Page 122 - 《中国药房》2025年6期
P. 122
·智慧药学·
基于机器学习模型预测抑郁症患者度洛西汀的血药浓度
Δ
谯 明 1, 2* ,靳 路 ,朱 毅 ,胡君萍 (1.新疆医科大学第一附属医院药学部,乌鲁木齐 830011;2.新疆药
4 #
3
1, 2
物临床研究重点实验室,乌鲁木齐 830011;3.新疆医科大学第一附属医院心理医学中心,乌鲁木齐 830011;
4.新疆医科大学药学院,乌鲁木齐 830017)
中图分类号 R971+.4 文献标志码 A 文章编号 1001-0408(2025)06-0752-06
DOI 10.6039/j.issn.1001-0408.2025.06.20
摘 要 目的 为临床尤其是无治疗药物监测条件的新疆基层医疗机构的抑郁症患者提供度洛西汀用药参考。方法 回顾性收集
2022年1月至2023年12月在新疆医科大学第一附属医院服用度洛西汀的281例抑郁症住院患者的病历资料,按7∶3比例划分为
训练集(196例)和测试集(85例)。通过随机森林(RF)模型中的“递归特征消除”程序进行特征选择,采用支持向量机、RF、极端梯
度提升(XGBoost)、人工神经网络 4 种机器学习算法构建度洛西汀血药浓度预测模型,并通过决定系数(R)、平均绝对误差
2
(MAE)、均方根误差(RMSE)评估比较4种模型的预测性能。通过夏普利加性解释方法对筛选出的最优模型的特征进行解释,确
定特征的重要性排序及其对度洛西汀血药浓度预测结果的影响大小与方向。结果 最终选择了29个特征变量,包括年龄、民族、
体重指数(BMI)等。XGBoost的R(0.808)最高,MAE(7.644)、RMSE(10.808)最低。影响度洛西汀血药浓度预测的特征重要性排
2
序为:BMI>年龄>其余20个特征集合(包括肝、肾功能和生化指标)>用药日剂量>合并疾病>联合用药>民族>白细胞计数>
血红蛋白>身高。结论 XGBoost模型预测度洛西汀血药浓度的预测性能最佳,BMI和年龄对度洛西汀血药浓度预测的影响较大。
关键词 抑郁症;度洛西汀;血药浓度;机器学习;治疗药物监测;夏普利加性解释
Prediction of duloxetine blood concentration in patients with depression based on machine learning
1, 2
4
QIAO Ming ,JIN Lu ,ZHU Yi ,HU Junping(1. Dept. of Pharmacy, the First Affiliated Hospital of Xinjiang
3
1, 2
Medical University, Urumqi 830011, China;2. Xinjiang Key Laboratory of Clinical Drug Research, Urumqi
830011, China;3. Psychological Medicine Center, the First Affiliated Hospital of Xinjiang Medical University,
Urumqi 830011, China;4. College of Pharmacy, Xinjiang Medical University, Urumqi 830017, China)
ABSTRACT OBJECTIVE To provide medication reference for duloxetine use in clinical settings, particularly for patients with
depression in primary medical institutions in Xinjiang that lack therapeutic drug monitoring conditions. METHODS The medical
records of 281 depression inpatients taking duloxetine in the First Affiliated Hospital of Xinjiang Medical University from January
2022 to December 2023 were retrospectively collected. They were divided into training set (196 cases) and test set (85 cases) in
the ratio of 7∶3. Feature selection was performed by encapsulating random forests (RF) with recursive feature elimination. Four
machine learning algorithms, namely support vector machine, RF, extreme gradient boosting (XGBoost) and artificial neural
network, were used to construct duloxetine blood concentration prediction model. The prediction performance of the models was
evaluated and compared by coefficient of determination (R), mean absolute error (MAE) and root mean squared error (RMSE).
2
The feature of the selected optimal model was explained by Shapley additive explanation method, and the importance ranking of the
features and the influence on the prediction results of duloxetine blood concentration were determined. RESULTS A total of 29
characteristic variables were selected, including age, ethnicity, body mass index(BMI), etc. XGBoost showed the highest R
2
(0.808), and the lowest MAE (7.644) and RMSE (10.808). The ranking of feature importance for predicting the blood
concentration of duloxetine was as follows: BMI>age>other 20 feature sets (including liver and kidney function and biochemical
indicators)>daily dosage>comorbidities>combination therapy>ethnicity>white blood cell count>hemoglobin>height.
CONCLUSIONS XGBoost model possesses the best prediction performance of duloxetine blood concentration; BMI and age have
a greater impact on the prediction of duloxetine blood
Δ 基金项目 新疆维吾尔自治区自然科学基金项目(No. 2022-
concentration.
D01C749) KEYWORDS depression; duloxetine; blood concentration;
*第一作者 副主任药师,博士。研究方向:临床药学。E-mail:
machine learning; therapeutic drug monitoring; Shapley
1532527694@qq.com
# 通信作者 教授,博士。研究方向:药物研究与开发。E-mail: additive explanation
hjp_yft@163.com
· 752 · China Pharmacy 2025 Vol. 36 No. 6 中国药房 2025年第36卷第6期