Deep Learning Model for Predicting Intradialytic Hypotension Without Privacy Infringement: A Retrospective Two-Center...

Area of Science:

• Nephrology and Biomedical Engineering
• Artificial Intelligence in Healthcare
• Patient Safety in Dialysis

Background:

• Intradialytic hypotension (IDH) is a common complication during hemodialysis.
• Existing IDH prediction models often rely on clinical variables that raise privacy concerns.
• There is a need for IDH prediction methods that ensure patient data confidentiality.

Purpose of the Study:

• To develop and validate a novel Intradialytic hypotension (IDH) prediction model.
• To utilize minimal, non-identifiable data from hemodialysis machines to predict IDH events.
• To ensure the developed model mitigates privacy infringement risks associated with traditional prediction methods.

Main Methods:

• Analysis of 63,640 unidentifiable hemodialysis sessions from Korean hospital databases.
• Development and comparison of machine learning (logistic regression, XGBoost) and deep learning (convolutional neural networks) models.
• Prediction of IDH events within a 10-minute window using 30-minute pre-event data, with three IDH definitions: Nadir90, Fall20, and Fall20/MAP10.
• Evaluation of model performance using Area Under the Receiver Operating Characteristic Curves (AUROCs) and precision-recall curves.

Main Results:

• The deep learning model demonstrated superior performance in predicting IDH across all three definitions compared to other models.
• Achieved high AUROCs: 0.905 for Nadir90, 0.864 for Fall20, and 0.863 for Fall20/MAP10.
• The model's effectiveness was confirmed using only data directly measured by the hemodialysis machine.

Conclusions:

• A deep learning model effectively predicts intradialytic hypotension (IDH) using solely hemodialysis machine monitoring data.
• This privacy-preserving approach eliminates the need for personal patient information, reducing the risk of privacy breaches.
• The findings support the clinical utility of machine-derived data for real-time IDH prediction and patient safety.