Epilepsy is a prevalent chronic neurological disorder, with a relatively high incidence rate among pediatric patients. The overall prevalence of epilepsy among children is approximately 151/100,000, a rate more than quadruple that of adults (Ding et al., 2021; Lin et al., 2021). At present, pharmacotherapy remains the cornerstone of seizure management (Liu et al., 2023). Among the available anti-seizure medications (ASMs), valproic acid (VPA) stands out as a broad-spectrum ASM, effective against various seizure types, including generalized, absence, partials, and myoclonic seizures. Its broad-spectrum efficacy has established VPA as a frontline clinical option for the treatment of pediatric epilepsy patients (PEPs) (Glauser et al., 2013). However, the therapeutic window for VPA, ranging from 50–100 mg/L, is relatively narrow and characterized by significant interindividual and intraindividual variability. These variability pose challenges for clinicians to manage dosages precisely and limiting the clinical application of VPA in treating PEPs (Johannessen Landmark et al., 2020).

In epilepsy management, medication safety is crucial. If VPA trough concentrations fall below the therapeutic range, seizures may not be adequately controlled, while exceeding this range may trigger severe adverse reactions in the hematopoietic and nervous systems, including anemia, thrombocytopenia, and ataxia, or even result in potentially lethal hepatotoxicity (Hernández García et al., 2023; Star et al., 2014). For adolescent girls, particularly those with childbearing potential, it is essential to reassess the risks associated with VPA exposure on an annual basis (Toledo et al., 2021). In the course of VPA dosage titration, it is imperative to align the trough concentration within the therapeutic range and to identify the most appropriate personalized concentration range for each PEP, for the therapeutic range is often more constrained, requiring further fine-tuning of blood concentrations to the desired therapeutic threshold within an ideal range (Johannessen and Landmark, 2008). Consequently, for PEPs necessitating prolonged therapeutic drug monitoring of VPA, the support of well-suited model-guided individualized VPA dosing tools can enhance the efficacy of monitoring VPA trough concentrations and adjusting VPA dosages, thereby enabling the establishment of more precise personalized treatment regimens.

Traditional methods for monitoring VPA trough concentrations, such as chromatography or immunoassay techniques, though accurate, are hampered by the problems of high cost and long detection cycles. In addition, these invasive detection methods can increase the psychological burden on PEPs, making them a less favorable option for those who require long-term monitoring. Therefore, model-guided individualized VPA dosing tools has currently surfaced as a heated research focus in the clinical arena, which may hold promises for reducing the number of invasive tests and producing an efficient and effective therapeutic range. Currently, the prevalent method for model-guided individualized dosing tools for VPA is the Nonlinear Mixed Effects Model (NONMEM). This method has been widely applied in population pharmacokinetics (popPK) studies (Gu et al., 2021). However, because it is based on parameter estimation, the operation and establishment process of the model is relatively complex, requiring a lot of time and efforts in constructing, validating, and optimizing the model. A recent systematic review meticulously curated and selected 10 published popPK models for PEPs, rigorously delineating their monitoring performance (Zhang et al., 2023). The findings indicated that the majority of these popPK models performed suboptimally in the context of pediatric epilepsy, which may potentially be attributed to the complexity of clinical data, insufficient sample sizes, deficiencies in model algorithms, and a lack of external data validation. Therefore, it is necessary to explore a more convenient and potent model for monitoring VPA trough concentrations in treating PEPs.

Currently, machine learning (ML) algorithms have exhibited significant potential in the medical field. They excel at processing complex and extensive clinical data and have been successfully applied in the monitoring of drug concentrations, including cyclosporine (Mao et al., 2022), mycophenolic acid (Shao et al., 2022), sertraline (Fu et al., 2024), and vancomycin (Huang et al., 2021). Moreover, preliminary studies have begun to emerge in the monitoring of VPA, such as the application of Fourier-transform infrared (FT-IR) spectroscopy combined with nonlinear support vector regression algorithms to construct a VPA monitoring model (El Orche et al., 2023) and the development of an XGBoost model to predict VPA trough concentrations by integrating covariates from various population pharmacokinetic models (Zhu et al., 2022a). Although ML-based VPA monitoring models have shown considerable potential in integrating real-world data and predicting VPA concentrations (Hsu et al., 2024), there are still limitations in external validation, comprehensiveness of evaluation metrics, data processing, and model interpretability. These are essential for guiding clinicians on which clinical indicators to monitor when treating PEPs, and they are what this study aims to address.

Ensemble models (Huang et al., 2021; Ma et al., 2022), by virtue of their ability to combine various ML algorithms to enhance the accuracy and robustness of monitoring models, are gaining increasing importance in the field of ML. Herein, we ranked the weights of each covariate across 3 ML models—Gradient Boosting Regression Trees (GBRT), Random Forest Regression (RFR), and Support Vector Regression (SVR) according to their SHapley Additive exPlanation (SHAP) (Ma et al., 2022). We then employed these algorithms to formulate the first ensemble model for monitoring VPA trough concentrations in PEPs. Our findings indicate that this ensemble model exhibits superior prediction performance, highlighting its significant potential in clinical applications.

The ensemble model is a robust ML technique (Supplementary Table S6) that enhances overall monitoring performance by combining the predictive characteristics of multiple models. This approach is recognized for reducing overfitting and enhancing generality, and often outperforms a single model in various scenarios (Ma et al., 2022). In this study, we developed the first ML-based ensemble model for monitoring VPA trough concentration in PEPs, with the goal of maximizing the R² value, relative accuracy, and absolute accuracy to optimize the performance. We selected three outstanding algorithms (GBRT, RFR, and SVR) and reallocated the weights of their R² values to reflect their importance and contribution in the ensemble model. By utilizing advanced ML techniques, we achieved automatic computation and model optimization, ultimately developing an ensemble model with an optimal comprehensive performance, with the R² weight ratio of GBRT, RFR, and SVR optimized at 5:2:3. Compared with an individual covariate model, this ensemble model demonstrated superior advantages across multiple dimensions and was further fully validated by two independent external sample groups (in-hospital and out-of-hospital datasets), demonstrating good generality.

SHAP plot visualization analysis plays a crucial role in the construction of ML models. It not only assigns a specific set of SHAP values to each covariate, clearly showing their importance and direction of influence on the monitoring results of the model, but also potentially explains the correlation between each covariate and its clinical significance (Cheng et al., 2023). Existing studies have shown that the daily dose and dosage form of VPA significantly affect VPA blood concentration (Zhu et al., 2022a; Xu et al., 2020). Consistently, in our study, the VPA daily dose ranked high in the SHAP visualization analysis [GBRT (second, 20.8%), RFR (second, 21.4%), and SVR (third, 13.6%)] (Figure 2); the VPA dosage form ranked first (24.6%) in SVR; and both the VPA daily dose and dosage form were positively correlated with VPA trough concentration (Figure 3), indicating that they are among the most important covariates. Therefore, in the clinical medication process, especially during a switch between VPA dosage forms (oral solution and sustained-release tablet), due to the differences in drug release and absorption between the two dosage forms, regular monitoring of VPA blood concentration in PEPs is necessary in case of VPA dosage adjustment.

VPA is commonly associated with hematological adverse reactions, particularly with thrombocytopenia and red blood cell dysplasia, which may be related to its impact on bone marrow hematopoietic function (Langlie et al., 2022; Acharya and Bussel, 2000). It is important to note that high concentrations of VPA may not only reduce platelets but also inhibit platelet aggregation, fibrinogen, and other factors, which can increase the bleeding risk in patients with epilepsy (Kumar et al., 2019; Beydoun et al., 1997). Additionally, although there is no evidence that VPA directly acts on RBCs, available data confirm that VPA may interact with RBCs through different mechanisms, including affecting the distribution of drugs within RBCs and enhancing the expression of erythropoietin (EPO) protein, which may affect RBC production (Mancl and Gidal, 2009; Rubiyana et al., 2020). This study showed that VPA trough concentration was inversely proportional to PLT and RBC levels (Figure 3), meaning that high VPA concentrations are more likely to cause a decrease in PLT and RBC, inducing the occurrence of thrombocytopenia and anemia, which is consistent with the existing research (Nasreddine and Beydoun, 2008; Nasreddine et al., 2022; Ma et al., 2019). Furthermore, SHAP plot analysis ranked RBC relatively high (consistently ranked fourth across the three models, with a contribution of 8.5%–10.3%), and PLT as one of the important covariates in different individual covariate models (Figure 2): GBRT (first, 24.8%), RFR (first, 37.2%), and SVR (second, 17.7%). For ALB, another important hematological indicator affecting VPA drug concentration, patients with hypoalbuminemia (below 35 g/L) may have significantly-increased free VPA concentrations (Drisaldi et al., 2019). However, the PEPs data analysis in this study showed that ALB fell within the normal range (44.98 ± 2.58 g/L) and was not correlated with VPA trough concentration (r = −0.026, p > 0.05). This phenomenon can explain why the weight of ALB in monitoring VPA trough concentration was relatively low. The above results indicate that the monitoring model for VPA trough concentration constructed in this study has strong clinical relevance. Therefore, in clinical practice, for PEPs receiving VPA, regular blood tests should be recommended to monitor potential adverse reactions in the hematological system, especially for PEPs with trough concentrations above the therapeutic window, in order to adjust VPA dosage in a timely manner.

CREA and UREA are two key clinical indicators for assessing kidney function. In the SHAP plot analysis, CREA ranked low in the three independent covariate models, indicating its relatively small impact on the model output. In contrast, UREA ranked third in the GBRT and RFR algorithms, highlighting the key role of UREA in VPA trough concentration (Figure 2). Existing studies have confirmed that VPA concentration is positively correlated with the severity of renal tubular damage (Knights and Finlay, 2014; Mazaheri et al., 2011) and that CREA and UREA levels are also positively correlated with VPA concentration (Yang et al., 2020). These findings suggest that increased levels of CREA and UREA generally indicate impaired kidney filtration function and can reflect VPA blood concentration levels to some extent. Changes in UREA levels need to be comprehensively assessed in conjunction with other indicators and clinical conditions. In some cases, UREA may not be as sensitive as CREA, leading to the clinical practice of using both to assess kidney function. For PEPs who use VPA on a long-term basis or have specific risk factors, it is essential to prescribe a regular monitoring of the kidney function and adjust the treatment plan correspondingly (Mazaheri et al., 2011). Therefore, incorporating CREA and UREA as covariates in our model has significant clinical implications for the effective management of VPA in treating PEPs.

Studies have shown that genetic variations in CYP2C9, CYP2C19, CYP2A6, and CYP2B6 affect the in vivo concentrations of VPA and its metabolites (Zhao et al., 2020). The risk of hepatotoxicity increases when VPA is used in children under 6 years old (Star et al., 2014). ALT is an important indicator for assessing liver function, in which elevated levels may indicate liver damage (Ghozzi et al., 2011). In this study, consistently, ALT was negatively correlated with VPA trough concentration (r = −0.103, p < 0.05), indicating that ALT is positively correlated with the concentration of VPA metabolites (Chen et al., 2012). In patients with strong metabolic capacity, especially children, VPA trough concentration is low, while the concentration of hepatotoxic metabolites is high, which may increase the risk of liver damage with elevated ALT (Zhao et al., 2020). However, in the pediatric patient population of this study, no cases of severe liver function damage were observed and ALT levels of most children remained within the normal range. This phenomenon can explain why the weight of ALT in monitoring VPA trough concentration was relatively low (Figure 2). Monitoring ALT levels is important for assessing and preventing liver damage.

In the GBRT, RFR, and SVR individual covariate models, the importance of the ten covariates varied to different extents (Figure 2). Daily dose, PLT, and RBC showed relatively consistent importance across these three models. However, the importance of VPA dosage form and UREA fluctuated significantly in these models. To balance the impact of each covariate on the output of the monitoring model and to enhance the role of important covariates, we decided to integrate these three independent models into an ensemble model to further improve the monitoring performance of the model. Our study results also confirmed this: the ensemble model showed consistent performance in both the training and testing datasets, indicating a proper model fitting. Additionally, compared with the three individual covariate models, this ensemble model demonstrated a clear superiority in monitoring performance (Figures 5, 6). The predictive results of this ensemble model are more in line with the clinical reality, which has profound practical value for clinical medication guidance and timely treatment.

The ensemble model showed significant differences in monitoring performance across different age groups. The in-depth analysis of the monitoring results of different age groups revealed that the model reported a relatively poor monitoring performance in patients under 3 years old but the best in the 3–10-years age group (Figure 7). The SHAP dependency plot illustrates the contribution of age feature values to the predictive outcome (Figure 3; Supplementary Figure S4). The data group for ages below 3 corresponds to SHAP values within the ±2 range (Supplementary Figure S4), indicating a lower contribution. Moreover, this data group has a significantly smaller sample size compared to other age subgroups, which may be one of the reasons for the relatively low relative and absolute accuracy of our model’s predictions. The metabolism of VPA is complex, involving three main pathways: mitochondrial-mediated β-oxidation, cytochrome P-450 (CYP450) catalyzed oxidation, and uridine diphosphate glucuronosyltransferase (UGTs) glucuronidation (Silva et al., 2008). Glucuronidation accounts for approximately 50% of VPA metabolism and is the most important factor affecting VPA blood concentration. However, UGTs gene polymorphism is developmentally regulated, with age affecting its metabolic rate, and young children have a higher metabolic capacity (Zhao et al., 2020). Some researchers believe this may be due to children having higher enzyme activity and a larger liver-to-body size ratio (Kearns et al., 2003). Therefore, the expression of UGTs gene polymorphisms in metabolic enzymes may significantly affect VPA metabolism, thereby affecting blood concentration and leading to significant inter-individual differences in blood concentration (Ghodke-Puranik et al., 2013). As this study did not include UGTs gene polymorphism as a covariate in the monitoring model, further research is needed to validate whether the poor monitoring effect observed in patients under 3 years old is related to the age-related differences in UGTs glucuronidation.

The ten covariates included in this study are not sufficient to fully represent all potential factors that may affect the monitoring of VPA trough concentration. Other factors that distribute VPA in the body, such as metabolic gene polymorphism and hypoalbuminemia, should be considered when constructing a monitoring model. To develop a mature monitoring model, more in-depth research work is needed. The current model has limited data outside the therapeutic window, which may, to some extent, limit the accuracy of model predictions. Especially in the patient population under 3 years old, the monitoring effect was not ideal, coupled with the relatively low rate of VPA trough concentration reaching the target range in this group, suggesting a need of due attention to this phenomenon. Therefore, follow-up studies need to acquire more data from the patient population under 3 years old to optimize the monitoring performance and generality of the model.

In translating the application of a ML-based ensemble model into clinical practice, several essential steps are involved (Zhu et al., 2022b): 1. collection of clinical data: gather more clinical sample datasets, incorporating a wider range of clinical feature variables including disease status, concomitant medications, and genetic polymorphisms, to further enhance the model’s clinical applicability. 2. external validation: require additional external validation datasets to assess the model’s accuracy and reliability in clinical settings. 3. drug clinical feature library: Establish a drug clinical feature library, similar to a disease database, capable of automatically capturing clinically relevant data for variables associated with valproic acid trough concentrations, especially data that can be aligned with testing times. 4. model transformation: transform the model into software, a website, or integrate it into electronic health records (EHRs) to enable automatic updates and dosage predictions, facilitating its use by clinical physicians. Correspondingly, potential barriers in clinical practice include (Tang et al., 2024): 1. quality and integrity of clinical data. 2. complexity and uncertainty of clinical data. 3. technical integration and compatibility. 4. ethical and privacy issues of clinical data.

The ensemble model proposed in this study, based on 3 ML techniques, can monitor the VPA trough concentrations in PEPs with a high degree of satisfaction. The advantage of this ensemble model lies in its ability to consider the importance of each covariate across different models, maximizing the utility of each covariate to enhance the overall monitoring performance. The monitoring results it yields are more clinically relevant, offering significant practical value for the individualized adjustment of clinical drug dosages and timely interventions in the precision treatment of PEPs.