Abstract: TH-PO016
Forecasting Intradialytic Hypotension: A Comparative Analysis of Machine-Learning and Deep-Learning Approaches
Session Information
- Augmented Intelligence for Prediction and Image Analysis
October 24, 2024 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Author
- Huang, Chun-Te, Taichung Veterans General Hospital, Taichung, Taiwan
Background
Intradialytic hypotension (IDH) poses significant risks to patients undergoing dialysis, necessitating reliable predictive tools for early intervention. This study evaluates the efficacy of machine learning and deep learning models in predicting IDH.
Methods
We developed predictive models on a Taichung Veterans General Hospital dataset using XGBoost (eXtreme Gradient Boosting) and RNN (Recurrent Neural Network) algorithms. The dataset encompasses the period from 2015 to 2020, consisting of 211,535 timestamps from 20,946 hemodialysis sessions across 2,118 patients. We defined intradialytic hypotension as a decrease in systolic blood pressure ≥20 mm Hg and/or a mean arterial pressure decrease ≥10 mm Hg. The dataset was split into 80% for training, employing 5-fold cross-validation, and 20% for testing. Features selected for training included age, sex, comorbidity, vital signs, laboratory data, medication, and specific hemodialysis characteristics.
Results
The XGBoost model exhibited superior performance in predicting intradialytic hypotension, achieving an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.9790 and an Area Under the Precision-Recall Curve (AUPRC) of 0.9825. While still effective, the RNN model showed slightly lower performance with an AUROC of 0.976 and an AUPRC of 0.978. These results emphasize the slight edge of XGBoost in predictive accuracy compared to the RNN model, demonstrating its enhanced capability in clinical forecasting without specifically relying on demographic variables like age and weight.
Conclusion
The findings from this study highlight the slightly superior performance of the XGBoost model over the RNN in predicting intradialytic hypotension. XGBoost demonstrated higher predictive accuracy and offers potential advantages in clinical implementation due to its less intensive computational demands. This makes it an especially viable option for real-time clinical applications with limited computational resources.
The model performance of XGBoost