Kidney Week

Abstract: FR-PO0013

Validation of Machine-Learning Predictions for Continuous Kidney Replacement Therapy Survival in Pediatric Patients

Session Information

• Artificial Intelligence and Digital Health at the Bedside
  November 07, 2025 | Location: Exhibit Hall, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Artificial Intelligence, Digital Health, and Data Science

• 300 Artificial Intelligence, Digital Health, and Data Science

Authors

• Feng, Jeffrey, University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States

Jeffrey Feng

• Tiu, Ryan Alyson-Yao, University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States

Ryan Alyson-Yao Tiu

• Srivastava, Rachana, University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States

Rachana Srivastava, MBBS

• Salusky, Isidro B., University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States

Isidro B. Salusky, MD, FASN

• Bui, Alex, University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States

Alex Bui, PhD

Background

While machine learning (ML) shows promise in predicting continuous kidney replacement therapy (CKRT) survival in adults (Zamanzadeh et al., 2024), pediatric prognostication remains undefined and limited due to sample size. To address this gap, we tested the adult-based model in a pediatric population to understand the potential for generalization to this unique population.

Methods

We extracted electronic health record data from 4,161 adult (age >=21) CKRT patients, 1,562 matched non-CKRT controls, and 206 pediatric (age <21) CKRT patients at UCLA from 2015-2021 (Fig. 1a). Longitudinal data up to 5 days preceding CKRT was used. A 60/20/20 split of the adult cohort was used to train, tune, and test ML models. The optimal gradient-boosting model was evaluated on the test set and pediatric cohort. Shapley (SHAP) values were used to interpret feature importance.

Results

The model achieved an area under the receiver operating characteristic curve of 0.92 (95% CI 0.90-0.94) and 0.81 (0.72-0.89), area under the precision-recall curve of 0.88 (0.84-0.91) and 0.86 (0.76-0.93), and Brier score of 0.11 (0.09-0.13) and 0.23 (0.17-0.30) on the adult and pediatric test sets, respectively (Fig. 1b). Higher O₂ saturation and hematopoeitic growth factor use favored survival, while leukocytosis and leukopenia predicted against survival (Fig. 1c). Performance declined with lower weight but stabilized above ~40kg, highlighting potential generalizability (Fig. 1d). When comparing features between adults and children above/below 40kg, we noted more differences for those <40kg, especially in important features such as white blood cell count (Fig. 1e).

Conclusion

The adult-based model shows potential in predicting pediatric CKRT survival, but performance drops with weight, reflecting age-specific disease etiologies. Incorporating pediatric-specific features for infants and younger children can improve generalizability.

Funding

• Other NIH Support

Digital Object Identifier (DOI)

• doi: 10.1681/ASN.2025sg5hek2x