Kidney Week

Abstract: PO0008

Predicting Intra-ICU Mortality Using Machine Learning Algorithms in Patients Who Require Acute Renal Replacement Therapy in a Critical Care Unit

Session Information

• AKI Epidemiology, Risk Factors, and Prevention: Clinical Research
  October 22, 2020 | Location: On-Demand
  Abstract Time: 10:00 AM - 12:00 PM

Category: Acute Kidney Injury

• 101 AKI: Epidemiology, Risk Factors, and Prevention

Authors

• Chang, Hsin-Hsiung, National Cheng kung College of Computer Science and Information Engineering, Tainan, Taiwan

Hsin-Hsiung Chang,

• Chen, Huiwen, University of Pittsburgh, Pittsburgh, Pennsylvania, United States

Huiwen Chen,

• Jhamb, Manisha, University of Pittsburgh, Pittsburgh, Pennsylvania, United States

Manisha Jhamb,

Background

The Acute Physiology, Age, Chronic Health Evaluation (APACHE) and Simplified Acute Physiology Score(SAPS) are not a fair mortality prediction tools for patients receiving acute renal replacement therapy (RRT) in the ICU. Our objective was to develop a series of machine learning models to predict intra-ICU mortality for patients requiring acute dialysis therapy using data obtained one day prior to RRT initiation.

Methods

We extracted data on patients who commenced acute RRT captured in eICU and MIMIC databases. We trained machine learning models by using the eICU and MIMIC dataset. The validity of those models was then evaluated by using MIMIC and eICU.

Results

The eICU cohort included 2,360 patients and MIMIC included 1,274 patients who met our eligibility criteria.Thediscrimination power of the model was evaluatedby calculating the area under the receiver operating characteristic curve (AUC). The intra-ICU mortality AUC of the Sequential Organ Failure Assessment (SOFA) score using data collected one day before RRT initiation was 0.803 (95% CI 0.777-0.829) in eICU and 0.683 (95% CI 0.636-0.729) in MIMIC. The intra-ICU mortality AUC of machine learning models using logistic regression (LR), XGBoost, random forest (RF) and multilayer perceptron (MLP) trained in eICU cohort were 0.858 (95% CI 0.850-0.867), 0.858 (95% CI 0.850-0.866), 0.859 (95% CI 0.848-0.870), 0.864 (95% CI 0.851-0.876) and validated in MIMIC were 0.799 (95% CI:0.775-0.824), 0.809 (95% CI:0.786-0.833), 0.814 (95% CI 0.791-0.837), 0.800 (95% CI 0.776-0.825), respectively. When training the models using MIMIC dataset, the intra-ICU mortality AUCof LR, XGBoost, RF and MLP were 0.818 95% CI (0.786-0.852], 0.821 (95% CI 0.787-0.856) 0.822 (95% CI 0.791-0.854), 0.827 (95% CI 0.795-0.858), respectively. Validating these models using eICU dataset, the AUC of LR, XGBoost, RF and MLP were 0.846 95% CI:0.828-0.864, 0.847 (95% CI 0.829-0.865), 0.853 (95% CI 0.835-0.870), 0.846 (95% CI 0.828-0.865), respectively.

Conclusion

In this study, we designed machine learning models to make intra-ICU mortality prediction for patients who required RRT. Our models correlated better than SOFA score in predicting the mortality of patients requiring RRT in ICU. All of the models almost had excellent performance in both databases.