Kidney Week

Abstract: FR-PO0054

Machine Learning-Based Prediction of Hidden Acute Kidney Disease in Outpatient Settings

Session Information

• AKI: Epidemiology, Risk Factors, and Prevention
  November 07, 2025 | Location: Exhibit Hall, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Acute Kidney Injury

• 101 AKI: Epidemiology, Risk Factors, and Prevention

Authors

• Chiang, Hsiu-Yin, Big Data Center, China Medical University Hospital, Taichung, Taiwan

Hsiu-Yin Chiang, MS, PhD

• Lin, Zi-Han, Big Data Center, China Medical University Hospital, Taichung, Taiwan

Zi-Han Lin, MS

• Wu, Minyen, Big Data Center, China Medical University Hospital, Taichung, Taiwan

Minyen Wu, MS

• Kuo, Chin-Chi, Big Data Center, China Medical University Hospital, Taichung, Taiwan

Chin-Chi Kuo, MD, PhD, MPH

Background

Outpatient acute kidney disease (AKD_OPT) often goes undiagnosed due to fragmented kidney function data. In 2017, we launched the Acute Kidney Injury Detection System (AKIDS) to automatically integrate national and local electronic medical records to screen for AKD_OPT. However, up to 40% of outpatients lack prior kidney function data, limiting AKIDS’s detection capability. This study aims to develop a machine learning (ML) model to predict AKD_OPT and evaluate its clinical impact.

Methods

From the AKIDS-screened outpatient visits during 2017-2022, we kept the first visits and excluded patients aged <18 or >90 years and those with end-stage kidney disease (ESKD) or cancer. Patients with ≥2 serum creatinine (SCr) measurements in the 180 days before the index visit, sufficient to assess AKD_OPT (>50% SCr or >35% eGFR change), were used to develop the model (cohort 1; train/test/validate=6:2:2). Patients without sufficient SCr data formed cohort 2 for evaluating clinical impact. An XGBoost model was built using clinical features--excluding SCr and eGFR. Baseline SCr was imputed using the population-based median stratified by age, sex, and chronic kidney disease (CKD), developed from 3 million SCr records in CMUH. Model performance was assessed by AUC and B statistic. Logistic regression was used to estimate the 1-year risk of composite kidney outcome (CKO: ESKD, SCr doubling, or >40% eGFR decline) and all-cause mortality for true and predicted AKD_OPT.

Results

About 16% of 94,733 patients in cohort 1 had AKD_OPT. ML model achieved an AUC of 0.82 and B statistic of 0.71 in the validation set. Top predictors included comorbidities (diabetes, hypertension, hyperlipidemia), medications (diuretics, insulin, statin, nephrotoxic antibiotics), frequency of inpatient or emergency visit, labs (hemoglobin, red blood cell count), and prior ECG record (yes vs no). In cohort 1, predicted AKD_OPT and true AKD_OPT had similar 1-year risks of CKO (odds ratio [OR] 4.3 vs 5.0) and mortality (OR 5.8 vs 5.5). For 331,998 patients in cohort 2, predicted AKD_OPT consistently showed clinical significance, with associations to 1-year CKO (OR 3.3) and mortality (OR 6.4).

Conclusion

Our findings support the role of ML model in predicting AKD_OPT for patients without prior kidney function data, enabling proactive AKD_OPT risk assessment. Further replication is needed to validate generalizability.

Funding

• Government Support – Non-U.S.

Digital Object Identifier (DOI)

• doi: 10.1681/ASN.2025gwwtsref