UNCERTAINTY-AWARE AI FOR RELIABLE AND CLINICALLY SAFE SEPSIS PREDICTION: CURRENT EVIDENCE AND RESEARCH GAPS

Abstract

Sepsis is a potentially life-threatening clinical state which needs to be detected as early as possible to ensure timely intervention and improved outcomes. The use of artificial intelligence (AI) and machine learning (ML) for the early identification of sepsis, based on routinely collected clinical data, has shown promise. Clinically safe or trustworthy decision-making is not a given with predictive performance. Despite the remarkable progress of AI sepsis predictive algorithms, they still have several drawbacks, including poor calibration, poor generalizability, uncertainty in the algorithm's prediction and absence of external validation. This work offers a thorough assessment of the various AI approaches for sepsis prediction, including their performance, uncertainty quantification, calibration, and clinical safety. The review's compass are three research questions regarding the AI models used, the performance of the models used for predictions and calibrations, and the main research gaps and future research directions. The suggested review strategy includes a planned literature search procedure, eligibility criteria, a method to select literature for review and a qualitative synthesis of data. From the practical clinical applications, the first literature review highlights the importance of independent validation, model robustness, uncertainty-aware prediction and assessment. Additionally, it shows that clinical reliability can be attained without relying just on conventional performance metrics. The most important areas for future study are the use of calibrated prediction probabilities, explicit statements of uncertainty, prospective and multicenter validation, and clinically important assessment of safety. The purpose of this review is to inspire the development of more trustworthy AI systems to be used in clinical practice and for early sepsis detection in a responsible way.

Keywords: Sepsis prediction, artificial intelligence, machine learning, uncertainty quantification, calibration, external validation, clinical safety.