Abstract: TH-PO1137
Prediction of Hyponatremia from Electronic Medical Records Using an Deep Learning Approach with an Artificial Neural Network Algorithm
Session Information
- Fluid, Electrolyte, Acid-Base Disorders
November 02, 2017 | Location: Hall H, Morial Convention Center
Abstract Time: 10:00 AM - 10:00 AM
Category: Fluid, Electrolytes, and Acid-Base
- 704 Fluid, Electrolyte, Acid-Base Disorders
Authors
- Jo, Young-Il, Konkuk University Medical Center, Seoul, Korea (the Republic of)
- Shin, Sug kyun, Ilsan Hospital NHIS, , GOYANG-SI, GYEONGGI-DO, Korea (the Republic of)
Background
Hyponatremia is associated with increased morbidity and mortality in both hospitalized and ambulatory patients. In the era of big data, analysis of electronic medical records (EMR) may have a significant impact on patient’s outcomes by identifying high-risk patients or supporting clinical decision making. The aim of this study was to predict hyponatremia from EMR using an deep learning approach with artificial neural network (ANN) algorithm.
Methods
A total 182,181 patients who measured serum sodium concentrations from 2010 to 2016 in a tertiary referral hospital were enrolled. Clinical, biochemical and medication data were obtained from EMR data warehouse. A total of 853 columns were presented along with basic patient information. For training, 500 random number trees were given to the random forest algorithm. The medication dataset was based on a neural network based on Dense Matrix. Learning and modeling have attempted to construct a combined dataset by using Ensemble model. The predictive value and diagnostic accuracy were calculated for the dataset based on serum sodium concentration less than 134 mEq/L.
Results
Using the confusion matrix and statistics for clinical and laboratory data set, the predicted value was 0.9104 (95% CI, 0.9040-0.9157). In the medication data set, he predicted value was 0.7515. Deep learning predictive approaches using the Ensemble model, the prediction probability for hyponatremia was 92.05% (Table).
Conclusion
Using an deep learning approach with the ensemble model, an artificial neural network can accurately predict hyponatremia from the laboratory dataset and the medication dataset of EMR.
Confusion matrix and statistics (Prediction/Real) for prediction of hyponatremia in the Ensemble model
0 | 1 | 2 | |
0 | 3199 | 64 | 69 |
1 | 170 | 3351 | 281 |
2 | 91 | 150 | 3014 |