Abstract: SA-PO165
Identification and Performance Evaluation of AKI Trajectory Subtypes Associated with Mortality and Kidney Recovery in Critically Ill Patients
Session Information
- AKI: Epidemiology, Risk Factors, Prevention - III
November 09, 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
- Neyra, Javier A., University of Kentucky Medical Center, Lexington, Kentucky, United States
- Smith, Taylor D., University of Kentucky, Lexington, Kentucky, United States
- Ortiz-Soriano, Victor M., University of Kentucky, Lexington, Kentucky, United States
- Li, Xilong, University of Texas Southwestern Medical Center, Dallas, Texas, United States
- Xie, Donglu, University of Texas Southwestern Medical Center, Dallas, Texas, United States
- Adams-Huet, Beverley, University of Texas Southwestern Medical Center, Dallas, Texas, United States
- Moe, Orson W., University of Texas Southwestern Medical Center, Dallas, Texas, United States
- Toto, Robert D., University of Texas Southwestern Medical Center, Dallas, Texas, United States
- Chen, Jin, University of Kentucky, Lexington, Kentucky, United States
Background
Few risk-prediction models focus on outcomes specific to critically ill patients with AKI. We developed and evaluated the performance of a novel machine learning model called Trajectory of Acute Kidney Injury (TAKI) for the prediction of mortality and kidney recovery.
Methods
Independent cohorts from two academic institutions were used: UK (discovery, n=37,095) and UTSW (validation, n=10,590). Exclusion criteria consisted of age <18, eGFR <15 or ESKD, kidney transplant, absence of ≥2 serum creatinine (SCr) measures, absence of SCr-criteria of AKI in the first 7 ICU days or ICU stay <48 h. First, a trajectory based on KDIGO-AKI SCr-severity classification was composed for every patient using repeated SCr measures up to 7 days. Second, for trajectories with different length, population-based dynamic time-warping was developed for alignment. Third, the distance between any two aligned trajectories was computed and then adjusted using AKI severity. Fourth, hierarchical clustering was adopted with a dynamic merging process to determine the final trajectory subtypes, which were used as features for predicting hospital mortality and major adverse kidney events (MAKE) at 90 days following discharge (composite of death, RRT dependence or inability to recover 50% of baseline eGFR).
Results
The incidence of AKI was 33.4% (UK) and 27.0% (UTSW). Hospital mortality rates were 24.4% and 13.7% and MAKE rates 38.3% and 36.2% in UK and UTSW cohorts, respectively. TAKI identified improving, stationary and worsening AKI subtypes associated with outcomes beyond severity classification. TAKI improved prediction of mortality and MAKE when added to severity classification of AKI or multiorgan failure scores in both cohorts [Table].
Conclusion
TAKI is a feasible method of AKI subtyping that informs risk-stratification of mortality and kidney recovery in critically ill adults with AKI beyond current AKI severity classification. Further validation is needed.
Performance metrics (95%CI) of TAKI for the prediction of mortality and MAKE (UK cohort)
| SOFA Mortality | SOFA+TAKI Mortality | KDIGO Mortality | KDIGO+TAKI Mortality | KDIGO MAKE | KDIGO+TAKI MAKE | |
| AUC | 0.66 (0.65-0.68) | 0.76 (0.74-0.77)* | 0.63 (0.61-0.64) | 0.73 (0.71-0.73)* | 0.64 (0.62-0.65) | 0.77 (0.76-0.78)* |
*p<0.001 for performance comparison to corresponding reference; absolute IDI% (9.2-10.9) and continuous NRI% (46.5-61.4)