Learning Patient Health Conditions From Physiological Responses to Vasoactive Agents.

This work analyzes patients' transient physiological responses to vasoactive infusions during their ICU stay, in order to assess their underlying health conditions. To achieve this, we developed a medical-aware representation model that captures patient health states from demographic information, bedside vital signs, and the administration of vasoactive infusions. In the representation space, we clustered patients with a self-organizing map and observed that the encoded representation captures critical information related to patients' future outcomes. This representation space can be integrated into various downstream tasks. For mortality prediction, it detects 85% of expired patients with a specificity of 0.58.Clinical relevance- Our model provides a medicine-aware patient monitoring that alerts when patients' vital signs fail to respond as expected after receiving vasoactive infusions.