Abstract: SA-PO365
Development of a Deep Learning Model for Predicting Intradialytic Hypotension Using Multicenter Clinical Data Warehouse
Session Information
- Hemodialysis and Frequent Dialysis: CV and Risk Prediction
November 05, 2022 | Location: Exhibit Hall, Orange County Convention Center‚ West Building
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 701 Dialysis: Hemodialysis and Frequent Dialysis
Authors
- Lee, Hanbi, Seoul Saint Mary's Hospital, Seocho-gu, Seoul, Korea (the Republic of)
- Chung, Sungjin, Catholic University of Korea Yeouido Saint Mary's Hospital, Yeongdeungpo-gu, Seoul, Korea (the Republic of)
- Yang, Chul Woo, Seoul Saint Mary's Hospital, Seocho-gu, Seoul, Korea (the Republic of)
- Koh, Eun Sil, Catholic University of Korea Yeouido Saint Mary's Hospital, Yeongdeungpo-gu, Seoul, Korea (the Republic of)
- Chung, Byung ha, Seoul Saint Mary's Hospital, Seocho-gu, Seoul, Korea (the Republic of)
Background
Intradialytic hypotension (IDH) is a serious complication of hemodialysis, and is associated with subsequent vascular access thrombosis, inadequate dialysis dose, cardiovascular morbidity, and mortality. Since the mechanism of IDH is multifactorial, its prediction is a clinical challenge. This study aims to develop a deep learning model to predict the occurrence of IDH using data from multicenter clinical data warehouse.
Methods
Data from 2,008 patients who underwent a total of 928,070 hemodialysis sessions at seven university hospitals in South Korea between Mar 2009 and Dec 2019 was used in this study. IDH was defined according to the following criteria: (i) nadir systolic blood pressure (SBP)<100mmHg when the initial SBP≥160mmHg, (ii) nadir SBP<90mmHg when the initial 90≤SBP<160mmHg, or (iii) ≥20mmHg intradialytic SBP fall when initial SBP<90mmHg. Patients were randomly divided into training, validation and test sets. The importance of features in the occurrence of IDH was calculated from logistic regression, random forest, and XGBoost models. A deep 1-dimensional convolutional neural network model was constructed to predict IDH and the prediction performance was compared with other machine learning models.
Results
The machine learning classifiers demonstrated that the important common features associated with IDH were the occurrence of IDH and the mean SBP of last hemodialysis session. The deep convolutional neural network model with medical records of the last session predicted IDH with recall of 51.2%, F1 score of 44.6%, and negative predictive value of 97.2%, underperforming than other machine learning classifiers. However, combining the medical records of last 3 sessions boosted the prediction performance by recall to 60.6% and F1 score to 58.3% with negative predictive value of 97.2%, outperforming all other classifiers.
Conclusion
The past hemodialysis information enables stronger classification performances on predicting IDH occurrences. Our deep learning model would be a reliable screening tool of IDH and allow clinicians to adjust hemodialysis settings before hemodialysis treatment to prevent IDH.