JMIR Publications

ABSTRACT

Background:

Given the critical role of real-time prediction models in the clinical diagnosis and management of sepsis, and the limitations of existing machine learning (ML)-based sepsis prediction models—such as poor real-time performance and insufficient interpretability—this study aims to develop a real-time sepsis prediction model that integrates high timeliness and clinical interpretability.

Objective:

The model is designed to dynamically predict the risk of sepsis in ICU patients and establish a practical, tailored sepsis prediction platform.

Methods:

Within a retrospective analysis framework, the model comprises a real-time prediction module and an interpretability module. The real-time prediction module leverages 3-hour dynamic temporal features derived from eight non-invasive, real-time physiological indicators: heart rate (HR), respiratory rate (RR), blood oxygen saturation (SpO2), mean arterial pressure (MAP), systolic blood pressure (SBP), diastolic blood pressure (DBP), body temperature (Temp), and blood glucose (BG). Three linear parameters (mean, standard deviation, and endpoint value) were calculated to construct the prediction model using multiple ML algorithms. The interpretability module employs the Tree-based Shapley Additive Explanations (TreeSHAP) method to enhance model transparency through both individual prediction and global explanations. Finally, a web-based platform was developed to integrate prediction and interpretability functions.

Results:

The sepsis prediction model demonstrated robust performance in the test cohort (224 patients), achieving an Accuracy of 0.70 (95% CI: 0.68–0.71), Precision of 0.69 (95% CI: 0.68–0.71), F1-score of 0.69 (95% CI: 0.67–0.70), and Area under the ROC curve (AUROC) of 0.76 (95% CI: 0.74–0.77). The six most influential features for sepsis risk were temperature fluctuation coefficient (var-TEM), MAP fluctuation coefficient (var-MP), mean heart rate (mean-HR), mean respiratory rate (mean-RR), SpO2 fluctuation coefficient (var-SPO2), and SBP fluctuation coefficient (var-SBP). The TreeSHAP method effectively visualized feature contributions, enabling clinicians to interpret the model’s prediction logic and identify anomalies. The web-based platform (https://sepsis-vision-explanation.streamlit.app/) significantly enhanced clinical utility by providing real-time risk assessment, statistical summaries, trend analysis, and actionable insights.

Conclusions:

This platform provides real-time dynamic warnings for sepsis risk in critically ill ICU patients, supporting timely clinical decision-making.