Kidney Week

Abstract: FR-PO0393

Dynamic Prediction of AKI Reversal in Critically Ill Patients Using a Transformer-Based Artificial Intelligence (AI) Model

Session Information

• AKI: Biomarkers, Diagnostics, and Risk Prediction
  October 23, 2026 | Location: Exhibit Hall A, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Acute Kidney Injury

• 102 AKI: Clinical, Outcomes, and Trials

Authors

• Oh, Wonsuk, Icahn School of Medicine at Mount Sinai, New York, New York, United States

Wonsuk Oh, PhD

• Shaikh, Ahmed, Icahn School of Medicine at Mount Sinai, New York, New York, United States

Ahmed Shaikh, MD

• Kohli-Seth, Roopa D., Icahn School of Medicine at Mount Sinai, New York, New York, United States

Roopa D. Kohli-Seth, MD

• Nadkarni, Girish N., Icahn School of Medicine at Mount Sinai, New York, New York, United States

Girish N. Nadkarni, MD, MPH

• Sakhuja, Ankit, Icahn School of Medicine at Mount Sinai, New York, New York, United States

Ankit Sakhuja, MBBS, MS, FASN

Background

Acute kidney injury (AKI) is a major complication of hospitalization, and early identification of those likely to recover may inform clinical decision-making. Current prediction models are static, which limits their usefulness for dynamic bedside application.

Methods

In this retrospective study, we developed a transformer based artificial intelligence model to predict AKI reversal using routinely collected clinical data from the Mount Sinai Health System (MSHS, 2022–2025), with external validation in MIMIC-IV (v3.1). AKI was defined by Kidney Disease: Improving Global Outcomes (KDIGO) criteria. The primary outcome, AKI reversal within 7 days of onset, was defined per the Acute Dialysis Quality Initiative (ADQI) 16 consensus as return of both serum creatinine and urine output to stage 0, without kidney replacement therapy, sustained for at least 48 hours. We evaluated model performance with time-dependent AUROC(t), AUPRC(t), and Brier score(t), benchmarking against logistic regression and XGBoost.

Results

The analysis included 34,625 AKI episodes from MSHS and 49,999 from MIMIC-IV. On the MSHS test set, the transformer achieved a mean AUROC(t) of 0.831 (95% CI 0.821–0.840), AUPRC(t) of 0.902 (0.894–0.910), and Brier(t) of 0.192 (0.189–0.195); external performance in MIMIC-IV was comparable, with AUROC(t) 0.827 (0.823–0.830), AUPRC(t) 0.922 (0.920–0.925), and Brier(t) 0.201 (0.200–0.202). Results remained consistent across age, sex and race subgroups.

Conclusion

A transformer-based AI model trained on routinely collected ICU data accurately predicts AKI reversal in critically ill patients and generalizes to an external cohort. Prospective validation is warranted to establish its utility for dynamic bedside risk assessment.

Acknowledgment

1. National Institutes of Health (NIH) grants K08DK131286 (AS)
2. AVI-SUMAN (AdVancIng Support and Mentorship for critical care AI research and Networking) Initiative – A collaboration between Windreich Dept of AI and Human Health, and Institute for Critical Care Medicine
3. Eric and Wendy Schmidt AI in Human Health Fellowship (WO)

Funding

• NIDDK Support