Abstract: To be useful in ICU settings, predictive models must be informed by temporally evolving features that align with changes in clinical status. We applied the SHAP method to identify features most informative to our NICU sepsis prediction model. Findings indicate the model has dependence on temporally changing lab values and vital signs, but age and central line presence were most informative. This suggests future modeling efforts should consider approaches to prioritize temporally evolving features.
Describe the new knowledge and additional skills the participant will gain after attending your presentation.: After attending this poster, attendees will learn how prediction explanation methods can improve machine learning models utilized in ICU clinical decision support tools.
Aaron Masino (Presenter)
Children's Hospital of Philadelphia
Mary Harris, Children's Hospital of Philadelphia
Robert Grundmeier, Children's Hospital of Philadelphia