Abstract: PO0077
Predicting Outcomes After AKI: Are You Better Than a Machine?
Session Information
- AKI Clinical, Outcomes, and Trials - 1
October 22, 2020 | Location: On-Demand
Abstract Time: 10:00 AM - 12:00 PM
Category: Acute Kidney Injury
- 102 AKI: Clinical, Outcomes, and Trials
Authors
- Arora, Tanima, Clinical and Translational Research Accelerator, Yale School of Medicine, New Haven, Connecticut, United States
- Biswas, Aditya, Clinical and Translational Research Accelerator, Yale School of Medicine, New Haven, Connecticut, United States
- Yamamoto, Yu, Clinical and Translational Research Accelerator, Yale School of Medicine, New Haven, Connecticut, United States
- Simonov, Michael, Clinical and Translational Research Accelerator, Yale School of Medicine, New Haven, Connecticut, United States
- Martin, Melissa, Clinical and Translational Research Accelerator, Yale School of Medicine, New Haven, Connecticut, United States
- Wilson, Francis Perry, Clinical and Translational Research Accelerator, Yale School of Medicine, New Haven, Connecticut, United States
Background
While multiple studies have used statistical models to predict outcomes after AKI, no studies have compared these models to physician intuition at the time of AKI consult. We studied the accuracy of physicians in predicting outcomes after AKI and compared it to the strength of predictive statistical models.
Methods
Our pilot study focuses on the prediction of 3 outcomes after AKI : Recovery,progression to dialysis and mortality. Postgraduate years 4 and 5 level Nephrology providers were asked, at the time of initial renal consult, to forecast outcomes at 3 timepoints : 24hr, 48hr and 7 days. We compared physician prognosis to a gradient boosted trees model trained using retrospective EHR data. Our primary measure of performance was area under the receiver operating characteristic curve (AUROC) at each time point.
Results
Data was captured from 56 patients with stage 2 AKI. Nephrology providers (n=7) were good to excellent at predicting dialysis at all three timepoints and death at 48 hours and 7 days. In contrast, their ability to predict recovery of AKI was relatively poor. The statistical model performed significantly better at predicting death at all timepoints, however was poorer at predicting dialysis (Figure 1.0).
Conclusion
Both physician clinical acumen and our statistical model showed good performance in predicting need for dialysis and death after AKI, however performed poorly when predicting recovery. This highlights the need to conduct further in-depth analysis into this area and implement strategies to enhance prediction of recovery after AKI.
Funding
- Other NIH Support