Abstract: SA-PO397
Prediction of Death in Dialysis Patients Using Artificial Intelligence
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
- Kim, Sungmin, Hallym University Sacred Heart Hospital, Anyang, Gyeonggi-do, Korea (the Republic of)
- An, Jung Nam, Hallym University Sacred Heart Hospital, Anyang, Gyeonggi-do, Korea (the Republic of)
- Lee, Dong Hee, Hallym University Sacred Heart Hospital, Anyang, Gyeonggi-do, Korea (the Republic of)
- Kim, Jwa-kyung, Hallym University Sacred Heart Hospital, Anyang, Gyeonggi-do, Korea (the Republic of)
- Kim, Sung Gyun, Hallym University Sacred Heart Hospital, Anyang, Gyeonggi-do, Korea (the Republic of)
Background
It is very important to predict the death and to detect and manage risk factors in dialysis patients. In this study, we aimed to build a prognostic prediction model that can predict the death of dialysis patients through a deep learning technique using dialysis-related clinical variables and vital signs during dialysis that change in real time in addition to traditional risk factors.
Methods
Data of patients who underwent maintenance dialysis at Hallym University Sacred Heart Hospital from January 2015 to December 2019 were extracted from electronic medical record. Changes in vital signs (before dialysis, during dialysis, immediately after dialysis), dry weight, weight gain between dialysis, ultrafiltration amount, blood test results, and in-hospital deaths were investigated. Out-of-hospital deaths were investigated using data from the National Statistical Office. Using refined data, a recurrent neural network-based long short-term memory deep learning model that can predict death from vital signs was trained.
Results
Of a total of 1,772 patients, 322 died in-hospital and 337 died out-of-hospital. Among these, patients with vital signs measured during dialysis within 72 hours were included. When learning with 4-fold cross validation for the prediction of death within 72 hours from each vital sign measurement time during dialysis (including both out-of-hospital and in-hospital death), the performance was AUROC 0.9591±0.0115 and AUPRC 0.2408±0.0290. Afterwards, the test set was composed of only 16 patients who died out of hospital within 100 hours from the time of the last vital sign measurement. As a result of checking the predictive performance, it was AUROC 0.8714 and AUPRC 0.1440.
Conclusion
In this study, a death prediction model using a deep learning technique that maximizes the correlation between data was constructed using vital signs and test results during dialysis, which are regularly measured in hemodialysis patients. A prospective study will be needed.