Abstract: TH-PO062
Prediction of AKI in ICU Using Routinely Collected Data by Machine-Learning Algorithms and Its Visualization
Session Information
- AKI: Epidemiology, Risk Factors, Prevention - I
November 07, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Acute Kidney Injury
- 101 AKI: Epidemiology, Risk Factors, and Prevention
Authors
- Sato, Noriaki, Department of Nephrology, Kyoto University Graduate School of Medicine, Kyoto, Japan
- Uchino, Eiichiro, Department of Nephrology, Kyoto University Graduate School of Medicine, Kyoto, Japan
- Kojima, Ryosuke, Kyoto University Graduate School of Medicine, Kyoto, Japan
- Hiragi, Shusuke, Kyoto University Hospital, Sakyo-ku, Kyoto, Japan
- Yanagita, Motoko, Department of Nephrology, Kyoto University Graduate School of Medicine, Kyoto, Japan
- Okuno, Yasushi, Kyoto University Graduate School of Medicine, Kyoto, Japan
Background
Acute kidney injury (AKI) is associated with high mortality, and occur frequently particularly in the intensive care unit (ICU), due to various conditions such as septic shock. It is clinically important to identify high-risk patients with AKI in advance and perform appropriate interventions. There have been many studies that predicted the onset of AKI using data like medication or laboratory data so far, but few studies have utilized time series vital data to predict AKI, and have assessed which time point is important for the prediction in each patient. In this study, we attempt to predict the onset of AKI by applying machine learning algorithms to vital data acquired routinely in most ICU patients from the publicly available dataset, followed by the visualization of the rationale behind the prediction.
Methods
We used publicly disclosed dataset named Medical Information Mart for Intensive Care three. AKI was defined based on KDIGO serum creatinine criteria. We included patients with stage 2 or higher AKI within 48 hours after admission to ICU. The onset of AKI was defined as when the high creatinine level was first detected. We included data of systolic, diastolic, and mean blood pressure, heart rate, respiratory rate, body temperature, and SpO2. In addition, age and the baseline serum creatinine level were collected. We utilized machine learning algorithm of XGBoost and 1-dimensional convolutional neural networks (1D-CNN). We predicted the onset of AKI in 24 to 48 hours using data collected in 0 to 24 hours after admission to ICU. We evaluated the performance of the model using area under receiver operating characteristic curve (AUROC), with five-fold cross-validation. We applied Grad-CAM to visualize the rationale behind the prediction in 1D-CNN.
Results
The highest performance was obtained by XGBoost when minimum, maximum and mean value of vital data and age, serum creatinine data was combined (AUROC: 0.793±0.34). Using only vital data, the performance was 0.666±0.046, 0.675±0.053 for XGBoost, 1D-CNN respectively. Using Grad-CAM to the 1D-CNN model, some visualization results suggested that the time period of lower blood pressure might contribute to the prediction.
Conclusion
It was suggested that XGBoost and 1D-CNN could predict the onset of AKI and visualize its rationale behind the prediction.
Funding
- Government Support - Non-U.S.