• Artificial Intelligence in Biosciences Symposium 3
  • Back

John Hopkins Uni, Suchi Saria
Developed a TREWS early warning for sepsis. Goal, action, receives score for action. Used offline ehrs and treat as traces. Large repositories. Making generalisations then attempting to improve estimates. Observed event stream, latent system state, desired output, detector output. Estimate event probabilities. Hard to model physiological signals and lab test results, different time schedules. Different data, different sampling generalisations, systematic bias. But trying to detect and create alerting system for septic shock. Could detect shock 24 hours prior to actual shock. Got worldwide attention. Published who is responsive to which strategy JMLR 2017. Got approval to go live in 5 hospitals.
Mentions that new products using patient data are risky for investors. Evaluators don’t understand what is good or not. Area starting to mature but tech moves fast to healthcare so not rigorous evaluations which need to be. Need to explain high and low quality AI – software. Reporting habits better suited to one off evaluations on offline datasets. 

Leave your comment

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.