整合机器学习方法预测肢端肥大症患者对第一代受体配体治疗的抵抗
本研究基于2010年1月至2024年7月的回顾性单中心数据,纳入111例肢端肥大症患者(第一代生长抑素受体配体敏感64例、抵抗47例),评估10种机器学习算法并构建预测计算器。CatBoost模型表现最优,AUROC为0.896(95%CI 0.751-0.990),准确率82.4%。作者结论为基于CatBoost的计算器可有效预测治疗反应,但临床实施前需前瞻性验证。
要点
- CatBoost模型预测第一代生长抑素受体配体抵抗的AUROC为0.896(95%CI 0.751-0.990)。
- SHAP分析识别的关键预测因素包括治疗前GH、Knosp分级、治疗前IGF-1指数、T2加权MRI密度和合并症负担。
- 作者结论为计算器可有效预测治疗反应,但临床实施前需前瞻性验证。
核心结论
- 基于CatBoost的计算器可有效预测肢端肥大症患者对第一代生长抑素受体配体(fgSRLs)的治疗反应,但临床实施前需前瞻性验证。
研究设计与样本
- 回顾性研究,利用Research Patient Data Registry评估10种机器学习算法并构建预测计算器,研究时间为2010年1月至2024年7月。
- 单中心研究,在Mass General Brigham附属医院开展。
- 共111例肢端肥大症患者符合纳入标准,分为fgSRLs敏感组(n=64)和fgSRLs抵抗组(n=47)。
主要发现
- CatBoost模型表现最优,AUROC为0.896(95%置信区间:0.751-0.990),准确率82.4%,精确率86.7%,特异性88.2%,F1分数81.2%。
- SHAP分析识别的fgSRLs抵抗关键预测因素包括治疗前GH、Knosp分级、治疗前IGF-1指数、T2加权磁共振成像密度和合并症负担。
- 模型校准良好(Brier分数0.131),决策曲线分析显示具有临床效用。
- 已开发基于网页的计算器供临床使用。
临床怎么用
- 作者结论为基于CatBoost的计算器可有效预测肢端肥大症患者对fgSRLs的治疗反应,但临床实施前需前瞻性验证。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
CONTEXT: Acromegaly, caused by excess GH and IGF-1 due to pituitary adenomas, often necessitates first-generation somatostatin receptor ligands (fgSRLs) therapy when surgery fails. However, responses to fgSRLs therapy vary widely. OBJECTIVE: To develop a machine learning (ML)-based calculator that predicts individual responses to fgSRLs therapy, enabling evidence-based acromegaly management. DESIGN: A retrospective study (January 2010-July 2024) utilizing the Research Patient Data Registry to evaluate 10 ML algorithms and create a predictive calculator. SETTING: Single-center study conducted at Mass General Brigham-affiliated hospitals. PATIENTS: One hundred eleven acromegaly patients met inclusion criteria, classified as fgSRLs-responsive (n = 64) or fgSRLs-resistant (n = 47). INTERVENTIONS: IGF-1 trajectories were analyzed using linear mixed-effects modeling. Ten ML algorithms were assessed to predict fgSRLs resistance. SHapley Additive exPlanations (SHAP) analysis identified key predictors for the development of a web-based clinical calculator. MAIN OUTCOME MEASURES: Model performance was primarily evaluated using area under the receiver operating characteristic curve (AUROC), along with accuracy, precision, recall, specificity, F1 score, and decision curve analysis (DCA). RESULTS: The CatBoost model exhibited optimal performance based on AUROC 0.896 (95% confidence interval: 0.751-0.990), with accuracy 82.4%, precision 86.7%, specificity 88.2%, and F1 score 81.2%. Key predictors of fgSRLs resistance identified via SHAP analysis included pre-fgSRLs treatment GH, Knosp grade, pre-fgSRLs treatment IGF-1 index, T2-weighted magnetic resonance imaging density, and comorbidity burden. The model demonstrated excellent calibration (Brier score 0.131) and clinical utility via DCA. A web-based calculator was developed for clinical use. CONCLUSION: The CatBoost-based calculator effectively predicts fgSRLs treatment response in acromegaly patients. Prospective validation is required before clinical implementation.
引用
- Lin W, Shi S, Zheng Y, et al. Integrative Machine Learning Approach for Predicting Resistance to First-generation Receptor Ligands in Acromegaly. J Clin Endocrinol Metab. 2026. PMID: 40569033. doi:10.1210/clinem/dgaf375
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。