New Open-Source Tool Aims to Make Sepsis Prediction Models More Transparent for Clinicians
A new open-source platform enhances the interpretability of CBC-based machine-learning models for sepsis prediction, offering improved transparency and clinical utility.
Laboratory leaders watching the evolution of clinical AI tools now have a new development to track: a fully open-source web application designed to make machine-learning–based sepsis prediction more interpretable and accessible.
A new study in The Journal of Applied Laboratory Medicine introduces SBC-SHAP, an interactive platform that visualizes how complete blood count (CBC) values influence individual patient risk scores.
The researchers say the tool was created to address a longstanding barrier: although CBC-based sepsis prediction models have shown promise, clinicians often struggle to understand why those models flag a patient as high-risk. As the study notes, prior lab approaches “do not explain how a specific value, such as white blood cell count, contributes to risk predictions.” Compounding the issue, some existing tools “required programming expertise that many clinicians lack.”
A New Way to View CBC-Based Sepsis Risk
Early sepsis detection remains a priority for hospitals, given that faster recognition leads to earlier intervention and improved outcomes. CBC parameters—white blood cells, red blood cells, hemoglobin, mean corpuscular volume, platelets—are routinely available and could serve as early indicators. Yet, the complexity of machine-learning (ML) algorithms has limited the usefulness of CBC-based prediction.
To tackle this issue, the research team developed a graph-based approach that incorporates time-series information into ML models, enabling predictions that account for trends rather than isolated values. The authors also evaluated whether adding specific ratios to a healthy reference baseline could bolster performance.
Photo credit: Image by Gerd Altmann from Pixabay.
According to the paper, the new approach “increased the sensitivity at 80% specificity across all ML models from 78.2% to 82.9% on an internal dataset.” When tested on an external dataset from an independent tertiary-care center, sensitivity improved “from 65.4% to 73.4%.”
Interpretable Outputs Meant for Real Clinical Use
The web tool itself, dubbed SBC-SHAP, was built to interpret the ML predictions. It offers clinicians an interactive view showing how age, sex, and specific CBC values contribute to a risk score.
In an example, the authors highlight that the tool “breaks out patient age, sex, and hemoglobin values and how white blood cells, red blood cells, and mean corpuscular volume indicate patient risk.” Another example shows how clinicians can adjust or “correct prior values for a more accurate estimate of sepsis risk.”
Users can drill into individual CBC measurements, view explanations for each predicted risk value, and filter cases depending on whether particular test results are available. As the authors put it, the platform allows clinicians to “investigate how specific feature values contribute to predicted sepsis risks” and tailor the view for “diverse use cases.”
The tool is open-source and freely accessible, which researchers say is central to making
ML-enhanced diagnostics feasible for everyday practice.
SBC-SHAP is available online at: mdoa-tools.bi.denbi.de/sbc-shap.
What Lab Leaders Should Take Away
For laboratory directors and diagnostic strategy teams, SBC-SHAP illustrates where AI in laboratory medicine is heading. Leaders are now seeing tools that not only predict, but also explain. As clinical teams grow increasingly cautious about “black-box” algorithms, transparent ML models that show their work may gain faster adoption.
This study also reinforces the ongoing importance of well-curated CBC data. By leveraging widely available hematology parameters, the authors demonstrate that meaningful AI-driven risk stratification does not always require advanced or specialty assays.
As labs plan their digital and clinical decision-support strategies, tools like SBC-SHAP signal a shift toward accessible, clinician-friendly ML applications. The researchers highlight the tool’s potential to support real-time decision-making, noting that it allows users to explore predictions and feature contributions without needing specialized programming skills.
—Janette Wider
