Abstract
OBJECTIVE: Sepsis remains a leading cause of mortality in healthcare, requiring rapid detection and intervention. This paper presents a scalable, serverless machine learning (ML) operations architecture for near real-time sepsis risk-stratification in Emergency Department (ED) waiting rooms, where pathology data is often unavailable and recognising sepsis presents the biggest opportunity for timely treatment. METHODS: The system integrates HL7 message processing through MuleSoft in a secure Amazon Web Services (AWS) cloud environment, leveraging AWS services such as Lambda for real-time data processing and SageMaker for ML model deployment. To optimise the model's performance, the receiver operating characteristic (ROC) curve was used to evaluate different cutoff thresholds of probability across different age groups (16-35, 35-65, and 65-115), aiming for >80% sensitivity and minimise false negatives. Processed data is stored in Aurora PostgreSQL Relational Database Service and visualised in an on-premises proprietary dashboard for clinical decision support. RESULTS: Despite high reliability, with 99.7% of HL7 messages successfully processed, limitations include occasional failures due to downtime and code set mismatches, as well as peak execution times of 10.5 s under heavy loads, highlighting areas for optimisation. Model development of eligible ED encounters (n = 484,617) using XG Boost was integrated as a real-time endpoint in SageMaker. The Extreme Gradient Boosting model achieved the highest overall accuracy (0.84) and F1-score (0.80), with balanced sensitivity and specificity for our specified limited features within an ED. ROC for age groups (16-35, 36-65, 66-115), showed strong performance in all cohorts (AUCs: 0.864, 0.867, 0.806). CONCLUSION: This paper outlines the system's design, implementation, and potential for enhancing early sepsis risk-stratification through near real-time monitoring in the ED waiting room.