A Rule out Approach to Sepsis Risk Stratification Using Masked Autoencoders in Low Resource Settings

Abstract

Sepsis continues to impose a heavy global burden, particularly in low-resource settings where early recognition is critical but difficult due to data limitations. Most predictive models require richly annotated, high-frequency inputs unavailable in such environments. This study reframes sepsis risk prediction as a rule-out task and proposes a self-supervised masked autoencoder (MAE) trained on the Sepsis Survival Minimal Clinical Records (SSMCR) dataset comprised of just age, sex, and survival outcome. Unlike traditional supervised approaches, the MAE learns latent structure from unlabelled data by reconstructing masked features, then fine-tunes on a minimal labelled subset to predict nine-day survival. We benchmark its performance against logistic regression, decision tree, random forest, and support vector machine classifiers using standard evaluation metrics, with specific emphasis on negative predictive value (NPV) and specificity. Results show that the MAE possesses the highest NPV (93.1%) and specificity (68.2%) and is therefore best for use in safely excluding sepsis in resource-limited settings. The results are evidence that even with few features, self-supervised models can generate clinically relevant predictions. The work paves the way for future real-world deployment in accordance with the Technology Acceptance Model and WHO SMART guidelines. This work makes AI-enabled diagnostics more accessible where conventional methods are not possible or unavailable.