(1046) Enhancing Sepsis Alerts With Machine Learning: Typing Profiles for ICU Admissions and Mortality

Introduction: Sepsis poses significant challenges in healthcare due to its complex nature and high mortality rates. Accurate early prediction of ICU admissions and mortality is crucial for effective intervention. This study employs machine learning techniques to type diverse patient profiles, including those without sepsis, to improve the predictive accuracy of sepsis alert systems.

Methods: A retrospective cohort study was conducted using data from 3862 sepsis patients. The dataset included key clinical variables such as age, NEWS score, SEP-2 parameters, and suspected infection status. The primary outcomes were ICU admission within 48 hours or death. Data preprocessing ensured consistency, and K-modes clustering identified three distinct phenotypes. Chi-square tests assessed the independence of clinical variables across clusters, and logistic regression models evaluated the predictive power of clusters for ICU admission or death. AI was utilized as a creative supplement in authoring this abstract.

Results: Three clusters were identified among sepsis patients. Cluster 1 had younger patients with elevated INR and low infection suspicion. Cluster 2 had higher infection suspicion, younger patients, and elevated platelet counts. Cluster 3 had older patients with elevated lactate and higher infection suspicion. Chi-square tests showed significant differences across clusters for age (Chi2 = 1252.98, p < 0.0001), NEWS score (Chi2 = 152.30, p < 0.0001), SEP-2 parameters (Chi2 = 2304.57, p < 0.0001), suspected infection (Chi2 = 1853.30, p < 0.0001), and ICU/death outcome (Chi2 = 48.16, p < 0.0001). Logistic regression indicated significant predictive power for ICU admission or death with odds ratios of 0.616 for Cluster 1 (p < 0.0001), 1.461 for Cluster 2 (p < 0.0001), and 1.268 for Cluster 3 (p = 0.005). Cluster 1 had the lowest ICU admission risk (25.55%), Cluster 2 the highest (36.43%), followed by Cluster 3 (34.82%).

Conclusions: This study showcases the potential of machine learning techniques in typing diverse risk profiles among sepsis and non-sepsis patients. The insights gained can guide personalized treatment strategies and improve the performance of sepsis alert systems in predicting ICU admissions and hospital mortality, ultimately enhancing patient care and outcomes.