Xie Jiangchuan, Ma Pan, Pan Xinmei, Cao Liya, Liu Ruixiang, Xiong Lirong, Wang Hongqian, Zhang Xin, Xie Linli, Chen Yongchuan
Department of pharmacy, The First Affiliated Hospital of Army Medical University, Chongqing, China.
Medical Big Data and Artificial Intelligence Center, the First Affiliated Hospital of Army Medical University, Chongqing, China.
Pharmacopsychiatry. 2025 Jun 2. doi: 10.1055/a-2593-3125.
To develop an optimal model to predict valproic acid (VPA) concentrations by machine learning, ensuring that the VPA plasma concentration is in the effective treatment range, and thus effectively control the patient's epilepsy.This single-center, retrospective study included patients diagnosed with epilepsy from January 2014 to January 2022. Patients receiving VPA and having undergone therapeutic drug monitoring were enrolled. Top three algorithms exhibiting superior model performance were selected to establish the ensemble prediction model, with Shapley Additive exPlanations (SHAP) employed for model interpretation. An independent dataset was collected as a clinical validation group to verify the prediction model performance.The algorithms chosen for the ensemble model-Light Gradient Boosting, Categorical Boosting, and Gradient Boosted Regression Trees-demonstrated high (0.549, 0.515, and 0.503, respectively). Post-feature selection, the final model incorporated 20 variables, proving superior in predictive performance compared to models considering all 24 variables. The , mean absolute error, mean square error, absolute accuracy (±20 mg/L), and relative accuracy (±20%) of external validation were 0.621, 10.67, 221.50, 78.98%, and 66.48%, respectively. The importance and direction of each variable were visually represented using SHAP values, with VPA administration and liver function emerging as the most significant factors.The innovative application harnesses advanced multi-algorithm mining methodologies to forecast VPA concentrations in adult epileptic patients. Furthermore, it employs SHAP to elucidate the nuanced influence of each feature within the integrated prediction model, thereby providing a robust and plausible explanation for the determinants affecting VPA concentration predictions.
为了通过机器学习开发一个预测丙戊酸（VPA）浓度的优化模型，确保VPA血浆浓度处于有效治疗范围内，从而有效控制患者的癫痫。这项单中心回顾性研究纳入了2014年1月至2022年1月期间被诊断为癫痫的患者。纳入接受VPA治疗并进行过治疗药物监测的患者。选择表现出卓越模型性能的前三种算法来建立集成预测模型，并采用夏普利值法（SHAP）进行模型解释。收集一个独立数据集作为临床验证组，以验证预测模型的性能。集成模型所选用的算法——轻梯度提升、分类提升和梯度提升回归树——均表现出较高的 （分别为0.549、0.515和0.503）。经过特征选择后，最终模型纳入了20个变量，与考虑所有24个变量的模型相比，其预测性能更优。外部验证的 、平均绝对误差、均方误差、绝对准确度（±20mg/L）和相对准确度（±20%）分别为0.621、10.67、221.50、78.98%和66.48%。使用SHAP值直观地展示了每个变量的重要性和方向，VPA给药和肝功能是最显著的因素。这项创新性应用利用先进的多算法挖掘方法来预测成年癫痫患者的VPA浓度。此外，它还采用SHAP来阐明综合预测模型中每个特征的细微影响，从而为影响VPA浓度预测的决定因素提供了有力且合理的解释。
