Kidney Week

Abstract: FR-OR72

Prospective and External Evaluation of an Artificial Intelligence Model for Continuous and Early Prediction of Moderate and Severe AKI in Critically Ill Patients

Session Information

• Innovative AKI Strategies and Drug Discoveries
  October 25, 2024 | Location: Room 7, Convention Center
  Abstract Time: 05:50 PM - 06:00 PM

Category: Acute Kidney Injury

• 101 AKI: Epidemiology, Risk Factors, and Prevention

Authors

• Kashani, Kianoush, Mayo Clinic Minnesota, Rochester, Minnesota, United States

Kianoush Kashani, MD, MS, FASN

• Alfieri, Francesca, U-Care Medical s.r.l, Torino, Italy

Francesca Alfieri, MSc

• Ancona, Andrea, U-Care Medical s.r.l, Torino, Italy

Andrea Ancona, PhD

• Zappalà, Simone, U-Care Medical s.r.l, Torino, Italy

Simone Zappalà, PhD

• Votta, Carmine Domenico, Ospedale Maggiore di Lodi - ASST Lodi, Lodi, Italy

Carmine Domenico Votta, MD

• Maderna, Laura, Ospedale Maggiore di Lodi - ASST Lodi, Lodi, Italy

Laura Maderna, MD

• Gomez, Josep Sr., Hospital Universitario de Tarragona Joan XXIII, Tarragona, Spain

Josep Sr. Gomez, PhD

• Esteban, Federico, Hospital Universitario de Tarragona Joan XXIII, Tarragona, Spain

Federico Esteban, MD

• Gilavert, Mari Carmen, Hospital Universitario de Tarragona Joan XXIII, Tarragona, Spain

Mari Carmen Gilavert, MD

• Bodí, Maria, Hospital Universitario de Tarragona Joan XXIII, Tarragona, Spain

Maria Bodí, MD, PhD

• Finazzi, Stefano, Istituto di Ricerche Farmacologiche Mario Negri IRCCS, Bergamo, Italy

Stefano Finazzi, MSc, PhD

• Bacci, Alessandro, U-Care Medical s.r.l, Torino, Italy

Alessandro Bacci, MSc

• Cauda, Valentina, U-Care Medical s.r.l, Torino, Italy

Valentina Cauda, PhD

Background

Acute Kidney Injury (AKI) is associated with clinical and economic burdens on patients and healthcare systems. We previously proposed an AI-based model for the early and continuous prediction of moderate and severe ICU-acquired AKI episodes (ICU-A-AKI-2/3). In this study, the AI model was able to handle missing data and it was validated it in retrospective and prospective cohorts.

Methods

The predictive performance of the AI model was first internally and externally validated in retrospective cohorts in three countries (US, Netherlands, Italy; N =70107) from 176 different ICUs. Subsequently, the model was integrated and prospectively validated into the production in 2 European hospitals (Italy, Spain; N=225) from May 2023 to October 2023. The model analyses spot and temporal profiles of clinical variables from ICU patients, and its output is hourly risk-probability of developing AKI Stages 2 and 3, according to the KDIGO definition.
The model only requires routinely-available clinical variables.

Results

In the multicenter external retrospective cohorts, the predictive model obtained the area under the receiver operating characteristic curve (auROC) of >0.89 for the early detection of ICU-A-AKI-2/3. In the prospective validation, the model achieved auROCs ranging from 0.82 (CI 95% 0.73, 0.92) up to 0.87 (CI 95% 0.802, 0.943) in the different hospitals, with a lead time for AKI-2/3 onset varying from 14 to 18 hours after the first day of ICU hospitalization.

Conclusion

In this study, the AI model for early prediction of ICU-A-AKI-2/3 episodes was successfully validated in both rettrospetive and prospective cohorts, achieving high predictive performances.This successful validation marks a significant milestone in integrating it into clinical workflows and represents an essential advancement in bringing the AI model to ICU settings.

Funding

• Commercial Support – U-Care Medical s.r.l.