Kidney Week

Abstract: TH-PO0715

A 30-Day Risk Prediction Model for Recovery, ESKD, or Death in Patients with Dialysis-Dependent AKI (AKI-D) Receiving Outpatient Hemodialysis

Session Information

• AKI: Prevention, Diagnostics, and Management
  October 22, 2026 | Location: Exhibit Hall A, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Acute Kidney Injury

• 101 AKI: Epidemiology, Risk Factors, and Prevention

Authors

• Nandorine Ban, Andrea, Renal Research Institute, New York, New York, United States

Andrea Nandorine Ban, MS, MSc

• Van Zandt, Carly R., Renal Research Institute, New York, New York, United States

Carly R. Van Zandt, MS

• Desai, Priya, Renal Research Institute, New York, New York, United States

Priya Desai, MPH

• Chaudhuri, Sheetal, Renal Research Institute, New York, New York, United States

Sheetal Chaudhuri, PhD

• Yi, Jun, Fresenius Medical Care Holdings Inc, Waltham, Massachusetts, United States

Jun Yi, PhD

• Neri, Luca, Renal Research Institute, New York, New York, United States

Luca Neri, MD, PhD

• Kotanko, Peter, Renal Research Institute, New York, New York, United States

Peter Kotanko, MD, FASN

• Kooman, Jeroen, Maastricht Universitair Medisch Centrum+, Maastricht, LI, Netherlands

Jeroen Kooman, MD, PhD

• Chatoth, Dinesh K., Fresenius Medical Care Holdings Inc, Waltham, Massachusetts, United States

Dinesh K. Chatoth, MD, FASN

• Usvyat, Len A., Renal Research Institute, New York, New York, United States

Len A. Usvyat, PhD, FASN

• Zhang, Hanjie, Renal Research Institute, New York, New York, United States

Hanjie Zhang, PhD

Background

Patients with acute kidney injury requiring dialysis (AKI-D) treated in outpatient dialysis units face competing risks of recovery, end-stage kidney disease (ESKD), or death. Existing models typically generate predictions at the time of AKI onset using baseline data. We developed a model to jointly estimate 30-day risks of these outcomes using recent clinical and machine data prior to the prediction.

Methods

We analyzed a U.S. cohort of in-center hemodialysis patients treated for AKI (Jan 2021–Jul 2023). Eligible patients were dialyzed via central venous catheter and monitored with the Crit-Line device. Patient- and treatment-level variables were derived from the 30 days prior to prediction, while ESA dose and laboratory variables were summarized over the prior 90 days, including most recent values. A multiclass XGBoost model was trained (80%) and evaluated on a held-out test set (20%) to predict 30-day outcomes (recovery, ESKD, death).

Results

Among 2,077 AKI-D patients, 568 (27.3%) recovered, 1,367 (65.8%) transitioned to ESKD, and 142 (6.8%) died. The test set included 321 patients with similar outcome distribution. Model performance achieved 0.71 accuracy, multiclass AUROC 0.76 (vs chance 0.50). Key predictors reflected multiple physiologic domains, including hemodynamic instability, fluid balance, dialysis treatment characteristics, and markers of residual kidney function (Figure 1). While creatinine was an important predictor of recovery and ESKD, the model also incorporated complementary signals across clinical domains.

Conclusion

We developed a multiclass model to estimate short-term outcomes in AKI-D patients. By integrating measures of kidney function, hemodynamics, and treatment characteristics, this approach enables clinically relevant risk stratification for competing outcomes and may support decision-making during outpatient dialysis care.

Acknowledgment

Artificial intelligence tools were used to assist in the writing of this abstract. The authors reviewed and edited the content and take full responsibility for the final version.

Figure 1. Top predictors of 30-day outcomes in AKI-D patients. Bars show mean absolute SHAP values (average contribution to prediction). (A) Recovery, (B) Transition to ESKD, (C) Death. ScvO₂ denotes central venous oxygen saturation.

Funding

• Commercial Support – Renal Research Institute, a wholly-owned subsidiary of Fresenius Medical Care.