Sepsis is a life-threatening condition that arises when the body’s response to infection results in organ dysfunction. Despite advances in modern medicine, sepsis remains a pressing healthcare issue, contributing to nearly 20% of all deaths worldwide in 2017. Its impact is especially severe among children under five and individuals living in low- and middle-income countries. Early and accurate detection remains a major challenge, largely due to the absence of reliable biomarkers and the often complex, nonspecific clinical signs associated with sepsis.
Conventional diagnostic methods – ranging from blood cultures to biochemical markers such as CRP and procalcitonin - are either too slow or insufficiently specific for timely intervention. The importance of early detection cannot be overstated, as each hour of delay increases the mortality risk.
This study investigates how Artificial Intelligence (AI) can be integrated into hematology workflows to address these challenges. By deploying the HORIBA Yumizen H series analyzers in combination with the Generative Manifold Learning (GML) framework developed by GeodAIsics, we aim to revolutionize sepsis detection, using only Complete Blood Count (CBC) data.
The AI model was trained using a multi-center dataset comprising 687 patients from Christian Medical College (India) and Bergonie Institute (France), spanning cases of Sepsis, SIRS, Septic Shock, and various hematological pathologies.
Key Components of the AI Framework:
Density-Based Probability Scoring: for every test sample, the model assigns a severity score. Scores above a user-defined threshold are classified as septic. Using lower thresholds will increase sensitivity, but reduce specificity.
Tri-Classification Atlas: the model differentiates between Sepsis, SIRS, and Septic Shock, offering a nuanced view of disease progression
Instrumentation:
HORIBA Yumizen H2500 & H1500 analyzers were used to collect CBC data. These devices are positioned to serve both high-end hospital labs and point-of-care settings in developing regions.
The AI model demonstrated high diagnostic accuracy, outperforming even FDA-authorized platforms such as the Sepsis ImmunoScore.
Model Performance Metrics:
| Comparison | Sensitivity | Specificity | AUC |
|---|---|---|---|
| Sepsis vs. Non-Sepsis | 91% | 86% | 0.89 |
| SIRS vs. Non-Sepsis | 84% | 90% | 0.87 |
| Sepsis/SIRS/Shock vs. Normal | 94% | 94% | 0.91 |
| Sepsis vs. Pathologies | 92% | 89% | 0.88 |
These results were validated across multiple instruments and patient cohorts, demonstrating robust generalizability.
Comparison with Sepsis ImmunoScore:
The ImmunoScore, based on immunological and biochemical markers, achieved an AUC of 0.81 in external validation. While effective, it requires specialized assays and higher costs, making it less accessible for widespread use.
In contrast, the AI model uses routine CBC data, enabling cost-effective, rapid, and scalable deployment, especially where resources are limited.
Clinical Applications:
Product Strategy:
HORIBA’s dual-instrument approach - Yumizen H2500 for mature markets and Yumizen H1500 for developing regions - ensures broad accessibility. The integration of AI flags and parameters into these devices enhances their diagnostic value without the need for additional reagents or hardware.
Regulatory Pathway:
Next Steps:
Biochemical Markers
Biochemical markers such as CRP, Procalcitonin (PCT), and Lactate, are widely used in clinical settings to detect inflammation and infection. While they offer rapid results, their specificity for sepsis is limited because elevated levels may also occur in other conditions such as trauma or autoimmune diseases.
| Marker | Description | Advantages | Limitations | Use Case |
|---|---|---|---|---|
| CRP | Acute-phase protein elevated in inflammation | Widely available, high sensitivity | Low specificity | General inflammation screening |
| Procalcitonin (PCT) | Rises in bacterial infections | Better specificity than CRP | Expensive, limited availability | ICU sepsis screening |
| Lactate | Marker of tissue hypoperfusion | Accessible, used in SOFA scoring | Non-specific | Severity assessment in ICU |
Microbiological Methods
Microbiological methods remain the gold standard for identifying pathogens responsible for sepsis. Blood cultures and site-specific cultures provide direct evidence of infection but are time-consuming and may yield false negatives in fastidious organisms.
| Method | Description | Advantages | Limitations | Use Case |
|---|---|---|---|---|
| Blood Cultures | Detects pathogens in bloodstream | Specific, guides therapy | Slow turnaround, may miss pathogens | Hospital-based diagnosis |
| Site-specific Cultures | Cultures from suspected infection sites | Targeted diagnosis | Requires clinical suspicion | Localized infection confirmation |
Immunological Assays
Immunological assays detect inflammatory mediators such as cytokines and soluble receptors. While these tests offer insights into immune response, they are not always available or routinely performed and require specialized laboratory facilities.
| Marker | Description | Advantages | Limitations | Use Case |
|---|---|---|---|---|
| IL-6, C5a | Inflammatory cytokines | Potential for early detection | Non-specific, costly | Research and specialized labs |
| Presepsin | Soluble CD14 subtype | Promising biomarker | Limited availability | Emerging diagnostic tool |
Hematological Parameters
Hematological parameters are part of routine CBC tests and include WBC count, Immature Granulocytes (IG%), and Monocyte Distribution Width (MDW). These markers are accessible but often lack specificity when used alone.
| Parameter | Description | Advantages | Limitations | Use Case |
|---|---|---|---|---|
| WBC Count | Elevated or depressed in infection | Routine test | Non-specific | Initial screening |
| IG% | Indicator of bone marrow response | Used in SIRS definition | Cannot differentiate sepsis | Supportive marker |
| MDW | FDA-approved early sepsis indicator | Promising in ED | Reagent-dependent | Emergency triage |
Imaging & Clinical Scoring Systems
Imaging and scoring systems such as SOFA and qSOFA are used to assess organ dysfunction and sepsis severity. While useful in clinical decision-making, they rely on multiple parameters and may not be feasible in all settings.
| Tool | Description | Advantages | Limitations | Use Case |
|---|---|---|---|---|
| SOFA Score | Sequential Organ Failure Assessment | Comprehensive severity assessment | Requires multiple lab values | ICU monitoring |
| qSOFA | Quick SOFA for bedside use | Simple, fast | Lower sensitivity | Emergency department triage |
| Imaging (CT, X-ray) | Detects infection source | Visual confirmation | Not diagnostic for sepsis | Supportive diagnosis |
| Criteria | Traditional Diagnostic Methods | AI-Driven Hematology (HORIBA + GeodAIsics) |
|---|---|---|
| Speed of Diagnosis | Blood cultures: 24–72 hours; Biomarkers: 1–3 hours; Imaging: variable | Real-time scoring during CBC analysis; Results within seconds |
| Specificity | CRP: Low; PCT: Moderate to High; MDW: Moderate; Cultures: High (if positive) | High specificity (up to 94%); Differentiates Sepsis, SIRS, and Septic Shock |
| Sensitivity | CRP: High; PCT: High; MDW: Moderate; Cultures: Variable | Very high sensitivity (up to 91%); Robust across patient cohorts |
| Cost per Test | PCT: $25–$50; Molecular diagnostics: $300–$3,000; Blood cultures: $30–$50 | Uses routine CBC data; No additional reagents; Cost-effective |
| Infrastructure Requirements | Biochemistry/immunology labs; Skilled technicians; Culture facilities | Only hematology analyzers; No special reagents; Deployable in primary care |
| Accessibility in LMICs | Limited due to cost and infrastructure | High accessibility; Ideal for decentralized healthcare |
| Scalability | Challenging due to reagent cost and lab dependency | Highly scalable; Software-based deployment |
| Clinical Workflow Integration | Requires multiple tests and coordination | Seamless integration into CBC workflow; Immediate flagging |
| Regulatory Status | MDW: FDA-approved; PCT: widely accepted; Cultures: gold standard | RUO phase; IVDR and FDA pathways in progress |
| Explainability & Interpretability | Biomarkers: known pathways; Cultures: direct pathogen ID | Digital twins and Z-score deviation; Transparent scoring |
| Patient Impact | Delayed diagnosis; Overuse of antibiotics | Early intervention; Reduces unnecessary antibiotics; Improves outcomes |
| Differentiation Capability | Often cannot distinguish between SIRS, Sepsis, and Septic Shock | Tri-classification model; Supports prognostic decisions |
| Data Requirements | Multiple sample types; Often invasive | Single CBC sample; Retrospective and prospective data compatible |
| Environmental & Operational Efficiency | High reagent usage; Energy-intensive processes | Green IT architecture; Minimal computational load |
The AI-driven hematology model represents a groundbreaking advance in sepsis detection, especially in low- and middle-income countries where access to advanced diagnostics is often limited. By leveraging existing infrastructure (CBC analyzers) and integrating AI algorithms, this solution democratizes early detection and empowers clinical decision-making at the point of care.
Key Advantages:
Cost-efficiency: eliminates the need for expensive reagents or specialized assays.
Speed: real-time scoring integrated into routine CBC analysis.
Scalability: easily deployable across primary health centers, emergency departments, and ICUs.
Regulatory Alignment: designed to meet IVDR and FDA standards
This case study illustrates how AI in hematology can transform sepsis detection from a reactive, delayed process, to a proactive, data-driven strategy. By leveraging routine CBC data, the model broadens access to advanced diagnostics, especially in low-resource environments where sepsis imposes the greatest burden.
The integration of Generative Manifold Learning, digital twins, and density-based scoring into hematology analyzers represents a paradigm shift in alignment with global health priorities and regulatory frameworks.
As the project moves toward clinical validation and regulatory approval, it stands to make life-saving impact, through earlier detection, improved differentiation, and smarter resource allocation.
Vous avez des questions ou des demandes ? Utilisez ce formulaire pour contacter nos spécialistes.
