 0.67, p < 0.001; MIMIC-IV: 0.68 vs. 0.67, p=0.01; eICU-CRD 0.73 vs. 0.70, p< 0.001, prospective cohort: 0.75 vs. 0.57, p=0.012) and APACHE II (internal validation: 0.80 vs. 0.66, p< 0.001; MIMIC-IV: 0.68 vs. 0.63, p=0.002, eICU-CRD: 0.73 vs. 0.70, p< 0.001). The risk scores were categorized into four sepsis risk levels with corresponding probability: low, medium, high, and very high. These risk categories predicted in-hospital mortality: low (6.4%), medium (21.7%), high (46.3%), and very high (69.5%) in the internal validation. Similar trends were observed for other metrics, such as 28-day mortality, length of ICU stay, ventilator day, and total cost.

Conclusions: The SRS system offers a rapid, interpretable, and infection-integrated risk stratification tool, which improves upon conventional scores by incorporating infection context and supports individualized treatment strategies.