Machine Learning to Predict Next-day Dialysis in Critically Ill Patients

Session Number

Project ID: MEDH 13

Advisor(s)

Nikolay Markov, Catherine A. Gao, Northwestern University Feinberg School of Medicine, Division of Pulmonary and Critical Care

Discipline

Medical and Health Sciences

Start Date

17-4-2024 10:00 AM

End Date

17-4-2024 10:15 AM

Abstract

Background: Some patients in the intensive care unit (ICU) experience acute kidney injury (AKI) or other conditions requiring dialysis as a treatment to support failing kidneys. Machine learning on electronic health record (EHR) data holds promise for tasks such as helping clinical teams with planning and for prognostic purposes.

Methods: Here we investigate the likelihood of a patient’s next-day dialysis status and complications by performing predictive analysis using an XGBoost classification model.

Results: Data from 12, 495 patient days in the ICU from 585 unique patients was used. Our best model using XGBoost model achieved excellent predictive performance with an AUROC of 0.89 at predicting next-day dialysis. On transition days, when a patient’s dialysis status switched, the model performed poorly with only 0.48 AUROC.

Discussion: A machine learning model was able to predict with high performance next-day dialysis, but had trouble with transition days. Future studies can focus on longer lead windows and optimizing prediction for transition days.

Share

COinS
 
Apr 17th, 10:00 AM Apr 17th, 10:15 AM

Machine Learning to Predict Next-day Dialysis in Critically Ill Patients

Background: Some patients in the intensive care unit (ICU) experience acute kidney injury (AKI) or other conditions requiring dialysis as a treatment to support failing kidneys. Machine learning on electronic health record (EHR) data holds promise for tasks such as helping clinical teams with planning and for prognostic purposes.

Methods: Here we investigate the likelihood of a patient’s next-day dialysis status and complications by performing predictive analysis using an XGBoost classification model.

Results: Data from 12, 495 patient days in the ICU from 585 unique patients was used. Our best model using XGBoost model achieved excellent predictive performance with an AUROC of 0.89 at predicting next-day dialysis. On transition days, when a patient’s dialysis status switched, the model performed poorly with only 0.48 AUROC.

Discussion: A machine learning model was able to predict with high performance next-day dialysis, but had trouble with transition days. Future studies can focus on longer lead windows and optimizing prediction for transition days.