Advancements in sepsis detection: A review of supervised and unsupervised machine learning methods versus traditional approaches

Sepsis remains one of the most common causes of mortality globally, demonstrating the urgent need for rapid and accurate diagnosis to improve patient outcomes. Recent advances in sepsis prediction have attempted to incorporate machine learning (ML) methods and traditional clinical diagnosis to identify and intervene early in the care process. This review compares supervised learning approaches, such as decision trees and support vector machines, with unsupervised approaches, including clustering and anomaly detection, to identify sepsis in complex, multimodal, real-time clinical datasets. Supervised learning approaches generally yield higher predictive accuracy when trained on labeled datasets. However, unsupervised learning approaches have the utility of identifying new patterns and subtle physiological changes without labeled training data, resulting in improved sensitivity of the system. Given the acute and time-sensitive nature of sepsis, sensitivity is the most critical performance measure, as missing a valid case can be deadly. Overall accuracy is also essential, as is model interoperability (the ability for various systems to integrate, likely due to the heterogeneous nature of health systems). These elements are crucial for scalability and trust in the clinical workspace. In this review, we further compare and contrast these ML approaches to traditional scoring/boundary biomarker approaches, discuss the difficulties of integration, and suggest ways for developing clinically deployable, interpretable, and sensitive sepsis detection systems.