应用机器学习挖掘奥卡西平致癫痫儿童低钠血症的危险因素

Machine learning to investigate risk factors of oxcarbazepine-induced hyponatremia in children with epilepsy

目的考察奥卡西平导致的低钠血症(OIH)在儿童患者群体中的患病率和关联危险因素。方法回顾性收集1555例口服奥卡西平治疗癫痫的病例和720例未服用奥卡西平的健康体检者资料,分别归为病例组和空白对照组。根据血清Na^(+)浓度将病例组分为Na^(+)正常(血清Na^(+)≥135 mmol·L^(-1))和OIH(Na^(+)<135 mmol·L^(-1))2个亚组,OIH又细分为轻度OIH(128 mmol·L^(-1)<Na^(+)<135 mmol·L^(-1))和严重OIH(Na^(+)≤128 mmol·L^(-1)),采用单因素分析比较各组与Na^(+)正常组在性别、年龄、肝肾功能、奥卡西平代谢物10-羟基卡马西平(MHD)血清稳态谷浓度、联合用药、合并感染等方面的差异。采用机器学习条件决策树模型筛选OIH的关联危险因素。采用受试者操作特征曲线、提升度曲线和代价曲线验证模型预测的准确性和泛化性能。结果病例组和空白对照组分别获得3208例次和733例次血清Na^(+)浓度监测资料。病例组OIH的检出频率为3.77%(121/3208),人群患病率为5.53%(86/1555);而空白对照组低钠血症的检出频率为0.41%(3/733),人群患病率为0.42%(3/720),2组间差异显著(P<0.05)。病例组严重OIH的检出频率为0.41%(13/3208),人群患病率为0.77%(12/1555),而空白对照组无严重低钠血症检出。病例组血清Na^(+)浓度显著低于空白对照组(P<0.05)。单因素分析发现OIH发生概率的影响因素有性别、年龄、尿素氮、血肌酐、MHD血清稳态谷浓度、合并感染、联用丙戊酸、联用左乙拉西坦或联用拉莫三嗪(P<0.05)。条件决策树模型筛选出OIH的发生风险与联用丙戊酸、肝肾功能情况和MHD稳态谷浓度水平相关。最终模型经验证预测准确,泛化性能强。结论在口服奥卡西平治疗癫痫儿童患者中,低钠血症的患病率高于健康人群,本研究建立的条件决策树模型适用于对OIH危险因素的识别。

AIM To estimate the frequency of oxcarbazepine(Oxc-induced hyponatremia(OIH)in pediatric patients with epilepsy and identify the risk factors for its development.METHODSA total of1555 pediatric patients with epilepsy treated by oral Oxc and 720 healthy people for health examination who did not take Oxc were collected retrospectively and classified into case group and blank control group respectively.According to the serum sodium concentrations,the patients in the case group were divided into normal subgroup(Na^(+)≥135 mmol·L^(-1)and OIH subgroup(Na^(+)<135 mmol·L^(-1))1I).Univariate analysis was used to compare the dfferences between each group and Na^(+)normal group in gender,age,liver and kidney function indexes,steady-state trough concentration of 10-hydroxycarbamazepine(MHD,the metabolite of Oxc),drug combinations and coinfections.Machine learning conditional decision tree model was used to screen the related risk factors of OIH.The accuracy and generalization of the model were verified by using the receiver operating characteristic curve,lift curve and cost curve.RESULTS The monitoring data of serum sodium concentration in the case group and the blank control group were obtained for 3208 times and 733 times respectively in this study.The frequency of OIH in case group was 3.77%(121/3208),and the prevalence was 5.53%(86/1555).The frequency of hyponatremia in the control group was 0.41%(3/733),and the prevalence was 0.42%(3/720).There was a significant difference between the two groups(P<0.05).Meanwhile,the frequency of severe 0IH in the case group was 0.41%(13/3208),and the prevalence was 0.77%(12/1555),while no severe hyponatremia was found in the healthy examinees.The serum sodium concentrations in the case group were significantly lower than those in the control group(P<0.05).Univariate analysis indicated that the probability of OIH occurrence was significantly affected by the factors such as genger,age,serum urea nitrogen(BUN),creatinine(CR),MHD serum steadystate trough concentration,coinfections,drug combinations with valproate,levetiracetam or lamotrigine(P<0.05).The conditional decision tree model screened that the risk of OIH was related to the combination of valproate,liver and kidney function and MHD steady-state trough concentration.The final model was verified to be accurate and had strong generalization performance.CONCLUSION The prevalence of hyponatremia in pediatric patients with epilepsy treated with oral Oxc is higher than that in healthy people.The conditional decision tree model established in this study is applicable to the identification of risk factors of OIH.