利用深度学习预测头颈部放疗期间是否需要积极营养补充
本研究评估271例接受放疗的头颈癌患者,利用16项临床特征、计划CT、三维剂量及治疗中锥形束CT构建模型预测放疗期间需喂养管或体重下降≥10%。逻辑回归临床模型AUC为0.69,医学网卷积神经网络影像模型AUC为0.73,融合模型AUC为0.75且优于单独临床模型(p=0.001)。
要点
- 融合临床、剂量与治疗中CBCT的人工智能模型预测喂养管需求AUC达0.75。
- 融合模型性能优于单独临床模型(p=0.001)。
- 影像模型采用迁移学习的医学网卷积神经网络,AUC为0.73。
核心结论
- 融合临床参数、剂量分布与治疗中锥形束CT的人工智能模型在识别需放置反应性喂养管方面取得最高性能。
研究设计与样本
- 评估271例接受放疗的头颈癌患者,收集16项临床特征、计划CT、三维剂量及治疗中锥形束CT。
- 预测结局为放疗期间需喂养管或体重下降≥10%。
- 采用逻辑回归、支持向量机、多层感知机三种常规分类器基于16项临床特征建模,卷积神经网络分析影像数据,进行五折交叉验证,以AUC比较模型性能,并用R-4.1.2中的配对DeLong检验进行ROC分析。
- 临床与影像模型结果通过证据推理规则融合进行最终预测。
主要发现
- 逻辑回归模型在临床数据集上表现最佳,AUC为0.69。
- 通过迁移学习训练的医学网卷积神经网络在影像数据集上表现最佳,AUC为0.73。
- 融合临床与影像的模型AUC为0.75,统计学上优于单独临床模型(p=0.001)。
关键数值与推荐
- 预测结局定义为需喂养管或体重下降≥10%。
- 逻辑回归临床模型AUC 0.69。
- 医学网卷积神经网络影像模型AUC 0.73。
- 融合模型AUC 0.75,与单独临床模型比较p=0.001。
临床怎么用
- 原文未据此给出推荐。
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
PURPOSE/OBJECTIVES: Radiation therapy (RT) for the treatment of patients with head and neck cancer (HNC) leads to side effects that can limit a person's oral intake. Early identification of patients who need aggressive nutrition supplementation via a feeding tube (FT) could improve outcomes. We hypothesize that traditional machine learning techniques used in combination with deep learning techniques could identify patients early during RT who will later need a FT. MATERIALS/METHODS: We evaluated 271 patients with HNC treated with RT. Sixteen clinical features, planning computed tomography (CT) scans, 3-dimensional dose, and treatment cone-beam CT scans were gathered for each patient. The outcome predicted was the need for a FT or ≥10% weight loss during RT. Three conventional classifiers, including logistic regression (LR), support vector machine, and multilayer perceptron, used the 16 clinical features for clinical modeling. A convolutional neural network (CNN) analyzed the imaging data. Five-fold cross validation was performed. The area under the curve (AUC) values were used to compare models' performances. ROC analyses were performed using a paired DeLong Test in R-4.1.2. The clinical and imaging model outcomes were combined to make a final prediction via evidential reasoning rule-based fusion. RESULTS: The LR model performed the best on the clinical dataset (AUC 0.69). The MedicalNet CNN trained via transfer learning performed the best on the imaging dataset (AUC 0.73). The combined clinical and image-based model obtained an AUC of 0.75. The combined model was statistically better than the clinical model alone (p = 0.001). CONCLUSIONS: An artificial intelligence model incorporating clinical parameters, dose distributions and on-treatment CBCT achieved the highest performance to identify the need to place a reactive feeding tube.
引用
- Dohopolski M, Wang K, Morgan H, et al. Use of deep learning to predict the need for aggressive nutritional supplementation during head and neck radiotherapy. Radiother Oncol. 2022. PMID: 35461951. doi:10.1016/j.radonc.2022.04.016
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。