|
| 基于实验室指标和机器学习的Wilson病营养风险模型 |
| A nutritional risk model for Wilson,s disease based on laboratory indicators and machine learning |
| 投稿时间：2025-06-05 |
| DOI：10.3969/j.issn.1000-0399.2025.11.003 |
| 中文关键词: Wilson病 机器学习 实验室指标 风险评估 列线图模型 |
| 英文关键词: Wilson′s disease Machine learning Laboratory indicators Risk assessment Nomogram model |
| 基金项目:大健康研究院新安医学与中医药现代化研究所专项资金资助(编号:2023CXMMTCM002), 安徽省高等学校科学研究项目(自然科学类)(编号:2023AH040104), 安徽省重点研究与开发计划(编号:S202204295107020135), 安徽中医药大学科研基金项目(编号:2021sfy1c03) |
|
| 摘要点击次数: 743 |
| 全文下载次数: 171 |
| 中文摘要: |
| 目的 基于实验室指标构建Wilson病（WD）营养障碍风险评估的机器学习模型，SHAP解释模型特征贡献并绘制列线图用于临床应用。方法 回顾性收集2020年9月至2024年9月安徽中医药大学神经病学研究所附属医院确诊为207例WD患者临床资料，按照8∶2拆分为训练集和验证集，采用主观全面营养评估（SGA）表、营养风险筛查NRS 2002评分、RFH-NPT量表综合判定是否存在营养风险。存在营养风险为结局变量，以年龄、性别、病程、身体质量指数（BMI）、UWDRS量表3部分得分以及血常规、肝肾功能、电解质、铜生化等共58个指标作为特征变量，采用XGBoost回归模型训练数据并提取特征的重要性，基于相关系数对特征进行筛选，剔除高度相关的特征，通过穷举法生成所有可能的特征组合，使用交叉验证计算每个组合的R2分数，选择最优特征组合，使用网格搜索（GridSearchCV）对XGBoost模型进行超参数优化，选择最佳的模型配置，SHAP解释模型特征贡献并基于最优特征组合绘制列线图。结果 XGBoost回归模型训练数据并提取特征的重要性，获得排名前15的特征分别为环氧化酶（COX）、BMI、降钙素原（PCT）、红细胞分布宽度-标准差（RDW_SD）、胆碱酯酶（CHE）、平均红细胞体积（MCV）、性别、血钠（NA）、碱性磷酸酶（ALP）、总胆汁酸（TBA）、三酰甘油（TG）、丙氨酸氨基转移酶（ALT）、统一帕金森病评定量表第2部分（UWDRS-2）、清蛋白（ALB）、锌（ZN），且排名前15的特征直接相关性系数均未超过0.95，穷举法获得最优特征组合为（COX, BMI, PCT, MCV, ALT, ALB, ZN），平均CV R2 0.577 078,BMI的SHAP值明显大于其他特征；基于最优特征组合的列线图（逻辑回归）模型训练集准确率0.8056（80.56%），测试集准确率：0.8254（82.54%），最佳Threshold指数为0.58。结论 基于相关系数剔除多重共线性与穷举法优化的XGBoost模型获得最优特征组合（COX, BMI, PCT, MCV, ALT, ALB, ZN）构建列线图可有效预测WD营养障碍风险，利于临床应用及广泛推广。 |
| 英文摘要: |
| Objective To construct a machine learning model for assessing the risk of nutritional disorders in Wilson’s disease（WD） based on laboratory indicators, interpret the feature contributions using SHAP, and develop a nomogram for clinical application. Methods A total of 207 patients with confirmed WD were retrospectively collected from the Affiliated Hospital of Neurology Institute, Anhui University of Chinese Medicine, and divided into a training set and a validation set at an 8:2 ratio. The presence of nutritional risk, defined as the outcome variable, was comprehensively determined using the Subjective Global Assessment（SGA）, Nutritional Risk Screening 2002（NRS2002）, and Royal Free Hospital-Nutritional Prioritizing Tool（RFH-NPT）. A total of 58 indicators were included as feature variables, including age, gender, disease duration, BMI, scores of three subscales of the Unified Wilson’s Disease Rating Scale（UWDRS）, blood routine parameters, liver and kidney function indices, electrolytes, and copper-related biochemical indicators. The XGBoost regression model was used to train the data and extract feature importance. Features were filtered based on correlation coefficients to eliminate highly correlated ones. All possible feature combinations were generated using an exhaustive method, and the R2 score of each combination was calculated via cross-validation to select the optimal feature set. Hyperparameter optimization of the XGBoost model was performed using grid search（GridSearchCV） to determine the best model configuration. SHAP was used to interpret feature contributions, and a nomogram was plotted based on the optimal feature combination. Results The XGBoost regression model extracted feature importance, and the top 15 features were COX, BMI, PCT, RDW-SD, CHE, MCV, sex, NA, ALP, TBA, TG, ALT, UWDRS-2, ALB, and ZN. The correlation coefficients between these top 15 features did not exceed 0.95. The exhaustive method identified the optimal feature combination as（COX, BMI, PCT, MCV, ALT, ALB, ZN） with a mean cross-validation R2 of 0.577 078. The SHAP value of BMI was significantly higher than that of other features. The nomogram（logistic regression） model based on the optimal feature combination achieved an accuracy of 0.805 6（80.56%） in the training set and 0.8254（82.54%） in the test set, with the optimal threshold index of 0.58. Conclusion The XGBoost model optimized by eliminating multicollinearity via correlation coefficients and using the exhaustive method identifies the optimal feature combination（COX, BMI, PCT, MCV, ALT, ALB, ZN）. The nomogram constructed based on this combination can effectively predict the risk of nutritional disorders in WD, facilitating clinical application and widespread promotion. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
