Abstract: FR-OR020
Development of Machine Learning Models for Predicting AKI Onset Using Electronic Medical Records
Session Information
- AKI: Epidemiology, Processes of Care, and Outcomes
November 08, 2019 | Location: 201, Walter E. Washington Convention Center
Abstract Time: 06:18 PM - 06:30 PM
Category: Acute Kidney Injury
- 101 AKI: Epidemiology, Risk Factors, and Prevention
Authors
- Uchino, Eiichiro, Kyoto University, Kyoto, Japan
- Kume, Kazuki, Fujitsu Ltd., Tokyo, Japan
- Iwamoto, Kazuki, Fujitsu Ltd., Tokyo, Japan
- Hayakawa, Tatsuo, Fujitsu Ltd., Tokyo, Japan
- Sato, Noriaki, Kyoto University, Kyoto, Japan
- Tamada, Yoshinori, Kyoto University, Kyoto, Japan
- Yanagita, Motoko, Kyoto University, Kyoto, Japan
- Okuno, Yasushi, Kyoto University, Kyoto, Japan
Background
Acute kidney injury (AKI) is a common disease associated with high morbidity and mortality, and the prediction of its onset will help the prevention and appropriate intervention. Recent studies have reported some machine learning models for predicting onset of AKI up to 72 hours in general patient population using electronic medical records (EMR), but studies for predicting in a relatively longer period such as one week have been limited.
Methods
We used the EMR data of adult patients who presented to Kyoto University Hospital, a tertiary teaching hospital in Japan, and received a measurement of renal function between January 2006 and November 2018. Based on the KDIGO guideline, the onset of stage 1 or severer AKI was determined by serum creatinine (sCr) values. Baseline sCr values were calculated by averaging within the windows according to KDIGO's 48-hour and 7-day AKI definitions. The comprehensive results of blood tests, medications, and vital signs were used as explanatory variables. By using random forest algorithm, seven models were constructed to classify whether or not to develop AKI during day 1 to day 7. The models were constructed and validated by 5-fold cross-validation in the cohort. The performance of the models was evaluated by area under the curve (AUC) of receiver operating characteristic curve.
Results
Of the 154,745 patients included in the analysis, it was determined that 10,460 patients (6.8%) had developed AKI. The amount of data with positive and negative labels was considered sufficient for training and validation of the seven models (positive labels, 2,528 ± 274; negative labels 8,571 ± 470 [mean ± standard deviation]). The AUC values were 0.910 ± 0.013, 0.885 ± 0.006, and 0.853 ± 0.012 in predicting onset of AKI after 1 day, 3 days, and 7 days, respectively.
Conclusion
Our models showed high performance equivalent to previous studies with AUC of more than 0.9 in prediction of onset after 1 day. In addition, the models achieved near performance even after periods of up to 7 days, which are longer in compared to previous studies. In future studies, implementing these predictive models in a clinical decision support system that presents risk scores may lead to appropriate interventions to prevent AKI.
Funding
- Commercial Support –