AI-Driven Human-Machine Interface for Early Detection and Differentiation of Sepsis, SIRS, and Septic Shock with Severity Assessment on Hematology Analyzers

► Click here to download the original poster (PDF)

Sukesh C Nair¹, Francoise Durrieu², Shubham Rastogi³.
¹Christian Medical College, Vellore, India. ²Bergonie Institute, Bordeaux, France. ³HORIBA ABX SAS, Montpellier, France

Introduction

Sepsis remains a major challenge internationally for healthcare systems with rising incidence due to poor public awareness, delay in early detection and subsequent management. In sepsis, mortality increases with every hour left untreated. The complexities of the definition of sepsis, as well as the lack of a reliable biomarker, hamper early detection of sepsis healthcare setup such as emergency department (ED). The modern HORIBA Yumizen H series (Yumizen H2500/ Yumizen H1500) augmented by AI from multi center data have the potential to provide early predictive Flagging/ Scoring system, that can be cost-efficient to quickly support the diagnosis of SEPSIS evolution and further management right from primary health centers.

Methods

To demonstrate the capabilities of AI-ML driven sepsis severity to identify patients with sepsis and measure performance of ‘Sepsis suspicion Flag / Scoring’ to reflect sepsis risk and/or severity on HORIBA Yumizen H Range hematology analyzers. Preprocessing the data and preparation of the model in identifying patients with sepsis in the cohort and measure performance. Attribute a ‘SEPSIS suspicion Flag / Scoring’ to reflect sepsis risk and/or severity. Retrospective data obtained from the hematology analyzers (n=687), patients with either sepsis, SIRS, or septic shock (Table 1) and controls from the participating centers, Christian Medical College, Vellore. Tamil Nadu. India and Bergonie Institute, Bordeaux, France between 2018 to 2023 subjected to AI-ML modeling.

Table 1: Sample data set for sepsis and other pathologies

Results

The severity and risk of sepsis are determined using a density-based probability approach. Each test data point is assigned a severity score, with scores above a user-defined threshold classified as septic (Fig. 1&2).

Figure 1 & 2: Severity scoring of sepsis determined by density-based probability

Notably, lower thresholds result in higher sensitivity but lower specificity. Our model has demonstrated the ability to predict and classify patients with Sepsis, SIRS, and Septic shock with an impressive sensitivity of 91% and specificity of 86% (AUC = 0.89) as presented in Table 2. The current model performance is created exclusively on CBC hematology parameters which also outperforms the established benchmark FDA authorized AL/ML- Sepsis Immunoscore system (Table 3). Furthermore, the model’s capability to differentiate between sepsis, SIRS, and septic shock will be discussed in detail.

Table 2: Model performance of AI-ML from CBC sample data performed on HORIBA hematology analyzer Yumizen H2500, Yumizen H1500 for sepsis differentiations v/s normal & non-sepsis samples

Table 3: Current model performance v/s FDA authorized AI/ML -Sepsis Immunoscore

Conclusion

AI-ML driven modeling analysis using only Complete Blood Count data, have shown potentially to revolutionize sepsis detection, diagnosis, and treatment in critical care settings at very early stages. AI algorithms on Hematology analyzers can help identify patients at high risk of developing sepsis, right from primary healthcare centers and doctor’s clinics to intervene earlier and prevent conditions progression.

PDF

1.58 MB