Anti-glutamic acid decarboxylase (GAD) antibody-associated neuroimmune disorders (GAD-ANDs) are autoimmune neurological conditions characterized by the presence of anti-GAD antibodies and typical clinical syndromes, including limbic encephalitis (LE), epilepsy (EP), stiff-person syndrome (SPS), autoimmune cerebellar ataxia (ACA), or overlapping phenotypes (1). Among these, GAD-EP and GAD-LE account for 52.9%-61.8% of GAD-ANDs, constituting the predominant clinical phenotypes (2, 3).

The association between anti-GAD antibodies and epilepsy was first reported in 1988 (4), with subsequent studies confirming their presence in patients with drug-resistant epilepsy by 1998 (5). Anti-GAD antibodies are present in 2.58% of all epilepsy patients (6), with higher prevalence in cryptogenic focal epilepsy (7.7%) (6), adult-onset temporal lobe epilepsy (TLE) (8.7%) (7), and refractory focal epilepsy (15.7%) (8). In 2020, the International League Against Epilepsy (ILAE) recognized GAD-EP as autoimmune-associated epilepsy, distinct from acute symptomatic seizures secondary to autoimmune encephalitis (9). Notably, patients with GAD-LE who experience seizures often progress to chronic drug-resistant epilepsy. Longitudinal studies suggest similar seizure outcomes between LE patients with seizures and isolated GAD-EP (3), supporting their combined analysis as a single clinical phenotype in prognostic modeling.

GAD-EP involves complex autoimmune mechanisms, including autoantibodies, B cells (10, 11), and T cells (12, 13). Cellular and serological parameters, such as lymphocyte counts and immunoglobulin levels, have been increasingly applied in neuroimmune disorders to guide individualized therapy. However, the relationship between peripheral immune cell counts or immunoglobulin levels and the prognosis of GAD-EP remains unclear.

Although 55.6%–90% of patients initially respond to immunotherapy and antiseizure medications (ASMs) (≥50% seizure frequency reduction), sustained seizure freedom is achieved in 0–36.7% (2, 3, 14–19), indicating that controlling seizures in these patients remains a major clinical challenge. Acute or subacute onset and early immunotherapy may improve outcomes, whereas focal or high-frequency seizures predict poor prognosis (19). While predictive models have been proposed for autoimmune encephalitis or epilepsy mediated by neuronal surface antibodies (20, 21), no such model currently exists for GAD-EP.

To address these gaps, we comprehensively collected demographic, clinical, immunological, neuroimaging, and electrophysiological data from patients with GAD-EP to identify prognostic factors and develop a logistic regression–based predictive model. The model supports risk-based patient stratification to inform individualized therapy.

This study included 91 patients with GAD-EP and aimed to identify prognostic factors and develop a predictive model. We found that focal seizures, TLE, and high seizure frequency were associated with poor outcomes, whereas a shorter disease duration, earlier initiation of immunotherapy, and higher CD8⁺T-cell levels were associated with seizure-freedom. TLE and seizure frequency emerged as independent predictors of poor prognosis. A prognostic model including these two variables showed good predictive performance. Based on this model, a risk scoring system was developed and demonstrated high accuracy in distinguishing low- and high-risk patients. Clinically, the score may aid early risk stratification, enabling timely use of standard immunotherapy in low-risk patients while prompting consideration of alternative or novel treatment strategies in high-risk patients rather than repeated application of ineffective regimens.

TLE is the most common subtype of GAD-EP, typically associated with poorer outcomes compared with non-GAD-related mesial TLE (14). Higher seizure frequency or number of seizures before remission are established predictors of poor seizure control, drug resistance, and recurrence in epilepsy patients (24–26). We developed a predictive model incorporating these factors, which demonstrated high sensitivity and excellent negative predictive value and accurately identifies patients likely to achieve seizure freedom while minimizing misclassification of those with poor outcomes.

In our cohort, 9 patients (10%) met the criteria for LE and 82 (90%) had GAD-EP, with no difference in seizure outcomes (p = 1.000). CSF pleocytosis tended to be associated with better outcomes, although this did not reach statistical significance (42% vs. 21%, p = 0.139). Including LE and CSF-inflammatory patients in the prognostic model maintained high accuracy (88.9% and 75.0%), supporting the combination of LE and EP for modeling. Elevated CSF WBC persisted up to 36 months, and disease duration ≤ 36 months was associated with better prognosis (p = 0.035), emphasizing that early diagnosis and treatment during the inflammatory phase—whether presenting as acute LE or chronic epilepsy—improves outcomes. We propose that patients with LE previously underwent earlier antibody testing and treatment, but growing recognition of GAD-EP now enables similarly timely diagnosis and therapy, reducing prognostic differences between the two presentations.

We found that higher peripheral CD8⁺T-cell counts at initial presentation were associated with favorable prognosis, suggesting they may reflect disease immune activity. Higher counts may indicate an active phase, whereas lower levels could reflect the chronic stage. Patients with disease duration ≤36 months had higher counts than those with longer disease (702.2 ± 404.5 vs. 529.3 ± 234.6 cells/µL), although the difference was not statistically significant (p = 0.141). These findings support the potential of CD8⁺T-cell counts as a biomarker, warranting larger studies with longitudinal monitoring to validate and define an optimal cutoff.

Brain MRI showed limited prognostic value in our cohort. Although patients without hippocampal atrophy or hypometabolism had higher rates of seizure freedom, the difference was not statistically significant. Several factors may explain this: (1) limited sample size, reducing statistical power; (2) heterogeneous timing of imaging, as patients were scanned at different disease stages, introducing variability and highlighting the need for standardized assessments at fixed time points in future studies; (3) limited sensitivity of conventional MRI, since subtle hippocampal or functional damage in GAD-EP may precede visible atrophy. Previous studies have shown more extensive hypometabolism on FDG-PET in GAD-TLE compared with hippocampal sclerosis (27), suggesting that FDG-PET may offer greater sensitivity; (4) GAD-EP involves a complex interplay between structural injury and immune-inflammatory processes, with brain inflammation potentially driving seizure persistence beyond structural damage, although noninvasive current methods to assess tissue inflammation remain limited (28).

Management of GAD-EP remains challenging, with conflicting immunotherapy evidence: some studies report benefits from early corticosteroids or IVIG (19, 29), while others show limited effects (15, 30). In our study, different immunotherapy regimens showed no significant differences in seizure freedom. To address indication bias—where severe cases receive intensive treatment—we applied inverse probability of treatment weighting (IPTW) with logistic regression, adjusting for age, disease duration, clinical phenotype, TLE, seizure frequency, imaging atrophy, CD8⁺T-cell counts, and ASMs number. After adjustment, combination first-line therapy showed no superiority over monotherapy (OR = 0.42, 95% CI: 0.18–0.93, p < 0.05), and rituximab for inadequate first-line response showed no benefit (OR = 0.70, 95% CI: 0.33–1.45, p > 0.05). These findings warrant cautious interpretation given potential unmeasured confounders (economic factors, HLA background,etc.). Symptomatic treatments like cenobamate-clobazam show modest benefit (31), while vagus nerve stimulation and epilepsy surgery show limited efficacy (32). Randomized controlled trials are needed for definitive immunotherapy comparisons. Novel therapeutic approaches are urgently required.

This study has several limitations. First, the retrospective design may introduce information bias. Second, despite representing a relatively large East Asian GAD-EP cohort, only 22 patients (24.2%) achieved seizure freedom, limiting statistical power. Following the 10 events-per-variable principle, we included two predictors, constraining model complexity. Consequently, the model generates four discrete probability values with limited discrimination for intermediate-risk patients. Third, although bootstrap validation addresses optimism bias, external validation is lacking. Ongoing patient enrollment may identify additional predictors to improve risk stratification. Currently, this model serves as a preliminary screening tool requiring validation in independent cohorts.

In conclusion, we analyzed patients with GAD-EP from multiple centers and evaluated multidimensional predictive variables. Focal seizures, TLE, and high seizure frequency were associated with poor outcomes, while early diagnosis and treatment and higher CD8⁺T-cell counts predicted favorable prognosis. TLE and seizure frequency were identified as independent predictors, forming a prognostic model with good discriminative performance. The resulting risk-scoring system effectively distinguishes high- and low-risk patients, providing a practical tool for clinical risk stratification and treatment decision-making in GAD-EP.