Abstract: PO1453
Machine Learning Models for Risk Prediction of Adverse Events in Hyperkalemic Patients
Session Information
- Fluid, Electrolyte, and Acid-Base Disorders: Clinical - 1
October 22, 2020 | Location: On-Demand
Abstract Time: 10:00 AM - 12:00 PM
Category: Fluid, Electrolyte, and Acid-Base Disorders
- 902 Fluid, Electrolyte, and Acid-Base Disorders: Clinical
Authors
- Kanda, Eiichiro, Kawasaki Medical School, Kurashiki, Japan
- Kohsaka, Shun, Keio University, Tokyo, Japan
- Okami, Suguru, AstraZeneca K.K., Osaka, Japan
- Okada, Masafumi, IQVIA Solution Japan K.K., Tokyo, Japan
- Ma, Xiaojun, IQVIA Solution Japan K.K., Tokyo, Japan
- Yajima, Toshitaka, AstraZeneca K.K., Osaka, Japan
Background
Hyperkalemia is a common electrolyte abnormality in heart failure (HF) and chronic kidney disease (CKD) patients. Although increased risks of adverse events in hyperkalemic patients have been well reported, there is limited information on causality of adverse events. Considering multifactorial conditions of hyperkalemic patients, we aimed to develop predictive models using novel machine learning algorithms.
Methods
We utilized a Japanese hospital claims registry, Medical Data Vision. We extracted hyperkalemic patients with either CKD and/or HF aged ≥18 years, defined as patients with ≥2 serum potassium values ≥5.1 mmol/L; from April 2008 to September 2018. Extracted dataset was split into 80:20 for training and validation. The risk of adverse clinical events including all-cause death, hospitalization for cardiac events, hospitalization for HF, and renal replacement therapy (RRT) introduction over 3 years after hyperkalemic episodes was modeled using gradient boosted tree (XG), neural network (NN), and logistic regression (LR) based on 81 clinical variables collected in 12 months before hyperkalemic episodes.
Results
Of 74,974 hyperkalemic patients, 8,480 patients were included. Mean age was 75.6 years and 53.7% were male. The ROC curve and calibration analyses showed excellent performance for death (AUC=0.841 [XG], 0.815 [NN], 0.838 [LR]), hospitalization for cardiac events (AUC=0.782 [XG], 0.718 [NN], 0.743 [LR]), HF (AUC=0.875 [XG], 0.850 [NN], 0.855 [LR]), and RRT (AUC=0.958 [XG], 0.917 [NN], 0.946 [LR]) (Table). Clinical variables with high importance were identified (Figure).
Conclusion
The machine learning model successfully identified high-risk hyperkalemic patients for adverse events. Despite the need for model validation, these results support the use of predictive models to select high-risk hyperkalemic patients.
Calibration analyses for 3-year mortality
| Model | # of event | ROC-AUC | Specificity | Sensitivity | PPV | NPV |
| Gradient boosted tree | 1,114 | 0.841 | 0.901 | 0.516 | 0.504 | 0.906 |
| Neural network | 1,114 | 0.815 | 0.902 | 0.444 | 0.467 | 0.893 |
| Logistic regression | 1,114 | 0.838 | 0.966 | 0.291 | 0.625 | 0.876 |
Top 20 important clinical variables for predicting 3-year mortality
Funding
- Commercial Support –