Chen Yue-Wen, Lin Xi-Kai, Chen Si, Zhang Ya-Lan, Wu Wei, Huang Chen, Rao Xin, Lu Zong-Xing, Liu Zhou-Jie
Department of Pharmacy, The First Affiliated Hospital, Fujian Medical University, Fuzhou, China.
Department of Pharmacy, National Regional Medical Center, Binhai Campus of the First Affiliated Hospital, Fujian Medical University, Fuzhou, China.
Front Pharmacol. 2024 Dec 18;15:1521932. doi: 10.3389/fphar.2024.1521932. eCollection 2024.
Few personalized monitoring models for valproic acid (VPA) in pediatric epilepsy patients (PEPs) incorporate machine learning (ML) algorithms. This study aimed to develop an ensemble ML model for VPA monitoring to enhance clinical precision of VPA usage.
A dataset comprising 366 VPA trough concentrations from 252 PEPs, along with 19 covariates and the target variable (VPA trough concentration), was refined by Spearman correlation and multicollinearity testing (366 × 11). The dataset was split into a training set (292) and testing set (74) at a ratio of 8:2. An ensemble model was formulated by Gradient Boosting Regression Trees (GBRT), Random Forest Regression (RFR), and Support Vector Regression (SVR), and assessed by SHapley Additive exPlanations (SHAP) analysis for covariate importance. The model was optimized for R, relative accuracy, and absolute accuracy, and validated against two independent external datasets (32 in-hospital and 28 out-of-hospital dataset).
Using the R weight ratio of GBRT, RFR and SVR optimized at 5:2:3, the ensemble model demonstrated superior performance in terms of relative accuracy (87.8%), absolute accuracy (78.4%), and R (0.50), while also exhibiting a lower Mean Absolute Error (9.87) and Root Mean Squared Error (12.24), as validated by the external datasets. Platelet count (PLT) and VPA daily dose were identified as pivotal covariates.
The proposed ensemble model effectively monitors VPA trough concentrations in PEPs. By integrating covariates across various ML algorithms, it delivers results closely aligned with clinical practice, offering substantial clinical value for the guided use of VPA.
针对小儿癫痫患者（PEP）丙戊酸（VPA）的个性化监测模型很少采用机器学习（ML）算法。本研究旨在开发一种用于VPA监测的集成ML模型，以提高VPA使用的临床精准度。
通过Spearman相关性和多重共线性检验（366×11）对包含来自252名PEP的366个VPA谷浓度以及19个协变量和目标变量（VPA谷浓度）的数据集进行优化。数据集按8:2的比例分为训练集（292）和测试集（74）。通过梯度提升回归树（GBRT）、随机森林回归（RFR）和支持向量回归（SVR）构建集成模型，并通过SHapley加性解释（SHAP）分析评估协变量重要性。对该模型进行R、相对准确度和绝对准确度优化，并针对两个独立的外部数据集（32个院内数据集和28个院外数据集）进行验证。
使用优化后的GBRT、RFR和SVR的R权重比为5:2:3，集成模型在相对准确度（87.8%）、绝对准确度（78.4%）和R（0.50）方面表现出卓越性能，同时外部数据集验证显示其平均绝对误差（9.87）和均方根误差（12.24）更低。血小板计数（PLT）和VPA每日剂量被确定为关键协变量。
所提出的集成模型可有效监测PEP中的VPA谷浓度。通过整合各种ML算法中的协变量，它得出的结果与临床实践紧密相符，为VPA的指导使用提供了重大临床价值。
