Kidney Week

Abstract: FR-PO0052

Artificial Intelligence-Enabled Prediction of Inpatient AKI: Analysis of 34 Million Hospitalizations Using the National Inpatient Sample, 2016-2020

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

• Sharma, Ashish, UConn Health, Farmington, Connecticut, United States

Ashish Sharma, MD

• Kumar, Harendra, Dow Medical College, Karachi, Sindh, Pakistan

Harendra Kumar, MBBS

• Tiwari, Angad, Maharani Laxmi Bai Medical College, Jhansi, UP, India

Angad Tiwari, MBBS

• Vishwanath, Manini, Nephrology Associates Inc, East Providence, Rhode Island, United States

Manini Vishwanath, MD

• Jailani, Mohamed, UConn Health, Farmington, Connecticut, United States

Mohamed Jailani, MD

• Jain, Nidhi, UConn Health, Farmington, Connecticut, United States

Nidhi Jain, MBBS

• Albeyoumi Mohammed, Hossam, UConn Health, Farmington, Connecticut, United States

Hossam Albeyoumi Mohammed, MD

Background

Acute kidney damage (AKI) is the leading cause of inpatient morbidity, death, and expense. Patient variability and data complexity makes early predictions difficult. Artificial intelligence (AI) offers scalable solutions, though real-world use with national datasets remains limited.

Methods

Adult hospitalizations from the National Inpatient Sample (2016–2020) were analyzed, with AKI defined via ICD-10 codes (N17.x). A Python-based ML pipeline using XGBoost and logistic regression was built. Predictors included demographics, comorbidities (Elixhauser), procedures, and hospital variables. Data were split 80/20 for training/testing. Performance was evaluated using AUC, Brier score, calibration plots, and SHAP values. Subgroup analyses involved CKD, sepsis, and major surgery cases.

Results

Among 4,192,847 (12.1%) of the 34,507,612 hospitalizations were for AKI. The XGBoost model surpassed logistic regression, with an AUC of 0.86 and a Brier score of 0.07. The key indicators were CKD, sepsis, congestive heart failure, contrast exposure, and mechanical ventilation. SHAP analysis showed consistent feature contributions across subgroups. The model retained efficacy in elderly (AUC 0.84) and surgical patients (AUC 0.81). External validation with 2020 data yielded AUC of 0.85.

Conclusion

AI models can accurately predict inpatient AKI using national dataset, enabling real-time risk stratification.

Receiver Operating Characteristic (ROC) Curve

Top Predictors of AKI in XGBoost Model

Digital Object Identifier (DOI)

• doi: 10.1681/ASN.2025w1qmkznx