Noh Dasom, Kwon Sunyoung, Cho Woo Hyun, Lee Jin Gu, Kim Song Yee, Park Samina, Jeon Kyeongman, Yeo Hye Ju
Department of Information Convergence Engineering, Pusan National University, Yangsan, South Korea.
School of Biomedical Convergence Engineering, Pusan National University, Yangsan, South Korea.
Clin Transplant. 2025 Aug;39(8):e70268. doi: 10.1111/ctr.70268.
In regions with limited donor availability, optimizing efficiency in lung transplant decision-making is crucial. Preoperative prediction of 1-year graft failure can enhance candidate selection and clinical decision-making.
We utilized data from the Korean Organ Transplantation Registry to develop and validate a deep learning-based model for predicting 1-year graft failure after lung transplantation. A total of 240 cases were analyzed using 5-fold cross-validation. Among 25 preoperative factors associated with 1-year graft failure, we selected the top 9 variables with coefficients ≥ 0.25 for model development.
Of the 240 lung transplant recipients, 55 (22.92%) developed graft failure within 1 year, while 185 survived. The final predictive model incorporated nine key pretransplant factors: age, bronchiolitis obliterans syndrome after hematopoietic cell transplantation, pretransplant bacteremia, bronchiectasis, creatinine, diabetes, positive human leukocyte antigen crossmatch, panel reactive antibody 1 peak mean fluorescence intensity, and pretransplant steroid use. The multilayer perceptron model demonstrated strong predictive performance, achieving an area under the curve of 0.780 and an accuracy of 0.733.
Our machine learning-based model effectively predicts 1-year graft failure in lung transplant recipients using a minimal set of pretransplant variables. Further validation is needed to confirm its clinical applicability.
在供体来源有限的地区，优化肺移植决策的效率至关重要。术前预测1年移植失败可改善候选者选择和临床决策。
我们利用韩国器官移植登记处的数据开发并验证了一种基于深度学习的模型，用于预测肺移植后1年的移植失败情况。使用5折交叉验证对总共240例病例进行了分析。在与1年移植失败相关的25个术前因素中，我们选择系数≥0.25的前9个变量用于模型开发。
在240例肺移植受者中，55例（22.92%）在1年内发生移植失败，185例存活。最终的预测模型纳入了9个关键的移植前因素：年龄、造血细胞移植后的闭塞性细支气管炎综合征、移植前菌血症、支气管扩张、肌酐、糖尿病、人类白细胞抗原交叉配型阳性、群体反应性抗体1峰值平均荧光强度以及移植前使用类固醇。多层感知器模型表现出强大的预测性能，曲线下面积为0.780，准确率为0.733。
我们基于机器学习的模型使用最少的移植前变量集有效地预测了肺移植受者1年的移植失败情况。需要进一步验证以确认其临床适用性。
