To map and critically synthesize the available evidence on artificial intelligence based predictive models for early sepsis detection in intensive care units, emphasizing the importance of clinical readiness.
Key Findings:
Thirty-seven studies were included, primarily using retrospective electronic health record data and machine learning techniques.
Reported performance of models varied widely, with limited external validation, raising concerns about their clinical applicability.
Few studies addressed clinical implementation, interpretability, or integration into real-time workflows, highlighting the need for further research.
Interpretation:
While AI-based models show potential for early sepsis detection, significant gaps exist in external validation, clinical integration, and real-world applicability, necessitating focused future research.
Limitations:
Most studies relied on retrospective data from single institutions or high-income countries, limiting generalizability.
Limited external or prospective validation of models raises questions about their reliability.
Inconsistent addressing of interoperability, clinician trust, and ethical considerations may hinder real-world application.
Conclusion:
Future research should prioritize methodological transparency and focus on implementation evaluations to enhance the clinical readiness of AI models for sepsis detection, addressing the urgent gaps identified.
