基于舌象客观化的转移性结直肠癌多模态特征融合诊断模型构建OA
Construction of A Multimodal Feature Fusion Diagnostic Model for Metastatic Colorectal Cancer Based on Objective Tongue Diagnosis
目的 本研究旨在探索一种基于舌诊客观化与中医特征的无创、便捷的用于诊断转移性结直肠癌(metastatic colorectal cancer,mCRC)的方法.方法 研究纳入282例结直肠癌患者(其中非转移性结直肠癌151例,转移性结直肠癌131例),使用ZK-S01便携式舌诊仪采集舌象并进行客观化分析,同时收集人口学、生活方式、疾病特征、脉象及中医症状评分等多维度信息.通过卡方检验或独立样本t检验对舌象特征和其他临床参数(生活方式、疾病特征、脉象、中医症状)进行差异性分析,以P<0.05作为入选标准筛选模型预测因子,基于Spearman分析方法对预测因子进行降维再筛选,将降维后的数据作为最终模型的预测因子.采用K-最近邻、随机森林、极度梯度提升树(extreme gradient boosting,XGBoost)、堆叠泛化(Stacking)等9种常用机器学习方法进行转移性结直肠癌多维度融合诊断模型的构建.将数据的80%作为训练集,20%作为测试集,采用5折交叉验证的方法对模型性能进行验证.以曲线下面积(area Under Curve,AUC)、F1值对模型性能进行评估.结果 区分非转移性结直肠癌与转移性结直肠癌的关键预测因子包含舌有瘀斑、RAS/BRAF基因突变、腻苔、疼痛、滑脉等;融合多种临床信息的多模态特征融合模型具有较好的准确性及泛化能力,其中Stacking算法性能最优(F1值79.25%、AUC 80.55%、敏感度80.77%、特异度80.33%、准确率80.53%、精确率82.10%).结论 融合舌象客观化特征与多维度临床信息的机器学习模型能有效区分非转移性结直肠癌与转移性结直肠癌,为临床提供了一种无创、便捷的分期诊断工具.
Objective This study aims to explore a non-invasive and convenient method for diagnosing metastatic colorectal cancer(mCRC)based on the objectification of tongue diagnosis and traditional Chinese medicine(TCM)characteristics.Methods A total of 282 colorectal cancer patients were enrolled in the study,including 151 cases of non-metastatic colorectal cancer and 131 cases of metastatic colorectal cancer.The ZK-S01 portable tongue diagnosis instrument was used to collect tongue images for objective analysis.Meanwhile,multidimensional information such as demographics,lifestyle,disease characteristics,pulse conditions,and TCM symptom scores was collected.Chi-square test or independent samples t-test was used to analyze the differences in tongue image features and other clinical parameters(lifestyle,disease characteristics,pulse conditions,TCM symptoms).Predictive factors for the model were screened with P＜0.05 as the inclusion criterion.Dimension reduction and re-screening of the predictive factors were performed using Spearman analysis,and the data after dimension reduction were used as the predictive factors for the final model.Nine common machine learning methods,including k-nearest neighbors,random forest,extreme gradient boosting(XGBoost),and Stacking,were adopted to construct a multi-dimensional fusion diagnostic model for mCRC.Eighty percent of the data was used as the training set,and 20%as the test set.Five-fold cross-validation was used to verify the model performance,and the area under the curve(AUC)and F1-score were used to evaluate the model performance.Results The key predictive factors for distinguishing between metastatic and non-metastatic colorectal cancer included tongue with ecchymoses,RAS/BRAF gene mutation,greasy tongue coating,pain,and slippery pulse.The multi-modal feature fusion model integrating various clinical information showed good accuracy and generalization ability,among which the Stacking algorithm achieved the optimal performance(F1-score:79.25%,AUC:80.55%,sensitivity:80.77%,specificity:80.33%,accuracy:80.53%,precision:82.10%).Conclusion The machine learning model integrating objective tongue image features and multidimensional clinical information can effectively distinguish between metastatic and non-metastatic colorectal cancer,providing a non-invasive and convenient staging diagnostic tool for clinical practice.
徐钰莹;杨宇飞;李芮;张继伟;张雪雪;宋婷婷;隗睿;许云;李秋艳
中国中医科学院西苑医院肿瘤科,北京 100091中国中医科学院西苑医院肿瘤科,北京 100091中国中医科学院西苑医院肿瘤科,北京 100091中国中医科学院西苑医院男科,北京 100091中国中医科学院西苑医院综合科,北京 100091中国科学院微电子研究所,北京 100029中国中医科学院西苑医院综合科,北京 100091中国中医科学院西苑医院肿瘤科,北京 100091中国中医科学院西苑医院综合科,北京 100091
医药卫生
结直肠癌分期舌象客观化多模态特征融合机器学习诊断模型
colorectal cancer stagingobjectification of tongue featuresmultimodal feature fusionmachine learningdiagnostic model
《中医肿瘤学杂志》 2026 (3)
37-52,16
首都卫生发展科研专项(编号:2022-1-4171)中国中医科学院科技创新工程(编号:CI2021A05012).
评论