Data-Driven Early Sepsis Prediction: A Systematic Review and Critical Appraisal of CBC-Based Models, Deep Time-Series Learning, Explainable AI, and Generalizability Challenges

Sepsis ranks among the major factors for in-hospital mortality. The delay in sepsis recognition has remained a major factor for preventable organ failures and mortality. The commonly used screening tools for sepsis include SIRS, qSOFA, and SOFA. These tools show poor sensitivity in the early stages of sepsis. In recent times, the development of electronic health records has improved machine learning and deep learning algorithms for the prediction of sepsis in its early stages. This systematic review aims to evaluate current data science methods for the early prediction of sepsis, focusing on five domains: [۱] complete blood count-based machine learning models, [۲] time-series deep learning algorithms, [۳] explainable artificial intelligence, [۴] methodology quality and bias assessment, and [۵] generalization and external validation. A systematic literature search was carried out through PubMed, Scopus, Web of Science, and IEEE Xplore databases for literature published from ۲۰۱۵ to January ۲۰۲۵. The systematic literature review included literature on machine learning-based predictive models for sepsis developed or validated using real-world clinical datasets. Results show that machine learning algorithms are superior to traditional methods for discrimination performance. The results show that machine learning algorithms have an AUROC > ۸۰%. However, it has also shown that there are many barriers for machine learning algorithms for the prediction of sepsis. These include methodology quality and bias assessment, calibration evaluation, and prospective impact assessment.

کلیدواژه ها:

Sepsis prediction ، Machine learning in healthcare ، Electronic health records ، Complete blood count biomarkers ، Deep learning time-series models ، Explainable artificial intelligence ، External validation