Abstract
Patients with β-thalassemia major (β-TM) face a wide range of complications as a result of excess iron in vital organs, including the heart and liver. Our aim was to find the best predictive machine learning (ML) model for assessing heart and liver iron overload in patients with β-TM. Data from 624 β-TM patients were entered into three ML models using random forest (RF), gradient boost model (GBM), and logistic regression (LR). The data were classified and analyzed by R software. Four evaluation metrics of predictive performance were measured: sensitivity, specificity, accuracy, and area under the curve (AUC), operating characteristic curve. For heart iron overload, the LR had the highest predictive performance based on AUC: 0.68 [95% CI (95% confidence interval): 0.60, 0.75]. The GBM also had the highest specificity (69.0%) and accuracy (67.0%). Most sensitivity is also acquired with LR (75.0%). For liver iron overload, the highest performance based on AUC was observed with RF, AUC: 0.68 (95% CI: 0.59, 0.76). The RF showed the highest accuracy (66.0%) and specificity (66.0%), while the LR had the highest sensitivity (84.0%). Ferritin, duration of transfusion, and age were determined as the most effective predictors of iron overload in both heart and liver. Logistic regression LR was determined to be the strongest method to predict cardiac and RF values for liver iron overload in patients with β-TM. Older thalassemia patients with a high serum ferritin (SF) level and a longer duration of transfusion therapy were more prone to heart and liver iron overload.
Acknowledgements
The authors thank Shiraz University of Medical Sciences for their approval and support. The present address of S. Haghpanah is Professor of Community Medicine, Nemazee Hospital, Zand Street, Shiraz, Iran. Authors’ contributions: S. Haghpanah and N. Asmarian had the idea for article; S. Haghpanah and M. Hosseini-Bensenjan performed the literature search; N. Asmarian performed the data analysis; M. Hosseini-Bensenjan S. Haghpanah prepared the original draft; S. Haghpanah, A. Kamalipour, M. Hosseini-Bensenjan and M. Karimi wrote, reviewed and edited the draft; A. Kamalipour critically proof read the draft, supervised by S. Haghpanah and N. Asmarian. All authors have read and agreed to the published version of the manuscript. All authors have contributed substantially to the reported study.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Data availability statement
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.