Predicting Pediatric Sepsis and Mortality Using Wearable Device Data and Machine Learning in Bangladesh.

Sepsis disproportionately impacts children in low-resource settings, where diagnostic tools like the Phoenix Sepsis Score (PSS) are constrained by reliance on laboratory testing. The objective of this research was to evaluate the use of continuous physiological data from low-cost wearable biosensors and machine learning models to predict pediatric sepsis, septic shock, and mortality in a low-resource, intensive care setting. This prospective observational single-site study analyzed 96 pediatric intensive care unit patients with suspected sepsis in Dhaka, Bangladesh. Physiological data were collected using a wearable biosensor patch, whereas clinical exams, laboratory tests, and PSS criteria identified sepsis, septic shock, and mortality. Least absolute shrinkage and selection operator (LASSO) regression models were developed and validated through leave-one-group-out cross-validation (LOGO-CV) using biosensor data. Our clinical diagnostic model for sepsis using biosensor-only features demonstrated an area under the receiver operating characteristic curve (AUROC) of 0.78 (null AUROC = 0.58). For septic shock, the model demonstrated an AUROC of 0.85 (null AUROC = 0.61). The mortality model demonstrated an AUROC of 0.87 (null AUROC = 0.64). Sensitivity analyses showed improvement of AUROC to 0.89 for prediction of sepsis with manual recorded oxygen saturation (SpO2) included. Although models were trained and tested retrospectively with internal validation, findings demonstrate the potential of wearable biosensors to support pediatric sepsis diagnosis without reliance on advanced diagnostics. These results encourage further external validation with larger, multisite cohorts and real-time mobile health (mHealth) integration to support clinical use in low-resource settings.