Impact of a Deep Learning Model Predicting Sepsis on Quality of Care and Survival

Prediction Model for Sepsis Detection

Prevention and recognition of sepsis are part of the daily care of patients in intensive care units. Sepsis is characterized by its heterogeneity, which makes it difficult to detect. The cited study summarizes the findings from the involvement of a deep learning model for early detection of sepsis with the aim of early warning of healthcare providers and subsequent intervention using “care bundles” to prevent sepsis.

Deep learning, or “deep learning,” is a branch of machine learning based on artificial neural networks and approaching artificial intelligence. In recent years, deep learning has found applications in many medical fields.

To improve early detection, a team from the University of California, San Diego, applied predictive analytics, specifically the deep learning model COMPOSER, which in real-time imports data from electronic health records to predict sepsis before its obvious clinical manifestation. COMPOSER was designed to verify its outputs with previous training samples to increase reliability. This significantly reduced the number of false-positive alerts, contributing to greater nurse confidence in the alerts and their compliance with subsequent recommended procedures.

Monitored Parameters

The aim of the quasi-experimental study before the implementation of the COMPOSER model in two different emergency departments within the UC San Diego health system and after its implementation was to assess the impact of this model in terms of early sepsis prediction and patient health status. Upon positive evaluation of sepsis status, COMPOSER triggered a notification in the electronic best practice advisory (BPA) system, which is delivered to nurses in the form of pop-up alerts. 

Hospital mortality, compliance with the “bundle” for sepsis prevention, 72-hour changes in organ dysfunction quantified by the SOFA score after sepsis onset, days without placement, and the number of interactions in the intensive care unit were evaluated before (705 days) and after intervention (145 days). The SOFA scoring system provides information on organ dysfunction, with an increased SOFA score significantly correlating with morbidity and mortality in septic patients. 

Study Results

In this way, 6217 adult septic patients (5065 in the pre-intervention and 1152 in the post-intervention phase) hospitalized from January 1, 2021, to April 30, 2023, were included in the study. Most patients had comorbidities, the median SOFA score at the time of sepsis was 2, and the baseline conditions of patients before and after the implementation of the prediction model intervention did not significantly differ.

Reduction in Mortality

The average sepsis mortality during the post-intervention period was 9.49%. The deployment of the COMPOSER model was associated with a 1.9% absolute reduction (i.e., 17% relative decrease) in hospital sepsis mortality (95% confidence interval [CI] 0.3–3.5%). A 4% decrease in the SOFA score (95% CI 1.1–7.1%) was also recorded, representing an improvement in the condition of septic patients in terms of mortality and morbidity. 

Improvement in Nurse Compliance

During the post-intervention period, the average compliance rate was 53.42%, which represents a 5% absolute (and 10% relative) improvement in compliance with the sepsis prevention “bundle” (95% CI 2.4–8.0%). Similar results were noted in both emergency departments, with significant improvements in the parameters of antibiotic administration, repeated lactate sampling, and fluid resuscitation. 

Conclusion

Early intervention implemented as soon as possible has significant benefits for the patient, making early recognition of sepsis crucial. This recently published study evaluating the application of a deep learning-based model in practice demonstrated significant increases in compliance with recommended protocols by the healthcare personnel and significant reductions in sepsis mortality. Therefore, the use of these models should be beneficial in everyday clinical conditions.

(lexi)

Sources:
1. Boussina A., Shashikumar S., Malhotra A. et al. Impact of a deep learning sepsis prediction model on quality of care and survival. NPJ Digit Med 2024; 7 (1): 14, doi: 10.1038/s41746-023-00986-6.
2. Holub M. Definition of sepsis and septic shock. Klinická biochemie a metabolismus 2018; 26 (2): 76–78.
3. Jiřík M., Moulisová V., Hlaváč M. et al. Artificial neural networks and computer vision in medicine and surgery. Rozhledy v chirurgii 2022; 101 (12): 564–570.
4. Litjens G., Kooi T., Bejnordi B. E. et al. A survey on deep learning in medical image analysis. Med Image Anal 2017; 42: 60–88, doi: 10.1016/j.media.2017.07.005.