Kidney Week

Abstract: TH-OR025

CRRT Machine Pressure Waveforms Improve Day-Wise Intensive Care Unit (ICU) Mortality Prediction in Patients with AKI

Session Information

• Artificial Intelligence in Kidney Care: From Pathology to Prediction
  October 22, 2026 | Location: Room 711, Convention Center
  Abstract Time: 05:30 PM - 05:40 PM

Category: Artificial Intelligence, Digital Health, and Data Science

• 300 Artificial Intelligence, Digital Health, and Data Science

Authors

• Pranto, Shehan Irteza, The University of Alabama at Birmingham, Birmingham, Alabama, United States

Shehan Irteza Pranto, MS

• Lambert, Joshua, University of Cincinnati, Cincinnati, Ohio, United States

Joshua Lambert, PhD

• Yang, Joanna, Icahn School of Medicine at Mount Sinai, New York, New York, United States

Joanna Yang, MS

• Kauffman, Justin, Icahn School of Medicine at Mount Sinai, New York, New York, United States

Justin Kauffman, MS

• Goldstein, Stuart, University of Cincinnati, Cincinnati, Ohio, United States

Stuart Goldstein, MD, FASN

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

Girish N. Nadkarni, MD, MPH

• Chan, Lili, Icahn School of Medicine at Mount Sinai, New York, New York, United States

Lili Chan, MD, MS

• Neyra, Javier A., The University of Alabama at Birmingham, Birmingham, Alabama, United States

Javier A. Neyra, MD, MS, FASN

• Chen, Jin, The University of Alabama at Birmingham, Birmingham, Alabama, United States

Jin Chen, PhD

Background

CRRT devices generate minute-level pressure waveforms (Access, Filter, Effluent, Return; 1,440 timesteps/patient/day) that reflect circuit function and troubleshooting but have not been utilized for clinical prognostication. We evaluated whether features derived from these pressure streams add prognostic value to clinical variables for rolling day-wise ICU mortality prediction.

Methods

In the CRRTnet cohort (976 patients; 5,193 treatment-days), we performed rolling day-wise ICU mortality prediction using a 1-day observation window, 1-day gap, and next-day target. Three data configurations (clinical-only, machine-only, combined) were compared under per-day and aggregated feature strategies, where aggregated features summarized all prior CRRT days. Features were selected by grouped mRMR (Figure, A); eight classifiers were tuned by 5-fold stratified search optimized for AUROC. Patients were split into training, validation, and held-out test sets (70:15:15) without patient-level overlap. SHAP quantified and explained feature contributions.

Results

Pressure features alone predicted next-day ICU mortality at AUROC 0.686–0.687, while clinical data had at AUROC of 0.730–0.776. Combined pressure and clinical features with aggregated summaries and Gradient Boosting (GB) achieved the best performance (AUROC 0.801 F1 0.759; Figure, B). Aggregated features outperformed per-day across all pipelines. In SHAP rankings, CRRT pressure features occupied 8 of the top 15 predictors (53%; Figure, C), led by minimum filter pressure, a known bedside marker of filter clotting and clogging, with all four pressure streams represented.

Conclusion

CRRT pressure waveforms carry independent prognostic signals for day-wise ICU mortality. Coupling routinely generated machine data with clinical variables yields improved performance of a daily prognostic assessment during CRRT. External validation and prospective calibration are next steps.

Acknowledgment

NIH-NIDDK R01DK133539
NIH-NIDDK U54DK137307

Feature selection, ICU mortality prediction performance, and top feature importances across six CRRT pipelines.

Funding

• NIDDK Support – Baxter/Vantive