Intradialytic Hypotension Frequency Prediction Using Generalizable Neighborhood Reasoning on Temporal Patient...

Area of Science:

• Nephrology
• Artificial Intelligence
• Data Science

Background:

• Intradialytic hypotension (IDH) is a frequent and serious complication in hemodialysis patients, impacting quality of life and increasing mortality risk.
• Accurate prediction of IDH allows for timely interventions, mitigating adverse events.
• Existing prediction models often focus on real-time data and may not fully capture the temporal dynamics of patient conditions.

Purpose of the Study:

• To develop a novel method for predicting monthly intradialytic hypotension (IDH) in hemodialysis patients.
• To leverage temporal knowledge graphs (KGs) for a more comprehensive representation of patient data over time.
• To enhance the generalizability of IDH prediction models by incorporating multi-month patient information.

Main Methods:

• Construction of monthly patient knowledge graphs (KGs) using data from 532 hemodialysis patients (Jan 2017 - Aug 2022).
• Creation of a temporal KG dataset by combining six sequential monthly KGs into observation windows.
• Application of neighborhood-based KG reasoning and a patient-centric graph convolution for representation learning, followed by sequential fusion and MLP prediction.

Main Results:

• The proposed temporal KG-based model significantly outperformed 7 classic machine learning models.
• Demonstrated superior performance in key metrics including accuracy and F1 score for predicting frequent IDH.
• Successfully utilized multi-month patient data within a temporal framework for improved prediction.

Conclusions:

• Temporal knowledge graphs provide a powerful framework for representing complex, multi-type patient data in hemodialysis.
• The developed patient-centric graph convolution and sequential fusion approach effectively captures temporal dependencies for IDH prediction.
• This approach offers a promising direction for improving the prediction and management of intradialytic hypotension.