Abstract：
OBJECTIVE: This study aimed to develop machine learning (ML) algorithms for the differential diagnosis of adrenocorticotropic hormone (ACTH)-dependent Cushing\'s syndrome (CS) based on biochemical and radiological features.
METHODS: Logistic regression algorithms were used for ML, and the area under the receiver operating characteristics curve (AUROC) was used to measure performance. We used Shapley Contributed Comments (SHAP) values, which help explain the results of the ML models to identify the meaning of each feature and facilitate interpretation.
RESULTS: A total of 106 patients, 80 with Cushing\'s disease (CD) and 26 with ectopic ACTH syndrome (EAS), were enrolled in the study. The ML task was created to classify patients with ACTH-dependent CS into CD and EAS. The average AUROC value obtained in the cross-validation of the logistic regression model created for the classification task was 0.850. The diagnostic accuracy of the algorithm was 86%. The SHAP values indicated that the most important determinants for the model were the 2-day 2-mg dexamethasone suppression test, the > 50% suppression in the 8-mg high-dose dexamethasone test, late-night salivary cortisol, and the diameter of the pituitary adenoma. We have also made our algorithm available to all clinicians via a user-friendly interface.
CONCLUSIONS: ML algorithms have the potential to serve as an alternative decision support tool to invasive procedures in the differential diagnosis of ACTH-dependent CS.
METHODS: Logistic regression algorithms were used for ML, and the area under the receiver operating characteristics curve (AUROC) was used to measure performance. We used Shapley Contributed Comments (SHAP) values, which help explain the results of the ML models to identify the meaning of each feature and facilitate interpretation.
RESULTS: A total of 106 patients, 80 with Cushing\'s disease (CD) and 26 with ectopic ACTH syndrome (EAS), were enrolled in the study. The ML task was created to classify patients with ACTH-dependent CS into CD and EAS. The average AUROC value obtained in the cross-validation of the logistic regression model created for the classification task was 0.850. The diagnostic accuracy of the algorithm was 86%. The SHAP values indicated that the most important determinants for the model were the 2-day 2-mg dexamethasone suppression test, the > 50% suppression in the 8-mg high-dose dexamethasone test, late-night salivary cortisol, and the diameter of the pituitary adenoma. We have also made our algorithm available to all clinicians via a user-friendly interface.
CONCLUSIONS: ML algorithms have the potential to serve as an alternative decision support tool to invasive procedures in the differential diagnosis of ACTH-dependent CS.
摘要：
目的：本研究旨在开发基于生化和放射学特征的促肾上腺皮质激素（ACTH）依赖性库欣综合征（CS）鉴别诊断的机器学习（ML）算法。
方法：逻辑回归算法用于ML，受试者工作特征曲线下面积(AUROC)用于测量性能。我们使用Shapley贡献评论(SHAP)值，这有助于解释ML模型的结果，以识别每个特征的含义并促进解释。
结果：共有106名患者，80例患有库欣病（CD），26例患有异位ACTH综合征（EAS），参加了这项研究。创建ML任务将ACTH依赖性CS患者分为CD和EAS。在为分类任务创建的逻辑回归模型的交叉验证中获得的平均AUROC值为0.850。该算法的诊断准确率为86%。SHAP值表明该模型最重要的决定因素是2天的2-mg地塞米松抑制试验，在8mg高剂量地塞米松试验中抑制>50%，深夜唾液皮质醇，和垂体腺瘤的直径。我们还通过用户友好的界面将我们的算法提供给所有临床医生。
结论：ML算法有可能作为ACTH依赖性CS鉴别诊断中侵入性程序的替代决策支持工具。
方法：逻辑回归算法用于ML，受试者工作特征曲线下面积(AUROC)用于测量性能。我们使用Shapley贡献评论(SHAP)值，这有助于解释ML模型的结果，以识别每个特征的含义并促进解释。
结果：共有106名患者，80例患有库欣病（CD），26例患有异位ACTH综合征（EAS），参加了这项研究。创建ML任务将ACTH依赖性CS患者分为CD和EAS。在为分类任务创建的逻辑回归模型的交叉验证中获得的平均AUROC值为0.850。该算法的诊断准确率为86%。SHAP值表明该模型最重要的决定因素是2天的2-mg地塞米松抑制试验，在8mg高剂量地塞米松试验中抑制>50%，深夜唾液皮质醇，和垂体腺瘤的直径。我们还通过用户友好的界面将我们的算法提供给所有临床医生。
结论：ML算法有可能作为ACTH依赖性CS鉴别诊断中侵入性程序的替代决策支持工具。
