Development of Artificial Intelligence for Clinical Decision Support to Optimizes Erythropoietin-Stimulating Agents in Patients with CKD

Background: The application of AI in predicting Hb levels and determining appropriate ESA doses for managing anemia of CKD is a promising development. Previous studies using AI for anemia of CKD treatment have demonstrated positive outcomes, such as increased percentages of patients achieving target Hb levels and reduced ESA dosages. The replication of these findings in the Thai population could have significant implications for managing anemia of CKD in Thailand. Methods: The development of the AI model included 4 steps: data collection, feature selection, data preparation, and model selection. CKD patients with eGFR <45 ml/min/1.73m2 who used ESA were included. The data was cleaned and prepared by removing outlier data and correcting missing values. Variables were used to create time series. Time series were divided into training and test set to create the best models (Decision Tree, Random Forest, XGBoost, and SVM). Finally, the models were evaluated by Pearson r correlation and RMSE. (Fig 1) Results: A total of 9,935 time series were collected . The median follow-up time was 9 months. The models did not perform well in predicting weekly change of Hb level on the entire dataset (RMSE for Decision Tree was 0.141, for Random Forest was 0.135, for XGBoost was 0.137, and for SVM was 0.135), and the difference in weekly change of Hb ranged from -0.2 to 0.2 g/dL, approximately 0-40%. The best models are Random Forest, XGBoost and SVM with similar results. (Fig 2) Conclusion: This study showed that AI model can predict weekly change in Hb level. Although the performance may not be able to accurately predict weekly changes in Hb levels. Future studies should aim to include a larger number of participants to improve the performance of the models. Funding: Government Support – Non-U.S.Trial Flow ChartThe Difference of weekly Hb change of the model