Abstract: FR-PO983

Machine Learning Model versus SOFA, APSIII, and OASIS in Prediction of Renal Replacement Therapy (RRT) within First 8 Days after ICU Admission in MIMIC-III (Medical Information Mart for Intensive Care III)

Session Information

Category: Bioengineering and Informatics

  • 101 Bioengineering and Informatics

Authors

  • Kiljanek, Lukasz, Drexel University College of Medicine, Philadelphia, Pennsylvania, United States
  • Aggarwal, Sandeep, Drexel University College of Medicine, Philadelphia, Pennsylvania, United States
Background

There is a paucity of prediction tools for new onset RRT in ICU population. Computer based Machine learning models are being employed to create risk prediction tools using large databases. In this study we attempted the use of Gradient Boosting Machine (GBM) - Machine learning algorithm to create a RRT prediction tool using MIMIC 3 database

Methods

GBM algorithm from H2O.ai R project library was used to model 2008-2012 MIMIC 3 database. We defined 750 clinical and laboratory variables from first 24 hours after ICU admission. Patients with history of dialysis or RRT related events charted within first 24 hours of ICU stay were excluded. 17045 patients were divided 30 times randomly into training (90% of patients) and testing (10% of patients) datasets. RRT within 8 days of ICU stay was defined as any RRT related event charted between 24 hours and 8 days mark. Incidence of RRT was 27.8 [95% CI 25.9-29.6] for 30 testing sets. Each training dataset was used to build GBM for predicting need for RRT, which was validated on testing dataset. Area under curve (AUC) of receiver-operator characteristics curve (ROC) was recorded. For each of 30 testing datasets, AUC for prediction of RRT by SOFA, OASIS and APS III was also recorded.

Results

For 30 testing datasets, AUC of ROC for; GBM was 0.87 [95% CI 0.85-0.89], and SOFA was 0.83 [95% CI 0.82-0.84], APSIII was 0.82 [95% CI 0.81-0.83] and OASIS was 0.71 [95% CI 0.70 0.72], in predicting RRT within 8 days. GBM vs SOFA, GBM vs APSIII and GBM vs OASIS revealed statistically significant higher AUC of GBM (p<0.05 Wilcoxon signed-rank test). Most important predictors, seen within first 24 hours of ICU stay, from 1 of 30 runs, with their scaled importances were : last serum creatinine recorded (1), maximal serum creatinine recorded (0.94)

Conclusion

In our analysis GBM model showed statistically superior accuracy compared to conventional ICU severity scores to predict patients who survive until and require RRT after ICU admission. Additionally, we discovered applicability of conventional ICU severity scores to predict RRT in this setting. This can be employed to EMRs in hospitals for as an RRT event prediction too, but more research is warranted to assess clinical aplicability and robustness of our methods.