Abstract: TH-PO665
An Explainable Model of ESA Prescription in the Hemodialysis Population
Session Information
- Anemia and Iron Metabolism
November 03, 2022 | Location: Exhibit Hall, Orange County Convention Center‚ West Building
Abstract Time: 10:00 AM - 12:00 PM
Category: Anemia and Iron Metabolism
- 200 Anemia and Iron Metabolism
Authors
- Chiu, Yi-Wen, Kaohsiung Medical University Chung Ho Memorial Hospital, Kaohsiung, Taiwan
- Lin, Ming-Yen, Kaohsiung Medical University Chung Ho Memorial Hospital, Kaohsiung, Taiwan
- Ku, Chan-Tung, National Sun Yat-sen University, Kaohsiung, Taiwan
- Yen, Hong-Ren, National Sun Yat-sen University, Kaohsiung, Taiwan
- Hsu, Chan, National Sun Yat-sen University, Kaohsiung, Taiwan
- Kang, Yihuang, National Sun Yat-sen University, Kaohsiung, Taiwan
Background
Our RCT (NCT04185519) has demonstrated the AI-assisted ESA prescription can maintain the Hb therapeutic target in HD patients, not inferior to physicians, and ESA dose-saving. However, the mechanism is untransparent.
Methods
A Generalized Linear Mixed-Effects Model (GLMM) was trained by ESA doses, Iron supplement, Hb levels, and demographic and biochemical data of HD patients between 2014 and 2020. Except for mean absolute error (MAE), the traditional ESA dose prescription algorithm (TEA) was compared for the model efficiency. An index named "Confidence” was calculated as the percentage of optimal ESA dose predicted by a model. The optimal ESA dose was defined as within one syringe dose different from the ones maintaining the Hb within 10.8-11.2 gm/dL.
Results
25,979 records of 316 ESKD patients were included, with 71.9% of Hb between 10-12mg/dL. The MAE were 0.4534 and 0.4964 gm/dl in training and testing data, respectively. Increasing the tree number and the depth of tree may further lower the MAE at the cost of model interpretability. The determinants of Hb changing by ESA dose included Hb, Hb changes, ESA dose, ESA dose changes, and iron supplement. There were 13 branches in our model to classify all records and demonstrate the ESA prescription algorithms. Compared with the TEA, our model showed similar efficiency in optimal ESA dose prediction. (Table)
Conclusion
A simple mixed-effect tree model with 13 leaf nodes can predict and explain the relationship between ESA dose and Hb changes. The "Confidence” used in our study helps evaluate the model. A further RCT (NCT05032651) is ongoing to confirm its clinical application.
| Records size | GLMM | GLMM | TEA | |
| N (%) | MAE, gm/dL | ESA Confidence,% | ESA Confidence, % | |
| node 1 | 1809 (7.0) | 0.49 | 86.2 | 95.7 |
| node 2 | 261 (1.0) | 0.46 | 94.6 | 96.6 |
| node 3 | 1347 (5.2) | 0.48 | 92.4 | 97.4 |
| node 4 | 217 (0.8) | 0.45 | 89.3 | 95.7 |
| node 5 | 789 (3.0) | 0.49 | 92.6 | 91.9 |
| node 6 | 3805 (14.7) | 0.45 | 78.8 | 89.1 |
| node 7 | 1128 (4.3) | 0.48 | 83.5 | 87.7 |
| node 8 | 2560 (9.9) | 0.43 | 85.9 | 86.9 |
| node 9 | 8224 (31.7) | 0.45 | 80.2 | 88.9 |
| node 10 | 2625 (10.1) | 0.45 | 86.5 | 73.5 |
| node 11 | 1059 (4.1) | 0.47 | 75.9 | 83.1 |
| node 12 | 1642 (6.3) | 0.51 | 96.6 | 100 |
| node 13 | 513 (2.0) | 0.50 | 100 | 99.4 |
Funding
- Clinical Revenue Support