ASN's Mission

ASN leads the fight to prevent, treat, and cure kidney diseases throughout the world by educating health professionals and scientists, advancing research and innovation, communicating new knowledge, and advocating for the highest quality care for patients.

learn more

Contact ASN

1401 H St, NW, Ste 900, Washington, DC 20005

email@asn-online.org

202-640-4660

The Latest on Twitter

Kidney Week

Abstract: PO0003

Early Prediction of Hospital-Acquired AKI from Electronic Health Records Using Machine Learning

Session Information

Category: Acute Kidney Injury

  • 101 AKI: Epidemiology, Risk Factors, and Prevention

Authors

  • Singh, Gurmukteshwar, Geisinger Medical Center, Danville, Pennsylvania, United States
  • Hu, Yirui, Geisinger Medical Center, Danville, Pennsylvania, United States
  • Riviello, David, Geisinger Medical Center, Danville, Pennsylvania, United States
  • Brown, Jason, Geisinger Medical Center, Danville, Pennsylvania, United States
  • Saunders, Sara, Geisinger Medical Center, Danville, Pennsylvania, United States
  • Garcia-Arce, Andres, Geisinger Medical Center, Danville, Pennsylvania, United States
  • Ho, Kevin, Sanofi Genzyme, Cambridge, Massachusetts, United States
  • Kirchner, H. Lester, Geisinger Medical Center, Danville, Pennsylvania, United States
Background

Hospital-acquired acute kidney injury (HA-AKI) leads to increased morbidity and mortality. Early prediction of HA-AKI using Electronic Health Records may enable clinicians to modify treatment to minimize risk and AKI severity

Methods

Inpatient admissions from 7/13/2012 – 7/11/2018 who had serum creatinine measured were included. Patients with end-stage renal disease, length of stay <48 hours and AKI at admission were excluded. A validated algorithm was used to determine baseline renal function. AKI was defined according to KDIGO guidelines. Machine learning algorithms were implemented to predict development of HA-AKI beyond the initial 24-hours of admission. 50 input variables to machine learning algorithms (random forest, XGBoost, logistic regression) included demographics, initial laboratory values taken within the first 24-hours of admission, active medications at time of admission, and prevalent comorbidities. Multiple imputation by chained equations (MICE) was used for missing variables. Univariate Feature Selection was utilized where variables were ranked by evaluating contribution to classification outcome. Randomized search strategy was performed to obtain the optimal hyperparameter set for each algorithm. Models were evaluated using a mean area under the receiver operating characteristic curve (AUC) over 5-fold cross-validation

Results

Among 209,300 inpatient admissions, 26,410 (12.6%) developed HA-AKI. For AKI prediction, the AUC of the full model was 0.88 for both random forest and XGBoost, and 0.86 for logistic regression. To balance the tradeoff between model simplicity and performance, 23 variables from univariate feature selection evaluated using random forest were selected in predicting HA-AKI (AUC = 0.87). The probability cut-off point of AKI prediction outcome was determined using Youden’s Index based on the balance between false positives and false negatives. A probability cutoff of > 0.23 provided sensitivity and specificity of 78% and 81%, respectively

Conclusion

Our machine learning algorithm applied at 24 hours of admission identifies patients at risk for HA-AKI with excellent accuracy. Significant variables included in this algorithm should be monitored in real-time to allow early identification and preventive interventions in patients at risk for HA-AKI

Funding

  • Clinical Revenue Support