Abstract
Current sepsis guidelines advocate for rapid antibiotic administration, yet the subjective nature of "possible" versus "probable" sepsis often leads to over-treatment. This multicenter study utilized two deep learning models, COMPOSER and Shock-Net, to objectively stratify over 34,000 emergency department patients into risk groups. Our findings suggest that while high-risk patients benefit significantly from antibiotics within one hour, low-risk patients (those unlikely to develop shock with only possible sepsis) show similar mortality rates regardless of whether antibiotics are administered within one or three hours. These results indicate that AI-driven stratification could allow for more judicious antibiotic use without compromising patient safety.
Introduction
Sepsis remains a primary driver of hospital mortality and healthcare expenditure. The Surviving Sepsis Campaign (SSC) guidelines recommend aggressive antibiotic windows (1–3 hours) based on clinical suspicion of sepsis and shock. However, clinical assessment of sepsis probability is notoriously subjective, often resulting in "defensive medicine" where broad-spectrum antibiotics are administered prematurely to low-risk patients. This practice contributes to the global crisis of antibiotic resistance and increases the risk of adverse drug events. This study investigates whether objective, AI-derived risk stratification can identify patient cohorts where a more deliberate diagnostic approach is safe.
Methods and Study Design
The researchers employed a dual-model deep learning approach to replace subjective clinical labels with objective data:
- COMPOSER: A multidimensional projection model that calculates a sepsis probability score (0–11). Scores $\ge 6$ indicate "probable sepsis," while scores $< 6$ indicate "possible sepsis."
- Shock-Net: A transformer-based neural network designed to predict the risk of a patient progressing to septic shock within a six-hour window.
By integrating these models, patients were categorized into four risk groups analogous to SSC tiers. The primary outcome measured was short-term mortality (death or transition to hospice) relative to the timing of the first antibiotic dose from triage.
Results
The study analyzed 29,765 patients at the development site (UCSD Health) and 4,322 patients at the validation site (UCI Health).
Risk Stratification and Mortality
The AI models successfully identified a clear mortality gradient across the four groups. At the development site, Group 1 (Shock likely/Sepsis probable) exhibited a 23.2% mortality rate, whereas Group 4 (Shock unlikely/Sepsis possible) had a mortality rate of only 1.9%. Shock-Net demonstrated exceptional predictive accuracy for septic shock, with an AUC of 0.959 at the development site and 0.937 at the validation site.
Impact of Antibiotic Timing
The association between antibiotic timing and mortality varied significantly by risk group:
- High-Risk (Group 1): Rapid intervention was critical. Mortality was 20.4% for antibiotics given within 1 hour, rising to 24.7% for those given between 1–3 hours (Adjusted Odds Ratio: 1.29).
- Low-Risk (Group 4): This group comprised approximately 40% of the total population. Despite a low confirmed sepsis rate (13–15%), the majority received antibiotics. Crucially, mortality remained stable regardless of timing (1.0% for <1 hour vs. 2.4% for 1–3 hours), with no statistically significant difference after adjusting for baseline characteristics.
Discussion
This study demonstrates that deep learning can provide an objective "yardstick" for sepsis risk, potentially refining the "one-size-fits-all" approach to antibiotic timing. For the substantial population of low-risk patients, clinicians may have the "luxury of time" to conduct more thorough diagnostic evaluations, thereby reducing unnecessary antibiotic exposure.
Limitations
As a retrospective observational study, these findings show association rather than definitive causation. Furthermore, the study does not account for the complexities of real-world human-AI interaction or how clinicians might alter their behavior when presented with these scores. Future research should focus on prospective randomized trials to validate whether AI-guided timing can safely improve antibiotic stewardship in the emergency department.
Summary updated