Lukasz R Kiljanek, Medstar Shah and Sandeep Aggarwal
Background: Prediction tool for incident renal replacement therapy (iRRT) use could potentially improve outcomes in ICU population. We used the data from the Medical Information Mart for Intensive Care III (MIMIC III) database to create artificial intelligence (AI) iRRT use prediction model.
Methods: Based on routinely collected data in ICU we identified and engineered 679 candidate predictors of iRRT use. The iRRT was defined as any dialysis-related event charted in the electronic medical record (EMR) within the seven days following the first 24 hours of ICU admission. ICU stays of patients on dialysis before ICU admission, and with dialysis-related events charted before the end of first 24 hours were excluded. Remaining 18379 ICU stays were randomly divided 400 times, into training and testing datasets. For each random training dataset, AI-model for iRRT prediction was trained. Predictions of AI, SOFA, OASIS, and APSIII, were validated on testing dataset against the known use of iRRT with the area under the curve (AUC) of receiver-operator characteristics curve (ROC) recorded.
Result: For all 400 iterations, AUC of ROC for AI-model was 0.88 [95% CI 0.88-0.89] and was higher than SOFA, APSIII and OASIS: 0.82 [0.82-0.82], 0.81 [0.81-0.81], and 0.7 [0.69-0.7] AUCs respectively (p<0.001).
Conclusion: AI-model was accurate in predicting patients who survive until, consent and undergo iRRT after ICU admission. High AUC for the AI model trained only on data from first 24 hours of ICU stay emphasizes the importance of initial ICU management on renal outcomes.
PDFShare this article
Journal of Health & Medical Informatics received 2128 citations as per Google Scholar report
