Aspiring Medical Student and current Critical Care RN. Enjoys everything outdoors but can also be found inside nerding out on her medical education artwork
The Pre-brief
58 y/o male with a history of OSA, COPD, CKD III, HTN, presents with RUQ abdominal pain x 4 days with associated nausea and vomiting. Hypotensive on admission, responsive to 30 cc/kg fluid bolus(110kg), not currently on vasopressors. Ultrasound findings were suggestive of volume intolerance (B-lines on lung ultrasound, and rising CVP estimated by IVC). Diagnosis of cholangitis is made based upon RUQ ultrasound, planned for ERCP. During morning rounds, objective measures were given for lactate, RR, MAP, but Urine output was described as “good urine.”
The Importance of Urine output.
KDIGO criterion for AKI includes Creatinine and Urine output1. As the KDIGO staging increases, mortality rises. Changes in urine output will precede changes in creatinine Oliguria is defined as .5 cc/kg/hr for > 6 hours. Though not overly complicated, the KDIGO urine output criterion may be enough of a cognitive load to promote the reductionist phrase “making good urine.” Could we simplify the assessment of urine output and maintain the ability to risk-stratify patients? How about admission 24-hour urine volume?
Let’s ask the Machine!
- Objectives
- Create a mortality prediction model using machine learning
- Assess and compare 24-hour admission output to other common ICU risk stratification values (RR, HR, Lactate, etc..).
- Methods
- I used the MIMIC database and 48629 patient encounters were extracted.
- Hospital Mortality was the primary outcome.
- I evaluated 29 common lab and vital values including Lactate, Heart Rate, and of course first 24-hour Urine output. Feature selection used to narrow down to the top 15 most important labs/vitals. Missing values imputed with Iterative computer. Data scaled with Standard Scaler (Figure 1).
- AUC was chosen as the primary metric.
- Finally, I used a Machine learning explainability method called SHAP, which ranks the features that were the most important for the prediction of mortality.
Results
The overall predictive ability of the Machine learning model (Xgboost) was 0.85, demonstrating a very good ability to predict mortality(Figures 2 ). Once I have determined that I have a strong model, I can now determine the most important features. The SHAP feature ranking demonstrates that UOP was the most important feature for the prediction of ICU mortality (Figure 3)!
Conclusions
Among all vital signs and labs, UOP was the single most important predictor of mortality. This quick analysis was retrospective, and machine learning feature selection is not without limitations, but it further corroborates the importance of urine output despite the simplified format. Don’t sleep on the Urine output!
References
Wang HE, Jain G, Glassock RJ, Warnock DG. Comparison of absolute serum creatinine changes versus Kidney Disease: Improving Global Outcomes consensus definitions for characterizing stages of acute kidney injury. Nephrology, dialysis, transplantation : official publication of the European Dialysis and Transplant Association – European Renal Association. 2013;28(6):1447-1454.
Johnson AEW, Pollard TJ, Shen L, et al. MIMIC-III, a freely accessible critical care database. Scientific Data. 2016;3(1):160035.
