AI for ICU analytics

Abstract

The healthcare system is overburdened. There are just 1.34 doctors per 1000 patients in India. Intensive Care imposes the highest burden on healthcare systems all across the globe. A significant fraction of lives lost and expenditure is attributed to complications such as hemodynamic shock, which are reversible if treated early. Artificial Intelligence-based early predictions can potentially save lives. Generalization across geographical contexts is a challenge. In this study, we approach generalizable, real-time prediction of new-onset shock in the ICU vitals data. The eICU dataset was used to develop generalizable models. We externally validated these models on a cohort on the MIMIC III database consisting of 5 million patient hours and a pediatric ICU in New Delhi with over 1.5 million patient hours. We predicted shock up to 8 hours in advance.