TL;DR
This study introduces two deep learning models to objectively stratify sepsis patients into risk groups based on their likelihood of developing sepsis and shock. Their findings suggest that for patients at low risk of shock, delaying antibiotic administration beyond one hour may not negatively impact mortality, potentially allowing for more judicious treatment.
What Problem Does This Paper Address?
Sepsis is a life-threatening condition that requires prompt treatment, with current guidelines from the Surviving Sepsis Campaign (SSC) recommending antibiotic administration within 1-3 hours based on the patient's probability of having sepsis and the presence of shock. However, stratifying patients into these categories has historically been subjective, leading to potential inconsistencies in care. The challenge lies in objectively identifying patients who are at high risk of developing severe sepsis or septic shock, and understanding how antibiotic timing affects outcomes within these distinct risk profiles. This paper aims to bridge this gap by using deep learning to provide objective risk stratification and analyze the impact of antibiotic timing on mortality across these groups.
Methods
The researchers developed and validated two deep learning (DL) models:
- COMPOSER: This model predicts the likelihood of a patient having sepsis, categorizing them as "sepsis probable" (score ≥ 6) or "sepsis possible" (score < 6). It's a conformal multidimensional projection model previously described and validated.
- Shock-Net: This transformer-based neural network predicts the likelihood of a patient developing shock. Scores < 0.88 indicate "shock unlikely to develop," while scores ≥ 0.88 indicate "shock likely to develop."
These models were used to stratify patients into four risk groups analogous to the SSC categories. The primary outcome measured was short-term mortality, defined as a composite of in-hospital mortality and transition to hospice care. The study then analyzed the association between antibiotic timing (within 1 hour, 1-3 hours, or >3 hours from triage) and this mortality outcome within each risk group.
Datasets
The study utilized prospectively collected patient data from two large health systems:
- Development Site: University of California, San Diego (UCSD) Health, with data from January 2016 to December 2023.
- Validation Site: University of California, Irvine (UCI) Health, with data from January 2023 to August 2024.
The total number of patients included in the primary analysis was 34,163. A sub-analysis was also performed excluding patients who developed shock within the first 3 hours of ED triage.
Key Insights and Ideas
- Objective Risk Stratification: The DL models provide a more objective and reproducible way to categorize sepsis patients compared to subjective clinical assessments, which is crucial for consistent treatment.
- Timing Flexibility for Low-Risk Patients: The most striking finding is that for patients identified as having probable sepsis but a low likelihood of developing shock, the timing of antibiotic administration (1 hour vs. 3 hours) did not significantly impact mortality. This suggests that a more lenient approach might be safe for this group.
- High-Risk Group Identification: The study successfully identified a high-risk group (probable sepsis, likely to develop shock) with significantly higher mortality, highlighting a population that warrants further investigation and potentially tailored management strategies.
- Potential for Overtreatment: In the "shock unlikely and sepsis possible" group, a high percentage of patients received antibiotics, suggesting a potential for overtreatment in this lower-risk category.
Key Results
- Overall Mortality: Short-term mortality was 6.1% at the development site and 9.3% at the validation site.
- Highest Mortality Group: Patients identified by Shock-Net as having a high likelihood of developing shock (Group 1) had the highest mortality rates (23.1% at the development site).
- Antibiotic Timing and Mortality (Low-Risk Group): For patients in Group 3 (shock unlikely to develop and sepsis probable), mortality rates were similar whether antibiotics were administered within 1 hour or within 3 hours from triage (e.g., Development Site: 5.9% vs. 5.5% vs. 5.9% for <1 hr, 1-3 hrs, >3 hrs respectively).
- No Significant Difference in Timing: Across most risk groups and at both sites, there was no statistically significant difference in short-term mortality between patients receiving antibiotics within 1 hour versus after 1 hour, or within 3 hours versus after 3 hours.
Limitations and Future Work
- Generalizability: The study was conducted at two academic medical centers, and findings may not be generalizable to all settings (e.g., rural, community hospitals).
- Physician Interaction: The study did not assess how physicians interact with the DL models or how their use impacts antimicrobial prescribing and patient outcomes.
- Causality: The study describes associations, not causation, between antibiotic timing and mortality.
- Definitions: The definitions of potential sepsis, sepsis, and shock were based on prior work, and using different constructs could alter results.
- Retrospective Nature: The investigation was retrospective, limiting the ability to infer causality.
The authors suggest that future prospective studies are needed to validate these findings in diverse practice settings and to further investigate how the probability of sepsis should impact antibiotic timing recommendations. They also highlight the need for further research into optimal management strategies for the high-risk group identified.
Summary updated