Impact of Deep Learning Sepsis Prediction Model on Nursing Quality and Patient Survival

Research Background

Sepsis is a systemic inflammatory response caused by infection, affecting approximately 48 million people globally each year, with around 11 million deaths. Due to the heterogeneity of sepsis, early identification often faces significant challenges. Early interventions, such as fluid resuscitation, antibiotic management, and infection source control, are notably effective in the early stages of the disease. Therefore, enhancing early detection of sepsis through predictive analysis is of great importance.

Research Origin

This study was conducted by Aaron Boussina, Supreeth P. Shashikumar, Atul Malhotra, Robert L. Owens, Robert El-Kareh, Christopher A. Longhurst, Kimberly Quintero, Allison Donahue, Theodore C. Chan, Shamim Nemati, and Gabriel Wardi. The authors are from the Department of Medicine and the Department of Emergency Medicine at the University of California, San Diego. The research article was published in the 2024 issue of the npj Digital Medicine journal.

Research Process

Study Design and Subjects

The study employed a quasi-experimental pre-and-post control design, conducted at two emergency departments of the UC San Diego Health System from January 1, 2021, to April 30, 2023. A total of 6217 adults were included to evaluate the impact of Best Practice Alerts (BPA) generated by the deep learning model Composer on patient outcomes.

Study Steps

1. Data Collection and Processing

   • Data available from electronic health records (EHR) was extracted, including emergency admissions, laboratory tests, vital signs, patient demographic information, comorbidities, and concurrent medications.
   • Using the Composer algorithm, a real-time data analysis was conducted. The algorithm, a feedforward neural network model, predicts the onset of sepsis within the next four hours. The model uses a “conformity prediction” method, marking data as “uncertain” when anomalies or unknown samples are detected.

2. Alert Generation and Nursing-Focused BPA

   • When the model predicts a high risk of sepsis, alerts are sent to nurses, along with the model’s primary recommendations. Nurses can choose appropriate actions (such as notifying a doctor immediately or initiating sepsis treatment/checks).

3. Evaluation of Results

   • The primary evaluation metric was in-hospital sepsis mortality rate. Secondary metrics included adherence to sepsis treatment protocols, changes in SOFA scores within 72 hours, ICU admission rates, and ICU-free days.

Data Analysis and Results

1. Baseline Characteristics

   • The study included 6217 subjects, with 5065 in the pre-intervention period and 1152 in the post-intervention period. There were no significant differences in baseline characteristics such as age, gender, race, and comorbidity index between the pre-and-post intervention groups.

2. Alert Data

   • During the post-intervention period, an average of 235 alerts were generated per month, of which 55% were marked by nurses as “immediate doctor notification.”

3. Causal Impact Analysis

   • Using a Bayesian structured time series model to evaluate the causal effect before and after the intervention, results showed an absolute reduction of 1.9% in in-hospital sepsis mortality rate (relative reduction of 17%), an absolute increase of 5.0% in adherence to sepsis protocols (relative increase of 10%), and a 4% decrease in changes to the SOFA score within 72 hours.

Summary of Results

After deploying the Composer algorithm, early prediction of sepsis significantly reduced in-hospital mortality rates and improved compliance with sepsis care processes. Specifically: - In-hospital mortality rate: Decreased from an expected 11.4% (95% confidence interval 9.8%-13.0%) to an actual 9.5%, a significant reduction of 1.9%. - Sepsis protocol adherence: Increased from 48.4% to 53.4%, a significant increase of 5.0%. - Organ function damage: The change in SOFA scores within 72 hours decreased from an expected 3.71 to an actual 3.56, a reduction of 4%.

Research Significance

This study is the first to validate the effectiveness of a deep learning model for early sepsis prediction in an emergency department environment, showing significant clinical relevance and application value. The results indicate that using AI models can enhance the early identification of sepsis, thereby improving patient management processes and outcomes.

Highlights

• Innovative Approach: The study introduces a deep learning and conformity prediction method that significantly reduces false alarms and improves model reliability.
• Significant Outcomes: This is the first demonstration of the clinical application of a deep learning model, achieving goals of reducing mortality rates and improving care processes.
• Broad Application Prospects: The study establishes a solid foundation for further application in multicenter and diverse patient populations, with strong potential for widespread implementation.