Abstract: FR-PO054
The Use of Machine Learning to Predict the Renal Replacement Therapy-Free Survival in Patients Who Require Continuous Renal Replacement Therapy
Session Information
- AKI: Clinical Outcomes, Trials
November 08, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Acute Kidney Injury
- 102 AKI: Clinical, Outcomes, and Trials
Authors
- Pattharanitima, Pattharawin, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Jaladanki, Suraj, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Duffy, Aine, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Paranjpe, Ishan, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Van Vleck, Tielman T., Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Coca, Steven G., Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Chan, Lili, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Nadkarni, Girish N., Icahn School of Medicine at Mount Sinai, New York, New York, United States
Background
AKI in critically ill patients is common and continuous renal replacement therapy (CRRT) is the preferential mode of renal replacement therapy patients who are hemodynamically unstable. Prior studies have yielded conflicting results for predictors of CRRT discontinuation and mortality. Therefore, we tested machine learning algorithms for predicting renal replacement therapy-free survival (RRTFS) in patients who required CRRT
Methods
We used the Medical Information Mart for Intensive Care III database to identify patients ≥18 years old, and who had AKI requiring CRRT for ≥24 hours. ESRD patients were excluded. RRTFS was defined as patients who were discharged alive and did not require RRT 7 days prior to hospital discharge. Five machine learning algorithms: the multi-layer perceptron neural network (MLP), random forest (RF), support-vector machine (SVM), logistic least absolute shrinkage and selection operator (LASSO) and logistic regression were trained. We evaluated model performance using area under the receiver operating characteristic (AUROC). Features included laboratory values and patient features and were selected for inclusion based off of prior published data.
Results
Out of 566 patients, 179 (31.6%) patients had RRTFS. Patients who had RRTFS were younger (60 years vs. 65 years, p =0.006), and more likely to be white (73% vs. 67%, p=0.001). MLP had the highest AUROC, 0.694 (95% CI 0.586-0.791), followed by RF 0.523 (95% CI 0.469 – 0.583), logistic regression 0.516 (95% CI 0.457 – 0.575), logistic LASSO 0.512 (95% CI 0.456 – 0.571), and SVM 0.500 (95% CI0.500-0.500).
Conclusion
Compared to the standard logistic regression model, machine learning models, particularly the MLP method, had better performance at predicting RRTFS in patients started on CRRT using features prior to CRRT initiation.
Funding
- NIDDK Support