JMIR Publications

ABSTRACT

Background:

Sepsis, a critical global health challenge, accounting for approximately 20% of worldwide deaths in 2017. Although the Sequential Organ Failure Assessment (SOFA) score standardizes the diagnosis of organ dysfunction, early sepsis detection remains challenging due to its insidious symptoms. Current diagnostic methods, including clinical assessments and laboratory tests, frequently lack the speed and specificity needed for timely intervention, particularly in vulnerable populations such as the elderly, intensive care unit (ICU) patients, and those with compromised immune systems. While bacterial cultures remain vital, their time-consuming nature and susceptibility to false negatives limit their effectiveness. Even promising existing machine learning approaches are restricted by reliance on complex clinical factors that could delay results, underscoring the need for faster, simpler, and more reliable diagnostic strategies.

Objective:

This study introduces innovative machine learning models using complete blood count with differential (CBC + DIFF) data—a routine, minimally invasive test that assesses immune response through blood cell measurements critical for sepsis identification. The primary objective was to implement this model within an artificial intelligence-clinical decision support system (AI-CDSS) to enhance early sepsis detection and management in critical care settings.

Methods:

This retrospective study at Tri-Service General Hospital (September-December 2023), analyzed 746 ICU patients with suspected pneumonia-induced sepsis (supported by radiographic evidence and a SOFA score increase of ≥2 points), alongside 746 stable outpatients as controls. Sepsis infection sources were confirmed via positive sputum, blood cultures, or FilmArray result. The dataset incorporated both basic hematological factors and advanced neutrophil characteristics (side scatter light intensity, cytoplasmic complexity, and neutrophil-to-lymphocyte ratio), with September-November data used for training and December for validation. Machine learning models, including Light Gradient Boosting Machine (LGBM), Random Forest Classifier (RFC), and Gradient Boosting Classifier (GBC), were developed using CBC + DIFF data, and were assessed using metrics such as area under the curve (AUC), sensitivity, and specificity. The best-performing model was integrated into the AI-CDSS, with its implementation supported through workshops and training sessions.

Results:

Pathogen identification in ICU patients found 243 FilmArray-positive, 411 culture-positive, and 92 undetected cases, yielding a final dataset of 654 sepsis cases out of 1,492 total cases. The machine learning models demonstrated high predictive accuracy, with LGBM achieving the highest AUC (0.90), followed by RFC (0.89) and GBC (0.88). The best-performing LGBM model was selected and integrated as the core of our AI-CDSS, which built on a web interface to facilitate rapid sepsis risk assessment using CBC + DIFF data.

Conclusions:

This study demonstrates that by providing streamlined predictions using CBC + DIFF data without requiring extensive clinical parameters, the AI-CDSS can be seamlessly integrated into clinical workflows, enhancing rapid, accurate identification of sepsis and improving patient care and treatment timeliness.