- AP AKI CRRT2026

Final Result

Oral Presentation

Upload E-Poster *

Format: jpg / png / jpeg / pdf

Abstract Details

Manuscript Type

Scientific Research Abstract

Abstract Category

Multidisciplinary

Please specify if you select 'Others' *

N/A

Author & Affiliation

Number of Co-Authors

Co-Author 1 *

Minh Thu Phung m142114003@tmu.edu.tw Taipei Medical University International Master Program in Medicine, College of Medicine Taipei Taiwan *

Co-Author 2 *

Tan Phat Huynh b11135035@mail.ntust.edu.tw National Taiwan University of Science and Technology International Advanced Technology Program Taipei Taiwan -

Co-Author 3 *

Dang Manh Ho dangmanh1811@gmail.com Taipei Medical University AIBiomed Research Lab Taipei Taiwan -

Co-Author 4 *

Nguyen Quoc Khanh Le khanhlee@tmu.edu.tw Taipei Medical University Professional Master Program in Artificial Intelligence in Medicine Taipei Taiwan -

Co-Author 5 *

Co-Author 6 *

Co-Author 7 *

Co-Author 8 *

Co-Author 9 *

Co-Author 10 *

Presenting Author

Presenting Author's First Name

Minh Thu

Presenting Author's Last Name

Phung

Presenting Author's Email Address

m142114003@tmu.edu.tw

Presenting Author's Country

Vietnam

Abstract Content

Abstract Title

Dual-Task Prediction of Acute Kidney Injury and Continuous Renal Replacement Therapy Initiation in the Intensive Care Unit: Multi-Center Development and External Validation

Introduction *

Acute kidney injury (AKI) and continuous renal replacement therapy (CRRT) initiation are clinically important renal endpoints in the intensive care unit (ICU). We developed and externally validated a dual-task machine learning framework to predict incident AKI within 24-48 hours and incident CRRT initiation within 24-96 hours after ICU admission.

Methods *

We used MIMIC-IV for model development/internal testing and the eICU Collaborative Research Database for external validation. Features were extracted from the first 24 hours and included vital signs, laboratory values, urine output, and support interventions. Race and admission type were excluded. AKI cohorts included 24,223 MIMIC-IV stays and 40,452 eICU stays; CRRT cohorts included 50,979 MIMIC-IV stays and 55,120 eICU stays. A calibrated logistic regression model was used for AKI prediction, and a calibrated XGBoost model was used for CRRT prediction.

Results *

For AKI prediction, the model achieved an AUROC of 0.7456 (95% CI 0.731-0.760) and an AUPRC of 0.2749 (95% CI 0.254-0.296) on the MIMIC-IV held-out test set. On eICU external validation, AUROC was 0.6972 (95% CI 0.688-0.707) and AUPRC was 0.2012 (95% CI 0.192-0.211). At a threshold of 0.101, eICU sensitivity was 68.49%, specificity 61.85%, positive predictive value (PPV) 16.05%, and negative predictive value 94.85%. For CRRT prediction, XGBoost achieved an AUROC of 0.9356 (95% CI 0.922-0.949), AUPRC of 0.2407 (95% CI 0.198-0.293), and Brier score of 0.00838 on MIMIC-IV held-out testing. Screening the top 5% highest-risk MIMIC-IV stays captured 71.48% of future CRRT starts, with PPV 13.97%. On eICU external validation, AUROC was 0.8832 (95% CI 0.866-0.900), AUPRC was 0.0706 (95% CI 0.054-0.093), and Brier score was 0.00737. Screening the top 5% highest-risk eICU stays captured 44.81% of future CRRT starts, with PPV 5.95% and 8.96-fold enrichment over baseline.

Conclusions *

Models trained on early ICU data predicted incident AKI and CRRT initiation with preserved external performance. Percentile-based risk stratification was especially useful for the low-prevalence CRRT endpoint and warrants prospective evaluation for ICU renal risk monitoring.

Keywords

Acute Kidney Injury, Continuous Renal Replacement Therapy, Intensive Care Unit, Machine Learning, Risk Stratification.

Figure

https://storage.unitedwebnetwork.com/files/1367/1256527-37821-Fig1_410048.jpeg

* JPG only