Factors associated with drug-resistant epilepsy in children with epilepsy: a retrospective cohort study

Abstract

Background

Drug‑resistant epilepsy (DRE) remains a major clinical challenge in children. Understanding clinical factors associated with drug resistance is important for guiding management. We aimed to investigate factors independently associated with DRE in a large cohort of children with epilepsy.

Methods

A retrospective cohort study was conducted on children with confirmed epilepsy who received regular follow‑up at two tertiary pediatric epilepsy centers. Patients were classified into DRE [failure of ≥2 adequate anti-seizure medications (ASMs) trials] and non‑DRE groups according to ILAE criteria. Baseline demographic, clinical, electroencephalographic, neuroimaging, and treatment‑related variables were collected. Univariate and multivariate logistic regression analyses were performed to identify factors independently associated with DRE. Treatment‑related variables (number of ASMs tried, maximum concurrent ASMs, and polytherapy at last follow‑up) were excluded from the multivariable model due to their intrinsic relationship with the definition and clinical consequences of DRE. Model discrimination was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), and calibration was assessed by the Hosmer‑Lemeshow test. The 95% confidence intervals (CIs) for performance metrics were obtained via bootstrap resampling.

Results

A total of 412 children were enrolled, including 100 (24.3%) with DRE. Children with DRE had significantly younger age at onset (4.30±3.40 vs. 6.01±3.50 years, P<0.001), higher baseline seizure frequency, more frequent status epilepticus, neurodevelopmental disorders, and psychiatric comorbidities (all P<0.05). In multivariable analysis excluding treatment‑related variables, generalized seizure type (vs. focal) [adjusted odds ratio (OR) =0.20; 95% CI: 0.04–0.62; P=0.01), history of status epilepticus (adjusted OR =4.86; 95% CI: 2.15–11.13; P<0.001), neurodevelopmental disorder (adjusted OR =3.40; 95% CI: 1.49–7.85; P=0.004), structural etiology (vs. genetic) (adjusted OR =0.25; 95% CI: 0.07–0.92; P=0.040), and older age at onset (adjusted OR =0.91; 95% CI: 0.82–1.00; P=0.046) were independently associated with DRE. The model showed moderate discriminative ability (AUC =0.802; 95% CI: 0.751–0.872).

Conclusions

In this retrospective cohort, seizure type, history of status epilepticus, neurodevelopmental disorder, structural etiology, and older age at onset were independently associated with DRE after excluding treatment‑related variables. These findings suggest that true baseline independently associated variables of DRE are limited, and the diagnosis of DRE itself—failure of two adequate ASM trials—remains the most practical indicator for considering non‑pharmacological therapies. External validation is needed before clinical application.

Highlight box.

Key findings

• Among 412 children with epilepsy, 24.3% developed drug-resistant epilepsy (DRE).

• After excluding treatment-related variables, generalized seizure type, history of status epilepticus, neurodevelopmental disorder, structural etiology, and older age at onset were independently associated with DRE.

• The model showed moderate discriminative ability (area under the receiver operating characteristic curve =0.802).

What is known and what is new?

• DRE is common in children, and previous studies often analyzed treatment-related factors together with baseline factors.

• After excluding treatment-related variables, the true baseline factors independently associated with DRE are limited, and the diagnosis of DRE itself remains the most practical indicator for considering non-pharmacological therapies.

What is the implication, and what should change now?

• Clinical decisions should not rely too heavily on baseline factors to predict DRE; instead, focus on the failure of two adequate anti-seizure medication trials as the key threshold.

• External validation is still needed before this model can be applied in clinical practice.

Introduction

Epilepsy represents one of the most frequently encountered neurological conditions in the pediatric population, with approximately 239 new cases diagnosed annually per 100,000 children (1). While many children achieve seizure freedom with appropriate anti‑seizure medications (ASMs), approximately 30% develop drug‑resistant epilepsy (DRE) as defined by the International League Against Epilepsy (ILAE)—failure of two tolerated, appropriately chosen, and used ASM schedules to achieve sustained seizure freedom (2). For children, DRE imposes a substantial multifaceted burden. Beyond its direct impact on seizure control, it is associated with elevated risks of premature mortality, cognitive impairment, psychosocial difficulties, and diminished quality of life—consequences that frequently extend well into adult life and generate considerable healthcare costs (3-5).

Early recognition of factors associated with DRE is clinically important because it may facilitate timely consideration of non‑pharmacological interventions—epilepsy surgery, ketogenic diet, or neuromodulation—that can significantly improve outcomes, cognitive development, and quality of life (6). Delays in recognition and referral are common, partly due to uncertainty about which clinical features reliably predict drug resistance. Over the past two decades, numerous studies have investigated potential risk factors for DRE. These include early age at seizure onset, high baseline seizure frequency, history of status epilepticus, structural brain lesions on magnetic resonance imaging (MRI), epileptiform discharges on etiologies electroencephalography (EEG), neurodevelopmental comorbidities, and specific etiologies (7-9). However, the findings across these studies have been inconsistent. This variability can be attributed to several factors, including the inclusion of heterogeneous populations (with varying age ranges and epilepsy syndromes), the use of differing definitions of DRE (pre-ILAE criteria vs. the current ILAE consensus definition), limited sample sizes, and insufficient adjustment for treatment-related confounders.

The relationship between treatment history and DRE remains incompletely understood. Specifically, it is unclear whether the number of ASMs tried and the use of polytherapy should be considered independently associated variables of pharmacoresistance or simply markers of established disease severity. While some investigators view the need for multiple ASM trials as an indicator of underlying epilepsy severity rather than a causal factor, others have suggested that early treatment complexity may itself predict eventual refractoriness (10,11). Disentangling these relationships requires rigorous multivariable modeling that accounts for disease severity, etiology, and comorbidities.

To address these unresolved questions, we designed a retrospective cohort study of children with epilepsy treated at two tertiary referral centers. The primary aim was to identify clinical factors independently associated with DRE in this population, with the goal of generating evidence to support early recognition of high-risk children and facilitate timely referral for non-pharmacological evaluation. We present this article in accordance with the STROBE reporting checklist (12) (available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0593/rc).

Methods

Study design and setting

This retrospective cohort study was conducted at two tertiary pediatric epilepsy centers (Taizhou Maternal and Child Health Hospital and The Second Affiliated Hospital of Zhejiang University School of Medicine) in China. The study spanned from January 2023 to April 2025, covering patient recruitment, exposure ascertainment, follow-up, and data collection. Patients were diagnosed between January 2023 and October 2024, ensuring a minimum follow-up of 6 months for all included patients before data collection ended in April 2025. Data were collected from electronic medical records, with the date of outcome assessment (final follow-up or DRE determination date) defined as the last documented clinical visit or the date when DRE criteria were met. At each participating center, patients were managed by the same dedicated pediatric epilepsy team throughout the study period, ensuring consistency in clinical decision-making and treatment adjustments. Standardized protocols were used for clinical evaluations, diagnostic investigations (including neuroimaging and EEG acquisition and interpretation), and data collection using a unified case report form. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committees of the two participating centers: Taizhou Maternal and Child Health Hospital (No. 2026L-05-127) and The Second Affiliated Hospital of Zhejiang University School of Medicine (No. 2026-0730). The requirement for informed consent was waived by both committees because the study posed minimal risk, used anonymized data from routine clinical practice, and could not be practically conducted without the waiver.

Participants

All children with a confirmed diagnosis of epilepsy who received regular follow-up at the two tertiary pediatric epilepsy centers during the study period were considered for inclusion. Eligibility criteria were: (I) diagnosis of epilepsy according to the ILAE criteria (13); (II) availability of complete medical records containing demographic, clinical, electroencephalographic, neuroimaging, treatment‑related, and comorbidity information (detailed variable definitions are provided in the “Variables and data sources” section below); and (III) a minimum follow-up duration of 6 months after diagnosis. In this cohort, the age at onset ranged from 0.2 to 17.2 years. Exclusion criteria were acute symptomatic seizures (e.g., febrile seizures, metabolic derangements, or acute central nervous system infections without subsequent epilepsy diagnosis), progressive neurological disorders (e.g., neurodegenerative diseases), and incomplete outcome data for DRE classification. No matching or stratification was applied in participant selection. These criteria were designed to ensure a well‑defined and representative cohort of children with epilepsy, suitable for identifying clinical factors associated with DRE.

Variables and data sources

The primary outcome was DRE status, defined according to the ILAE consensus definition, which requires failure of at least two adequately chosen and dosed ASM schedules to achieve sustained seizure freedom. The variables used to establish DRE status included: the number of ASMs tried, ASM names and dosages, treatment duration, documented treatment failure (due to lack of efficacy or tolerability), and seizure frequency and patterns over time. Exposures and independently associated variables included age at onset, sex, seizure type (categorized as focal, generalized, mixed, or unknown), baseline seizure frequency (seizures per month during the 3 months before enrollment), history of status epilepticus, history of febrile seizures, etiology (classified per ILAE categories as structural, genetic, infectious, metabolic, immune, or unknown), presence of MRI‑detected structural lesions, presence of EEG epileptiform discharges, neurodevelopmental disorder (intellectual disability or developmental delay), psychiatric comorbidity, and medication adherence (good vs. poor, based on medical records and refill patterns). Treatment‑related variables—including the number of ASMs tried, maximum number of concurrent ASMs, polytherapy at last follow‑up, and prior epilepsy surgery or neurostimulation—were collected for descriptive purposes only. They were not included as factors in the multivariable regression model, as the number of ASMs tried is inherently linked to the ILAE definition of DRE (failure of ≥2 ASMs), and polytherapy at last follow‑up largely represents a consequence rather than a cause of DRE. All variables were extracted from electronic medical records using a standardized case report form by two trained research assistants independently, and any disagreements were adjudicated by a third senior investigator. The same assessment methods were applied across all participants.

Bias

Several measures were implemented to minimize potential bias. Selection bias was addressed by enrolling consecutive eligible patients from both centers without preselecting based on outcome status. Information bias was mitigated by using standardized data extraction forms and predefined variable definitions. To minimize outcome misclassification, DRE status was classified using the ILAE criteria whenever possible (13), based on documented medication trials and seizure outcomes from medical records. Although this approach improves standardization, it cannot completely eliminate misclassification when records are incomplete.

Study size

No formal sample size calculation was performed a priori, as this was an exploratory retrospective analysis using available data. The final sample size (n=412, including 100 DRE cases) was determined by the number of eligible patients meeting inclusion criteria during the study period. In the multivariable logistic regression model, the events‑per‑variable (EPV) ratio was approximately 7.7 (100 events/13 variables), which is within an acceptable range for logistic regression analysis.

Statistical analysis

All analyses were performed using R software (version 4.3.0). Continuous variables were expressed as mean ± standard deviation (SD) or median [interquartile range (IQR)] as appropriate. Normally distributed variables were compared using independent samples t-tests; non-normally distributed and discrete count variables were compared using the Mann-Whitney U test. Continuous forms were retained in regression analyses to preserve information and statistical power, and to avoid bias associated with categorization. The linearity assumption for continuous factors in the logistic regression model was assessed by including quadratic terms and visually inspecting the relationship between each continuous predictor and the logit of the outcome; all continuous factors satisfied this assumption. No categorization was applied to continuous factors in the regression models. Categorical variables were presented as frequencies (percentages) and compared using Pearson’s Chi-squared test or Fisher’s exact test as appropriate. Univariate logistic regression was conducted for each candidate variable. Variables with a P value <0.10 in univariate analysis, together with clinically relevant factors (age, sex, MRI lesion), were entered into a multivariate logistic regression model using the enter method. Multivariable logistic regression was used to adjust for potential confounders, including age at onset, sex, seizure type, baseline seizure frequency, history of status epilepticus, history of febrile seizure, etiology category, MRI lesion, EEG epileptiform discharges, neurodevelopmental disorder, and psychiatric comorbidity. Treatment‑related variables (number of ASMs tried, maximum concurrent ASMs, polytherapy at last follow‑up, and prior surgery/neurostimulation) were excluded from the multivariable model to avoid circular reasoning and reverse causation, as these variables are either part of the DRE definition or consequences of the disease course. Multicollinearity was assessed by variance inflation factor (VIF), with VIF >5 indicating significant collinearity. Model calibration was evaluated by the Hosmer-Lemeshow goodness-of-fit test. Discriminative ability was quantified by the area under the receiver operating characteristic (ROC) curve (AUC). The optimal cut-off value was determined using the Youden index (maximizing sensitivity + specificity − 1). Sensitivity, specificity, accuracy, and positive predictive value (PPV) were calculated at this cut-off. To provide robust estimates of precision, 95% confidence intervals (CIs) for performance metrics were derived by bootstrap resampling with 2,000 replications. Missing values were primarily confined to baseline seizure frequency, seizure type, and etiology category, with no missing data for DRE_status, age at onset, sex, MRI_lesion, EEG_epileptiform, neurodevelopmental disorder, psychiatric comorbidity, or treatment-related variables. The overall proportion of missing values was <5%, and missingness appeared to be random with respect to DRE status and other key clinical variables. Therefore, complete-case analysis was performed. No subgroup analyses or sensitivity analyses were prespecified. A two-sided P value <0.05 was considered statistically significant for all analyses.

Results

Participant characteristics

A total of 412 children with epilepsy were included in the analysis, of whom 100 (24.3%) were classified as having DRE according to the ILAE criteria. The clinical and demographic characteristics of the DRE and non‑DRE groups are summarized in Table 1. Children with DRE was a significantly younger age at onset compared with those without DRE (4.30±3.40 vs. 6.01±3.50 years, P<0.001). Sex distribution did not differ significantly between the two groups. Seizure type distribution was markedly different (P<0.001): mixed seizure types were more frequent in the DRE group (17.0% vs. 4.5%), whereas generalized seizures were less common in DRE (7.0% vs. 17.6%). Baseline seizure frequency was significantly higher in the DRE group {median [IQR]: 1 [0–10] vs. 0 [0–1] per month, P<0.001}. A history of status epilepticus was more prevalent among DRE patients (25.0% vs. 7.1%, P<0.001), while the frequency of febrile seizures did not differ between groups.

Table 1. Clinical and demographic characteristics of study participants by DRE status.

Characteristics	Non-DRE (n=312)	DRE (n=100)	P value
Age at onset (years)	6.01±3.50	4.30±3.40	<0.001†
Female	133 (42.6)	50 (50.0)	0.24†
Seizure type			<0.001†
Focal	243 (77.9)	75 (75.0)
Generalized	55 (17.6)	7 (7.0)
Mixed	14 (4.5)	17 (17.0)
Unknown	0 (0.0)	1 (1.0)
Seizure frequency at baseline (per month)	0 [0–1]	1 [0–10]	0.001‡
History of status epilepticus	22 (7.1)	25 (25.0)	<0.001‡
History of febrile seizure	59 (19.0)	24 (24.0)	0.28‡
Etiology category			<0.001†
Genetic	17 (5.4)	17 (17.0)
Infectious	1 (0.3)	1 (1.0)
Other/unknown	257 (82.4)	64 (64.0)
Structural	37 (11.9)	18 (18.0)
MRI lesion present	103 (33.0)	44 (44.0)	0.046‡
EEG epileptiform discharges	299 (95.8)	93 (93.0)	0.32‡
Neurodevelopmental disorder	34 (10.9)	43 (43.0)	<0.001‡
Psychiatric comorbidity	12 (3.8)	11 (11.0)	0.01‡

Data are presented as mean ± SD or median [IQR] for continuous variables and n (%) for categorical variables. ^†, Pearson’s Chi‑squared test (or Fisher’s exact test for variables with expected frequencies <5). ^‡, Mann-Whitney U test. DRE, drug-resistant epilepsy; EEG, electroencephalography; IQR, interquartile range; MRI, magnetic resonance imaging; SD, standard deviation.

Regarding etiology, there was a significant overall difference between groups (P<0.001). Genetic etiologies were more common in the DRE group (17.0% vs. 5.4%), whereas the proportion of “other/unknown” etiology was lower (64.0% vs. 82.4%). Structural lesions on MRI were present in 44.0% of DRE patients compared with 33.0% of non-DRE patients (P=0.046). The presence of EEG epileptiform discharges did not differ significantly (93.0% vs. 95.8%, P=0.32). Comorbidities were more frequent in the DRE group: neurodevelopmental disorder (43.0% vs. 10.9%, P<0.001) and psychiatric comorbidity (11.0% vs. 3.8%, P=0.01).

As expected, given the definition of DRE, patients in the DRE group had tried a greater number of ASMs {median [IQR]: 5 [3–6] vs. 1 [1–2], P<0.001}, used a higher maximum number of concurrent ASMs {median [IQR]: 3 [3–4] vs. 1 [1–2], P<0.001}, and were almost universally on polytherapy at last follow‑up (68.0% vs. 23.0%, P<0.001). These treatment‑related differences are inherent to the DRE definition and clinical course; they were not included as factors in the multivariable model (Table 2).

Table 2. Treatment-related characteristics of study participants by DRE status.

Characteristics	Non-DRE (n=312)	DRE (n=100)	P value
Number of ASMs tried	1 [1–2]	5 [3–6]	<0.001‡
Maximum concurrent ASMs	1 [1–2]	3 [3–4]	<0.001‡
Polytherapy at last follow-up	72 (23.0)	68 (68.0)	<0.001‡
Poor adherence	6 (2.0)	4 (4.0)	0.42†
Prior surgery/neurostimulation	3 (1.0)	4 (4.0)	0.02‡

Data are presented as median [IQR] or n (%). These variables are presented for descriptive purposes only and were not included as factors in the multivariable regression model due to their intrinsic relationship with the DRE definition and clinical consequences. ^†, Pearson’s Chi‑squared test (or Fisher’s exact test for variables with expected frequencies <5). ^‡, Mann-Whitney U test. ASM, anti-seizure medication; DRE, drug-resistant epilepsy; IQR, interquartile range.

Univariate logistic regression

Table 3 presents the results of univariate logistic regression for DRE. The largest effect estimates were observed for neurodevelopmental disorder [odds ratio (OR) =6.37; 95% CI: 3.44–11.91; P<0.001], history of status epilepticus (OR =4.24; 95% CI: 2.16–8.31; P<0.001), and mixed seizure type (vs. focal) (OR =28.61; 95% CI: 4.98–539.34; P=0.002). Unknown etiology (vs. genetic) was associated with lower odds of DRE (OR =0.39; 95% CI: 0.22–0.69; P=0.001), indicating that genetic etiology was associated with higher odds of DRE compared with unknown etiology. This finding is consistent with the higher proportion of genetic etiologies observed in the DRE group (17.0% vs. 5.4% in the non-DRE group). Generalized seizure type (vs. focal) was also associated with lower odds (OR =0.22; 95% CI: 0.05–0.62; P=0.01). Other factors showing borderline significance included psychiatric comorbidity (OR =2.44; 95% CI: 0.93–6.04; P=0.057), baseline seizure frequency (OR =1.05; 95% CI: 1.01–1.12; P=0.040), history of febrile seizure (OR =1.74; 95% CI: 0.95–3.10; P=0.07), and MRI lesion (OR =1.61; 95% CI: 0.96–2.69; P=0.07).

Table 3. Univariate logistic regression analysis for DRE (DRE as positive outcome).

Variables	OR (95% CI)	P value
Age at onset (per 1-year increase)	0.84 (0.77–0.91)	<0.001
Sex (female vs. male)	1.32 (0.79–2.20)	0.28
Seizure type
Generalized vs. focal	0.22 (0.05–0.62)	0.01
Mixed vs. focal	28.61 (4.98–539.34)	0.002
Seizure frequency at baseline (per 1/month increase)	1.05 (1.01–1.12)	0.040
History of status epilepticus	4.24 (2.16–8.31)	<0.001
History of febrile seizure	1.74 (0.95–3.10)	0.07
Etiology
Structural vs. genetic	0.89 (0.37–1.93)	0.78
Unknown vs. genetic	0.39 (0.22–0.69)	0.001
MRI lesion present	1.61 (0.96–2.69)	0.07
EEG epileptiform discharges	0.71 (0.20–3.29)	0.62
Neurodevelopmental disorder	6.37 (3.44–11.91)	<0.001
Psychiatric comorbidity	2.44 (0.93–6.04)	0.057

Treatment-related variables (number of ASMs tried, maximum concurrent ASMs, polytherapy at last follow-up, and prior surgery/neurostimulation) were excluded from the univariate and multivariable analyses due to their intrinsic relationship with the DRE definition and clinical consequences. DRE was used as the positive outcome (coded as 1) in the logistic regression model; ORs represent the odds of DRE vs. non‑DRE. ASM, anti-seizure medication; CI, confidence interval; DRE, drug-resistant epilepsy; EEG, electroencephalography; MRI, magnetic resonance imaging; OR, odds ratio.

Multivariate logistic regression

After adjusting for potential confounders (age, sex, seizure type, baseline seizure frequency, history of status epilepticus, history of febrile seizure, etiology category, MRI lesion, EEG epileptiform discharges, neurodevelopmental disorder, and psychiatric comorbidity), several factors remained independently associated with DRE (Table 4). Generalized seizure type (compared with focal) was associated with significantly lower odds of DRE (adjusted OR =0.20; 95% CI: 0.04–0.62; P=0.01). History of status epilepticus (adjusted OR =4.86; 95% CI: 2.15–11.13; P<0.001), neurodevelopmental disorder (adjusted OR =3.40; 95% CI: 1.49–7.85; P=0.004), and older age at onset (adjusted OR =0.91; 95% CI: 0.82–1.00; P=0.046) were also independently associated with DRE. Structural etiology (compared with genetic) showed lower odds (adjusted OR =0.25; 95% CI: 0.07–0.92; P=0.040). Treatment‑related variables were excluded from the multivariable model as prespecified.

Table 4. Multivariate logistic regression analysis of factors independently associated with DRE (DRE as positive outcome).

Variables	Adjusted OR (95% CI)	P value
Age at onset (per 1-year increase)	0.91 (0.82–1.00)	0.046
Sex (female vs. male)	1.01 (0.54–1.85)	0.98
Seizure type
Generalized vs. focal	0.20 (0.04–0.62)	0.01
Mixed vs. focal	9.06 (1.13–195.55)	0.07
Seizure frequency at baseline (per 1/month increase)	1.04 (0.98–1.12)	0.17
History of status epilepticus	4.86 (2.15–11.13)	<0.001
History of febrile seizure	0.59 (0.25–1.29)	0.20
Etiology
Structural vs. genetic	0.25 (0.07–0.92)	0.040
Unknown vs. genetic	0.42 (0.16–1.13)	0.08
MRI lesion present	1.08 (0.52–2.16)	0.83
EEG epileptiform discharges	0.85 (0.22–3.28)	0.82
Neurodevelopmental disorder	3.40 (1.49–7.85)	0.004
Psychiatric comorbidity	1.84 (0.57–5.48)	0.29

Model adjusted for: age, sex, seizure type, baseline seizure frequency, history of status epilepticus, history of febrile seizure, etiology category, MRI lesion, EEG epileptiform discharges, neurodevelopmental disorder, and psychiatric comorbidity. Treatment-related variables were excluded from the multivariable model to avoid circular reasoning and reverse causation. DRE was used as the positive outcome (coded as 1) in the logistic regression model; ORs represent the odds of DRE vs. non‑DRE. CI, confidence interval; DRE, drug-resistant epilepsy; EEG, electroencephalography; MRI, magnetic resonance imaging; OR, odds ratio.

Model performance

The final multivariate model demonstrated moderate discriminative ability, as shown in Table 5 and Figure 1. The AUC was 0.802 (95% CI: 0.751–0.872). Using the optimal cut-off value of 0.139, the model achieved an accuracy of 69.4% (95% CI: 66.6–86.6), a sensitivity of 82.9% (95% CI: 54.0–90.1%), a specificity of 65.7% (95% CI: 61.0–93.5%), and a PPV of 39.4% (95% CI: 35.6–71.9%).

Table 5. Model performance for predicting DRE.

Metric	Value	95% CI
AUC	0.802	0.751–0.872
Optimal cut-off value	0.139	–
Accuracy (%)	69.4	66.6–86.6
Sensitivity (%)	82.9	54.0–90.1
Specificity (%)	65.7	61.0–93.5
PPV (%)	39.4	35.6–71.9

AUC, area under the ROC curve; CI, confidence interval; DRE, drug-resistant epilepsy; PPV, positive predictive value; ROC, receiver operating characteristic.

Figure 1.

ROC curve of the multivariate logistic regression model for predicting DRE. The shaded area represents the 95% CI. The optimal cut‑off point (determined by the Youden index) is marked, with corresponding sensitivity and specificity of 82.9% and 65.7%, respectively. AUC =0.802 (95% CI: 0.751–0.872). The ROC curve was derived from internal validation using bootstrap resampling (2,000 replications). AUC, area under the ROC curve; CI, confidence interval; DRE, drug‑resistant epilepsy; ROC, receiver operating characteristic.

Discussion

In this retrospective cohort study of 412 children with epilepsy, we identified several factors independently associated with DRE. Generalized seizure type (vs. focal) was associated with significantly lower odds of DRE (adjusted OR =0.20; P=0.01). History of status epilepticus (adjusted OR =4.86; P<0.001), neurodevelopmental disorder (adjusted OR =3.40; P=0.004), and older age at onset (adjusted OR =0.91; P=0.046) were independently associated with increased and decreased odds of DRE, respectively. Structural etiology (vs. genetic) also showed lower odds (adjusted OR =0.25; P=0.040). The multivariate model demonstrated moderate discriminative ability (AUC =0.802; 95% CI: 0.751–0.872).

The inverse association between generalized seizure type and DRE is consistent with the known heterogeneity of drug responsiveness across epilepsy types. Focal epilepsies, particularly those associated with structural brain lesions such as focal cortical dysplasia, hippocampal sclerosis, and low‑grade tumors, are well recognized to carry a higher risk of drug resistance (9). In contrast, many generalized epilepsies, especially idiopathic generalized epilepsies, respond favorably to broad‑spectrum ASMs. This differential responsiveness likely explains why generalized seizure type was associated with lower odds of DRE in our model. However, it is important to note that certain generalized epilepsy syndromes that may follow a fluctuating course—with some patients achieving initial seizure control before later developing drug resistance—can also become refractory, and the observed association should not lead to complacency in managing generalized epilepsies (8).

Mixed seizure type showed a trend toward higher odds of DRE (adjusted OR =9.06; P=0.07), although this did not reach statistical significance. Children with mixed seizure types often present with more severe and complex epilepsy phenotypes, including multiple semiologies and broader epileptogenic networks, which may contribute to a higher likelihood of pharmacoresistance (9). The wide CI (1.13–195.55) reflects the relatively small number of patients with mixed seizure types in our cohort (n=29) and should be interpreted with caution.

History of status epilepticus emerged as one of the strongest independent factors associated with DRE in our cohort (adjusted OR =4.86; P<0.001). This finding is consistent with previous studies demonstrating that status epilepticus, particularly when it occurs early in the course of epilepsy, may reflect a more severe underlying epileptogenic process and is associated with an increased likelihood of pharmacoresistance (9). The biological plausibility is supported by experimental evidence suggesting that prolonged seizures can induce alterations in drug transporter expression and receptor subunit composition at the blood–brain barrier, potentially reducing ASM efficacy over time. Clinically, a history of status epilepticus should prompt closer monitoring and earlier consideration of non‑pharmacological options if initial ASM trials fail.

Neurodevelopmental disorder was strongly and independently associated with DRE (adjusted OR =3.40; P=0.004). This finding is consistent with established literature demonstrating that children with intellectual disability or developmental delay are at substantially higher risk of drug resistance (9). Several mechanisms may explain this association. First, children with neurodevelopmental disorders often present with more severe epilepsy phenotypes, higher baseline seizure burden, and a greater likelihood of structural brain abnormalities. Second, neurodevelopmental comorbidities may complicate medication management through challenges in adherence, polytherapy due to multiple seizure types, and higher susceptibility to adverse effects that limit dose escalation. Third, the underlying genetic or structural causes of neurodevelopmental disorders may directly contribute to epileptogenicity and pharmacoresistance. The substantial burden of neurodevelopmental comorbidity in DRE is further illustrated by population-based data showing that 86% of children with early-onset DRE have global developmental delay or intellectual disability (5), and that DRE occurs in up to 77.7% of children with developmental and epileptic encephalopathies (14). These factors collectively underscore the importance of early, comprehensive evaluation in this vulnerable population.

In our cohort, structural etiology was associated with significantly lower odds of DRE compared with genetic etiology (adjusted OR =0.25; P=0.040), while unknown etiology showed a trend toward lower odds (adjusted OR =0.42; P=0.08). This finding is somewhat unexpected, as structural epilepsies—particularly those associated with malformations of cortical development, hippocampal sclerosis, and tumors—are generally considered to carry a high risk of drug resistance (9). Indeed, population-based studies have consistently reported that structural factors are the most common etiology in pediatric DRE, accounting for 23.4% to 29% of cases, followed by genetic factors at 19.2% (5,15,16). First-drug efficacy is significantly lower, and DRE rates are higher in children with structural etiologies compared with those with unknown etiologies (17). However, several factors may explain the seemingly lower odds of DRE for structural vs. genetic etiologies in our cohort. First, the genetic etiology group in our cohort included a substantial proportion of children with severe developmental and epileptic encephalopathies (e.g., SCN1A-related Dravet syndrome, KCNQ2 encephalopathy), which are notoriously drug-resistant. A recent study of early-onset epilepsy found that definitive genetic diagnoses were established in 71.4% of DRE cases vs. only 37.5% of non-refractory cases (18). Genetic causes have also been associated with an increased risk of poor prognosis compared with structural causes (OR =4.754; P=0.053) (19). Second, some patients with structural etiologies in our cohort had undergone epilepsy surgery, which may have improved seizure outcomes and reduced the likelihood of DRE classification. Third, the relatively small sample size in the genetic and structural subgroups may have affected the stability of the estimates. Furthermore, the age of onset differs significantly between etiological categories—children with genetic causes are a significantly earlier age of onset than those with structural etiologies (P<0.001) (15), and early onset is itself an independent risk factor for DRE. These findings highlight the heterogeneity within each etiological category and underscore the need for more granular etiological subtyping in future studies.

Older age at onset was independently associated with lower odds of DRE (adjusted OR =0.91 per 1‑year increase; P=0.046). This finding is consistent with the well‑established observation that early‑onset epilepsies are often more severe and more likely to be drug‑resistant (7,14,20). Epilepsies with onset in infancy or early childhood are frequently associated with structural brain abnormalities, genetic mutations, or developmental encephalopathies that confer a higher risk of pharmacoresistance (5,9). In contrast, epilepsies with later onset in childhood are more likely to be self‑limited or more responsive to ASMs. While the effect size is modest, this factor may contribute to risk stratification when combined with other clinical variables.

The model demonstrated moderate discriminative ability (AUC =0.802; 95% CI: 0.751–0.872), which is comparable to several previously published DRE prediction models that used only baseline clinical variables. A systematic review by Wang et al. reported that most existing models had AUCs ranging from 0.70 to 0.85 (9). The sensitivity and specificity at the optimal cut‑off were 82.9% and 65.7%, respectively, suggesting that the model may be more suitable for screening (identifying potential DRE cases) than for definitive diagnosis. The relatively wide CIs for sensitivity (54.0–90.1%) and specificity (61.0–93.5%) reflect the limited number of DRE cases (n=97) and should be interpreted with caution. Notably, the AUC dropped substantially compared with models that included treatment‑related variables, underscoring the critical importance of methodological rigor in DRE prediction research. Many previously published models may have overestimated their predictive performance by including variables intrinsically linked to the outcome definition—such as the number of failed ASM trials—rather than true baseline factors (20).

Our findings, together with the ILAE definition of DRE, support a practical approach to clinical management. For any child who fails two adequately chosen and dosed ASM regimens, the diagnosis of DRE is established by definition, and the likelihood of achieving seizure freedom with additional medication trials is low. This is a well‑established principle in epilepsy management, supported by landmark studies demonstrating that patients who fail two ASM schedules have less than a 5% chance of becoming seizure‑free with subsequent medications (21). Therefore, in line with current ILAE guidelines and established literature, after failure of two appropriate ASM trials, clinicians should consider non-pharmacological options, including epilepsy surgery evaluation, ketogenic diet, or neuromodulation, depending on the individual patient’s epilepsy type, etiology, and surgical candidacy (20).

The factors independently associated with DRE identified in this study—particularly seizure type, history of status epilepticus, neurodevelopmental disorder, and age at onset—may help clinicians recognize children who are more likely to follow a drug‑resistant trajectory early in their disease course. However, these factors should be used to inform clinical judgment and monitoring intensity, rather than as a standalone decision‑making tool. The diagnosis of DRE itself remains the most practical and evidence‑based indicator for initiating non‑pharmacological evaluation (20).

This study has several limitations that should be acknowledged. Most importantly, a fundamental limitation of this study—and of many DRE prediction studies—is the inherent circularity between the outcome and certain treatment‑related variables. The ILAE definition of DRE itself requires failure of at least two ASM regimens; therefore, the number of ASMs tried is not an independent predictor but rather a diagnostic criterion. Similarly, polytherapy at last follow‑up is largely a consequence rather than a cause of DRE. To avoid circular reasoning and reverse causation, we excluded these treatment‑related variables from our multivariable model. This design choice, while methodologically necessary, substantially reduced the model’s discriminative ability (AUC from 0.956 to 0.802), highlighting that previously published models may have overestimated their performance by including definition‑overlapping variables. Future prospective studies with baseline‑only factors are needed to identify true independent risk factors for DRE. Second, the retrospective design introduces the potential for selection and information bias, despite our efforts to use standardized data extraction and predefined definitions. Third, the sample size, although adequate for the number of variables included (event per variable ratio of approximately 7.7), may not have been sufficient to detect smaller effect sizes or interactions. The relatively wide CIs for sensitivity (54.0–90.1%) and specificity (61.0–93.5%) reflect the limited number of DRE cases and should be interpreted with caution. Fourth, we did not perform external validation; the model’s moderate performance may be overfitted to our cohort, and generalizability to other populations remains unknown. Fifth, we could not account for several potentially important confounders, including ASM dosing, serum drug levels, adherence over time, and specific epilepsy syndromes, due to incomplete documentation. Sixth, the study was conducted at two tertiary referral centers, which may have introduced referral bias (more severe or complex cases), limiting the generalizability of our findings to community‑based pediatric epilepsy populations. Seventh, the classification of etiology as “unknown” in a large proportion of patients (64.0% in the DRE group and 82.4% in the non‑DRE group) reflects the diagnostic limitations in clinical practice but may have diluted the effect of specific etiologies. Finally, we emphasize that our findings should be interpreted as retrospective associations, not as predictive or causal relationships. External validation in prospectively designed cohorts with baseline‑only factors is essential before any clinical application of this model. Additionally, the number of patients with specific epilepsy syndromes (e.g., juvenile myoclonic epilepsy, Lennox‑Gastaut syndrome) was extremely limited in our cohort, and these diagnoses were not systematically captured in a dedicated data field, precluding reliable subgroup‑specific model performance analyses. While DEEs were identifiable through genetic and clinical notes, the sample size remained too small for meaningful stratified analyses. Our findings regarding seizure type should therefore be interpreted at the level of broad seizure type categories rather than specific epilepsy syndromes. Future large-scale, multicenter studies with sufficient sample sizes are needed to evaluate whether prediction models perform differently across distinct epilepsy syndromes. Additionally, specific types of epilepsy surgery or neuromodulation procedures were not systematically recorded due to the retrospective nature of the dataset and could not be analyzed. Neurodevelopmental and psychiatric comorbidities were based on clinical diagnoses documented in medical records, as standardized assessment tools were not routinely used. This may have contributed to variability in the classification of these comorbidities.

Prospective multicenter studies with external validation are urgently needed to confirm our findings and to identify true baseline factors associated with DRE. Future research should focus on: (I) prospective cohort studies that collect all potential predictors before ASM initiation, avoiding the inclusion of variables that overlap with the outcome definition; (II) the role of genetic biomarkers, advanced neuroimaging [e.g., positron emission tomography (PET), magnetoencephalography (MEG)], and quantitative EEG features in informing DRE risk assessment; (III) the development and validation of clinical risk scores using baseline-only variables, with rigorous methodological standards to avoid circularity; and (IV) randomized controlled trials comparing early surgical intervention vs. continued medical therapy in children who have already failed two ASM schedules, which would provide the highest level of evidence to guide clinical practice (20).

Conclusions

In this retrospective cohort of 412 children with epilepsy, several factors were identified as independently associated with DRE after excluding treatment-related variables that are definitionally or consequentially linked to the outcome. Generalized seizure type (vs. focal), structural etiology (vs. genetic), and older age at onset were associated with lower odds of DRE, whereas history of status epilepticus and neurodevelopmental disorder were associated with higher odds. The multivariate model demonstrated moderate discriminative ability (AUC =0.802), substantially lower than models that included treatment-related variables, underscoring a critical methodological concern: the strong associations of ASM number and polytherapy with DRE largely reflect the diagnostic criteria and clinical consequences of DRE rather than true predictive effects. The diagnosis of DRE itself—failure of two adequate ASM trials—remains the most practical and evidence-based indicator for considering non‑pharmacological options. Future prospective studies should focus on identifying baseline factors measured before treatment initiation, avoiding the inclusion of variables that overlap with the outcome definition. Our findings should not be interpreted as supporting a diagnostic or predictive model without external validation.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committees of the two participating centers: Taizhou Maternal and Child Health Hospital (No. 2026L-05-127) and The Second Affiliated Hospital of Zhejiang University School of Medicine (No. 2026-0730). The requirement for informed consent was waived by both committees because the study posed minimal risk, used anonymized data from routine clinical practice, and could not be practically conducted without the waiver.

Data Sharing Statement

Available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0593/dss

tp-15-9-376-dss.pdf ^{(71.3KB, pdf)}

DOI: 10.21037/tp-2026-0593