Mulugeta Getahun, Zewotir Temesgen, Tegegne Awoke Seyoum, Muleta Mahteme Bekele, Juhar Leja Hamza
Department of Statistics, Bahir Dar University, Bahir Dar, Ethiopia.
School of Mathematics, Statistics & Computer Science, KwaZulu Natal University, Durban, South Africa.
BMC Med Inform Decis Mak. 2025 Feb 3;25(1):54. doi: 10.1186/s12911-025-02906-y.
Renal transplantation is a critical treatment for end-stage renal disease, but graft failure remains a significant concern. Accurate prediction of graft survival is crucial to identify high-risk patients. This study aimed to develop prognostic models for predicting renal graft survival and compare the performance of statistical and machine learning models.
The study utilized data from 278 renal transplant recipients at the Ethiopian National Kidney Transplantation Center between September 2015 and February 2022. To address the class imbalance of the data, SMOTE resampling was applied. Various models were evaluated, including Standard and penalized Cox models, Random Survival Forest, and Stochastic Gradient Boosting. Prognostic predictors were selected based on statistical significance and variable importance.
The median graft survival time was 33 months, and the mean hazard of graft failure was 0.0755. The 3-month, 1-year, and 3-year graft survival rates were found to be 0.979, 0.953, and 0.911, respectively. The Stochastic Gradient Boosting (SGB) model demonstrated the best discrimination and calibration performance, with a C-index of 0.943 and a Brier score of 0.000351. The Ridge-based Cox model closely followed the SGB model's prediction performance with better interpretability. The key prognostic predictors of graft survival included an episode of acute and chronic rejections, post-transplant urological complications, post-transplant nonadherence, blood urea nitrogen level, post-transplant regular exercise, and marital status.
The Stochastic Gradient Boosting model demonstrated the highest predictive performance, while the Ridge-Cox model offered better interpretability with a comparable performance. Clinicians should consider the trade-off between prediction accuracy and interpretability when selecting a model. Incorporating these findings into the clinical practice can improve risk stratification and personalized management strategies for kidney transplant recipients.
肾移植是终末期肾病的关键治疗方法，但移植肾失败仍是一个重大问题。准确预测移植肾存活情况对于识别高危患者至关重要。本研究旨在开发预测肾移植存活的预后模型，并比较统计模型和机器学习模型的性能。
该研究使用了2015年9月至2022年2月期间埃塞俄比亚国家肾脏移植中心278例肾移植受者的数据。为解决数据的类别不平衡问题，应用了SMOTE重采样。评估了各种模型，包括标准和惩罚Cox模型、随机生存森林模型和随机梯度提升模型。根据统计学意义和变量重要性选择预后预测因素。
移植肾存活时间的中位数为33个月，移植肾失败的平均风险为0.0755。发现3个月、1年和3年的移植肾存活率分别为0.979、0.953和0.911。随机梯度提升（SGB）模型表现出最佳的区分度和校准性能，C指数为0.943，Brier评分为0.000351。基于岭回归的Cox模型紧随SGB模型的预测性能，且具有更好的可解释性。移植肾存活的关键预后预测因素包括急性和慢性排斥反应发作、移植后泌尿系统并发症、移植后不依从、血尿素氮水平、移植后定期运动和婚姻状况。
随机梯度提升模型表现出最高的预测性能，而岭回归 - Cox模型在性能相当的情况下具有更好的可解释性。临床医生在选择模型时应考虑预测准确性和可解释性之间的权衡。将这些发现纳入临床实践可改善肾移植受者的风险分层和个性化管理策略。
