基于机器学习的决策模型在胃癌手术切除前辅助临床医生决定淋巴结清扫范围(D1与D2)的性能
本研究回顾性分析557例胃癌患者的临床影像特征,用逻辑回归、支持向量机和自编码器算法构建决策模型,在371例训练集和186例测试集中预测D1与D2淋巴结清扫范围。自编码器模型诊断性能最高(测试集AUC 0.946),使过度治疗从全部治疗模式的21.7%降至0.7-0.9%。
要点
- 自编码器决策模型在测试集中区分D1与D2的AUC为0.946(95%CI 0.925-0.978)。
- 新方法使过度治疗从全部治疗模式的21.7%(121/557)降至0.7-0.9%(4-5/557)。
- 日本第4版胃癌治疗指南标准使过度治疗降至15.1%(84/557),新方法优于该标准。
核心结论
- 基于机器学习的决策模型在识别适合D1与D2切除的患者方面表现出高准确性。
- 与全部治疗模式和日本第4版指南标准相比,新方法使过度治疗分别改善约14-20%,可能增加可安全避免不必要淋巴结清扫的局部晚期胃癌患者数量。
研究设计与样本
- 回顾性研究,纳入557例胃癌患者,由对全部组织病理学信息设盲的专家小组解读常规临床可获得的临床影像特征。
- 全部患者接受标准D2切除术,以术后病理分期为参考标准区分D1(pT1+N0)与D2(≥pT1+≥pN1)淋巴结清扫。
- 采用逻辑回归、支持向量机和自编码器算法在371例训练数据和186例测试数据中分别开发决策模型。
诊断标准与关键数值
- 自编码器模型区分D1与D2的AUC:训练集0.965(95%CI 0.948-0.978),测试集0.946(95%CI 0.925-0.978)。
- 支持向量机模型AUC:训练集0.925(95%CI 0.902-0.944),测试集0.942(95%CI 0.922-0.973)。
- 逻辑回归模型AUC:训练集0.886(95%CI 0.858-0.910),测试集0.891(95%CI 0.891-0.952)。
- 过度治疗率:全部治疗模式21.7%(121/557),日本第4版指南标准15.1%(84/557),新方法0.7-0.9%(4-5/557)。
临床应用与阅片要点
- 作者认为,机器学习决策模型可高精度识别适合D1与D2切除的患者。
- 作者指出,与全部治疗模式和日本第4版指南标准相比,新方法使过度治疗改善约14-20%,可能增加可安全避免不必要淋巴结清扫的局部晚期胃癌患者数量。
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)。
英文原文(摘要)· 点击展开
BACKGROUND: Controversy still exists on the optimal surgical resection for potentially curable gastric cancer (GC). Use of radiologic evaluation and machine learning algorithms might predict extent of lymphadenectomy to limit unnecessary surgical treatment. We purposed to design a machine learning-based clinical decision-support model for predicting extent of lymphadenectomy (D1 vs. D2) in local advanced GC.
METHODS: Clinicoradiologic features available from routine clinical assignments in 557 patients with GCs were retrospectively interpreted by an expert panel blinded to all histopathologic information. All patients underwent surgery using standard D2 resection. Decision models were developed with a logistic regression (LR), support vector machine (SVM) and auto-encoder (AE) algorithm in 371 training and tested in 186 test data, respectively. The primary end point was to measure diagnostic performance of decision model and a Japanese gastric cancer treatment guideline version 4th (JPN 4th) criteria for discriminate D1 (pT1 + pN0) versus D2 (≥ pT1 + ≥ pN1) lymphadenectomy.
RESULTS: The decision model with AE analysis produced highest area under ROC curve (train: 0.965, 95% confidence interval (CI) 0.948-0.978; test: 0.946, 95% CI 0.925-0.978), followed by SVM (train: 0.925, 95% CI 0.902-0.944; test: 0.942, 95% CI 0.922-0.973) and LR (train: 0.886, 95% CI 0.858-0.910; test: 0.891, 95% CI 0.891-0.952). By this improvement, overtreatment was reduced from 21.7% (121/557) by treat-all pattern, to 15.1% (84/557) by JPN 4th criteria, and to 0.7-0.9% (4-5/557) by the new approach.
CONCLUSIONS: The decision model with machine learning analysis demonstrates high accuracy for identifying patients who are candidates for D1 versus D2 resection. Its approximate 14-20% improvements in overtreatment compared to treat-all pattern and JPN 4th criteria potentially increase the number of patients with local advanced GCs who can safely avoid unnecessary lymphadenectomy.
引用
- Liu C, Qi L, Feng QX, et al. Performance of a machine learning-based decision model to help clinicians decide the extent of lymphadenectomy (D1 vs. D2) in gastric cancer before surgical resection. Abdom Radiol (NY). 2019. PMID: 31201432. doi:10.1007/s00261-019-02098-w
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。