Division of Cardiology, Department of Internal Medicine, Chi-Mei Medical Center, Zhonghua Road, Yongkang District, 901, Tainan, Taiwan, ROC.
Department of Biotechnology, Southern Taiwan University of Science and Technology, Tainan, Taiwan, ROC.
Arch Toxicol. 2022 Oct;96(10):2731-2737. doi: 10.1007/s00204-022-03341-y. Epub 2022 Jul 25.
Although anti-cancer therapy-induced cardiotoxicity is known, until now it lacks a reliable risk predictive model of the subsequent cardiotoxicity in breast cancer patients receiving anthracycline therapy. An artificial intelligence (AI) with a machine learning approach has yet to be applied in cardio-oncology. Herein, we aimed to establish a predictive model for differentiating patients at a high risk of developing cardiotoxicity, including cancer therapy-related cardiac dysfunction (CTRCD) and symptomatic heart failure with reduced ejection fraction. This prospective single-center study enrolled patients with newly diagnosed breast cancer who were preparing for anthracycline therapy from 2014 to 2018. We randomized the patients into a 70%/30% split group for ML model training and testing. We used 15 variables, including clinical, chemotherapy, and echocardiographic parameters, to construct a random forest model to predict CTRCD and heart failure with a reduced ejection fraction (HFrEF) during the 3-year follow-up period (median, 30 months). Comparisons of the predictive accuracies among the random forest, logistic regression, support-vector clustering (SVC), LightGBM, K-nearest neighbor (KNN), and multilayer perceptron (MLP) models were also performed. Notably, predicting CTRCD using the MLP model showed the best accuracy compared with the logistic regression, random forest, SVC, LightGBM, and KNN models. The areas under the curves (AUC) of MLP achieved 0.66 with the sensitivity and specificity as 0.86 and 0.53, respectively. Notably, among the features, the use of trastuzumab, hypertension, and anthracycline dose were the major determinants for the development of CTRCD in the logistic regression. Similarly, MLP, logistic regression, and SVM also showed higher AUCs for predicting the development of HFrEF. We also validated the AI prediction model with an additional set of patients developing HFrEF, and MLP presented an AUC of 0.81. Collectively, an AI prediction model is promising for facilitating physicians to predict CTRCD and HFrEF in breast cancer patients receiving anthracycline therapy. Further studies are warranted to evaluate its impact in clinical practice.
尽管已知抗癌治疗会导致心脏毒性，但目前缺乏一种可靠的预测模型来预测接受蒽环类药物治疗的乳腺癌患者随后发生心脏毒性的风险。人工智能（AI）和机器学习方法尚未应用于心脏肿瘤学领域。在此，我们旨在建立一个预测模型，以区分发生心脏毒性（包括癌症治疗相关心功能障碍和射血分数降低的心力衰竭）风险较高的患者。这项前瞻性单中心研究纳入了 2014 年至 2018 年期间准备接受蒽环类药物治疗的新诊断乳腺癌患者。我们将患者随机分为 70%/30%的比例，分别用于 ML 模型的训练和测试。我们使用了 15 个变量，包括临床、化疗和超声心动图参数，构建了一个随机森林模型，以预测 3 年随访期间（中位数为 30 个月）的 CTRCD 和射血分数降低的心力衰竭（HFrEF）。还比较了随机森林、逻辑回归、支持向量聚类（SVC）、LightGBM、K-最近邻（KNN）和多层感知机（MLP）模型的预测准确性。值得注意的是，与逻辑回归、随机森林、SVC、LightGBM 和 KNN 模型相比，使用 MLP 模型预测 CTRCD 的准确性最高。MLP 模型的曲线下面积（AUC）为 0.66，灵敏度和特异性分别为 0.86 和 0.53。值得注意的是，在这些特征中，曲妥珠单抗的使用、高血压和蒽环类药物剂量是逻辑回归中发生 CTRCD 的主要决定因素。同样，MLP、逻辑回归和 SVM 也显示出更高的 AUC 来预测 HFrEF 的发生。我们还使用另一组发生 HFrEF 的患者对 AI 预测模型进行了验证，MLP 呈现的 AUC 为 0.81。总之，人工智能预测模型有望帮助医生预测接受蒽环类药物治疗的乳腺癌患者发生 CTRCD 和 HFrEF。需要进一步的研究来评估其在临床实践中的影响。
