Abstract: TH-PO260
Real-Time Dual Prediction of Intradialytic Hypotension and Hypertension Using an Explainable Deep Learning Model
Session Information
- Hemodialysis: Volume, Metabolic Complications, Clinical Outcomes
November 02, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 801 Dialysis: Hemodialysis and Frequent Dialysis
Authors
- Yun, Donghwan, Seoul National University Hospital, Jongno-gu, Seoul, Korea (the Republic of)
- Kim, Dong Ki, Seoul National University Hospital, Jongno-gu, Seoul, Korea (the Republic of)
- Oh, Kook-Hwan, Seoul National University Hospital, Jongno-gu, Seoul, Korea (the Republic of)
- Han, Seung Seok, Seoul National University Hospital, Jongno-gu, Seoul, Korea (the Republic of)
Background
Intradialytic hypotension (IDH) and hypertension (IDHTN) are associated with poor outcomes in hemodialysis patients. However, there is currently no real-time predictive model for dual outcomes. This study aims to develop an explainable deep learning model using a sequence-to-sequence-based attention network to simultaneously predict IDH and IDHTN.
Methods
Electronic health records of 11,110 hemodialysis patients were utilized, comprising 302,774 sessions. The data was divided into training (70%), validation (10%), and test (20%) sets using randomization. IDH-1 was defined as nadir systolic blood pressure (BP) <90 mmHg, IDH-2 as a decrease in systolic BP ≥20 mmHg and/or a decrease in mean arterial pressure ≥10 mmHg, and IDHTN as an increase in systolic BP ≥10 mmHg within 1 hour. The temporal fusion transformer (TFT)-based model was developed and compared with other machine learning models, including recurrent neural network, light gradient boosting machine, random forest, and logistic regression, in terms of model performance measured by receiver operating characteristic curve (AUROC) and area under the precision-recall curves (AUPRC).
Results
The TFT-based model outperformed other models with AUROCs of 0.953 (0.952–0.954), 0.892 (0.891–0.893), and 0.889 (0.888–0.890) for predicting IDH-1, IDH-2, and IDHTN, respectively. The AUPRCs of the TFT-based model for the outcomes were higher compared to other models. Key predictors included age, previous session, systolic BP, and elapsed time.
Conclusion
The developed TFT-based model enables real-time prediction of both IDH and IDHTN while providing explainable variable importance.
Mean weights of time-invariant and time-varying features from the attention module in the model. (A), Weights of time-invariant features. (B), Weights of time-varying features. BP, blood pressure.