This study was designed as a retrospective secondary analysis based on a publicly available dataset. The original data were derived from a study published by Liu et al. (8), which aimed to construct a prediction model for secondary epilepsy within 1 year in patients with AIS. According to the original authors’ data availability statement, the complete and de-identified dataset is deposited in the Dryad Digital Repository (22). The dataset aggregates comprehensive clinical data from AIS patients across multiple medical centers in Chongqing, China, spanning from 2013 to 2022. The reliability and quality of the data have been confirmed in subsequent independent validation studies (23). Building upon this high-quality data source, we included a total of 21,459 patients from the original cohort for analysis (screening process shown in Figure 1). During the screening, we strictly adhered to the criteria of the original study: inclusion criteria encompassed a diagnosis of AIS confirmed by head CT or MRI, with complete follow-up records and relevant clinical data; exclusion criteria included a history of epilepsy, hemorrhagic transformation, other non-ischemic brain diseases, and missing key research variables.

Regarding ethical compliance, as this study utilized only publicly available and fully de-identified secondary data, it was exempt from ethical review and patient informed consent requirements in accordance with relevant regulations and the principles of the Declaration of Helsinki. Data acquisition and processing strictly complied with the terms of use and data sharing agreements of the Dryad platform.

This study is reported in strict accordance with the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis) statement (24). Furthermore, the risk of bias and model applicability were rigorously evaluated using the PROBAST (Prediction model Risk Of Bias ASessment Tool) framework (25). The completed TRIPOD checklist and the detailed PROBAST assessment are provided in Supplementary Tables S1, S2, respectively.

To comprehensively assess patient clinical characteristics, we extracted detailed baseline variables from the original dataset. Demographic characteristics included age and gender; medical history and comorbidities covered hypertension, diabetes mellitus, coronary heart disease, atrial fibrillation, hyperlipidemia, and fatty liver. Regarding stroke clinical characteristics and complications, admission NIHSS scores, cerebral herniation, and hydrocephalus were recorded. Neuroimaging features detailed the anatomical location of the infarct, including the frontal lobe, temporal lobe, parietal lobe, occipital lobe, basal ganglia, and brainstem. To validate anatomical risk, involvement of any region among the frontal, temporal, parietal, or occipital lobes was defined as “cortical involvement.” Laboratory indicators extracted included routine blood tests (white blood cell count, platelet count), biochemical markers (CRP, serum albumin, blood lactate, glycosylated hemoglobin, creatinine), coagulation markers (D-dimer), and electrolytes (sodium, calcium) collected at admission. To quantify the “Two-Hit” effect, we calculated two core composite indices based on raw laboratory metrics: the Immuno-Nutritional Index was defined as the ratio of CRP (mg/L) to serum albumin (g/L) (CAR); the Hypoxia-Nutritional Index was defined as the ratio of blood lactate (mmol/L) to serum albumin (g/L) (LAR). The primary endpoint of this study was defined as secondary epilepsy occurring within 1 year of follow-up.

For baseline characteristics, continuous variables with non-normal distribution were expressed as medians (interquartile range [IQR]) and compared between groups using the Mann–Whitney U test; categorical variables were expressed as frequencies (percentages) and compared using the Chi-square test. Given the narrow numerical ranges and skewed distributions of CAR and LAR, Z-score standardization was applied to these metrics to enhance the interpretability and numerical stability of regression coefficients. Consequently, odds ratios (OR) in all subsequent regression models represent the change in risk per 1-standard deviation (SD) increase in the index.

To rigorously validate the selection of biomarkers and exclude potential post hoc optimization bias, we employed a dual-method feature selection strategy prior to multivariable modeling: (1) Least Absolute Shrinkage and Selection Operator (LASSO) regression was utilized to penalize model complexity and identify non-redundant predictors with non-zero coefficients; and (2) Random Forest (RF) analysis was performed to rank variables based on the Mean Decrease in Gini impurity to assess their predictive importance. All clinical variables showing statistical significance (p < 0.05) in the univariate analysis were entered into these screening models.

We employed multivariable logistic regression models to evaluate the independent associations of CAR and LAR with PSE. The multivariate models were adjusted for age, gender, NIHSS score, cortical involvement, and other potential confounders identified by the feature selection process. To explicitly address potential concerns regarding mathematical coupling arising from the shared denominator (albumin), we calculated the Variance Inflation Factor (VIF) to assess multicollinearity between CAR and LAR. Furthermore, sensitivity analyses were performed by comparing the “Two-Hit” model (utilizing ratio indices) against a “Decomposed Model” utilizing individual biomarkers (CRP, lactate, and albumin modeled as separate covariates) to validate that the ratio formulations captured synergistic information beyond their constituent components.

To explore non-linear relationships, restricted cubic spline (RCS) models were applied to analyze the dose–response relationship between CAR/LAR and epilepsy risk, with four knots specified. Additionally, Spearman correlation analysis was used to assess the correlation between CAR and LAR, and a heatmap of epilepsy incidence was constructed by grouping both indices into quartiles to visually validate the synergistic effect of the “Two-Hit” hypothesis.

To evaluate the incremental predictive value, three multivariate prediction models were constructed: a Baseline Model, a Single-Marker Model, and a Two-Hit Model. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC), along with the Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) to measure improvements in risk stratification. To ensure robust confidence intervals, NRI and IDI calculations utilized 1,000 bootstrap resampling iterations.

To rigorously assess model performance and correct for potential overfitting, we performed internal validation using the bootstrap method with 1,000 resamples. We calculated the optimism-corrected AUC, calibration slope, and Brier score. Clinical utility was measured via Decision Curve Analysis (DCA) to assess net benefit across different threshold probabilities. Finally, based on the final multivariate logistic regression model, we developed an HTML-based interactive web risk calculator to facilitate clinical translation.

All statistical analyses were performed using Python (version 3.11.5) and R (version 4.4.2), with a two-sided p-value of less than 0.05 considered statistically significant.

This study ultimately included 21,459 patients with AIS for analysis. During the one-year follow-up period, 936 patients (4.36%) developed PSE, while the remaining 20,523 patients did not. As shown in Table 1, compared to the non-epilepsy group, patients in the epilepsy group were slightly younger (median age: 66 vs. 67 years, p = 0.01) and exhibited more severe neurological deficits at admission (median NIHSS score: 11 vs. 7, p < 0.001). Regarding anatomical distribution, the epilepsy group demonstrated significantly higher rates of cortical involvement, specifically in the frontal (15.4% vs. 3.8%), temporal (11.6% vs. 2.6%), and parietal lobes (10.3% vs. 2.8%) (all p < 0.001). In terms of immuno-metabolic biomarkers, the epilepsy group presented a “Two-Hit” pattern characterized by concurrent systemic inflammation and metabolic stress. Specifically, white blood cell count, CRP, and lactate levels were significantly elevated, while albumin levels were significantly reduced in the epilepsy group (all p < 0.05). Consequently, the derived composite indices were markedly higher in epilepsy patients: the median Immuno-Nutritional Index (CAR) was nearly four times that of the non-epilepsy group (0.81 vs. 0.22, p < 0.001), and the Hypoxia-Nutritional Index (LAR) also showed a significant upward trend (0.07 vs. 0.06, p < 0.001).

The selection of CAR and LAR was primarily driven by our specific pathophysiological hypothesis: the “Two-Hit” mechanism involving immuno-nutritional and hypoxia-metabolic disturbances. While other hematological parameters (e.g., leukocyte count) were also elevated in the epilepsy group, we prioritized CRP and lactate ratios due to their specific mechanistic roles in blood–brain barrier disruption and neuronal hyperexcitability, which are central to epileptogenesis.

Crucially, this biological rationale was corroborated by our machine learning feature selection analysis (Supplementary Figure S1). In the LASSO regression, both CAR and LAR retained non-zero coefficients at the optimal lambda value, indicating non-redundancy. In the Random Forest model, CAR ranked as the third most important predictor (Importance Score: 0.155), and LAR ranked fifth (0.108), outperforming established risk factors such as age and D-dimer. The fact that CAR was identified as a top predictor alongside WBC confirms that the composite “Immuno-Nutritional” status—reflecting the balance between systemic inflammation (insult) and nutritional reserve (defense)—offers distinct prognostic information not fully encompassed by standard metrics of infection (WBC) or stroke severity (NIHSS).

To assess the independent predictive value of the biomarkers, univariate and multivariate logistic regression analyses were conducted (Table 2). Given the relatively narrow ranges of these indices, data were standardized, and reported OR represent the change in risk per 1-SD increase. In unadjusted models, both CAR (OR 2.03, 95% CI 1.95–2.11) and LAR (OR 1.62, 95% CI 1.55–1.70) demonstrated strong correlations with PSE risk. After adjusting for potential confounders including age, gender, NIHSS score, cortical involvement (frontal/temporal/parietal), and electrolyte levels, these two indices retained independent predictive value. Multivariate analysis revealed that for every 1-SD increase in CAR, epilepsy risk increased by 1.96-fold (adjusted OR 1.96, 95% CI 1.86–2.06, p < 0.001); for every 1-SD increase in LAR, risk increased by 1.71-fold (adjusted OR 1.71, 95% CI 1.62–1.81, p < 0.001). For clinical translation, the full specifications of the parsimonious model used for the web calculator—including the model intercept and regression coefficients—are explicitly provided in Supplementary Table S3.

To rigorously assess the conceptual independence of CAR and LAR and rule out mathematical coupling due to the shared denominator (albumin), we performed specific validation analyses. First, Variance Inflation Factor (VIF) analysis revealed low values for CAR (1.42) and LAR (1.12), well below the threshold of 5, thus ruling out significant multicollinearity. Second, in a comparative sensitivity analysis, the “Two-Hit Ratio Model” (utilizing composite indices, AUC 0.888) demonstrated predictive performance comparable to a “Decomposed Model” incorporating CRP, lactate, and albumin as separate covariates (AUC 0.892). This finding confirms that the composite indices effectively capture the synergistic prognostic information of their components while offering greater clinical interpretability (Supplementary Table S4).

Restricted cubic spline models were utilized to visualize the non-linear dose–response relationships between these indices and epilepsy risk (Figure 2). The immuno-nutritional hit (CAR) exhibited a typical “J-shaped” non-linear association (Figure 2A): epilepsy risk remained stable at lower levels but accelerated sharply once the index exceeded the population mean (0 SD), suggesting a threshold effect of systemic inflammation on the epileptogenic process. Notably, the hypoxia-nutritional hit (LAR) displayed a unique “bell-shaped” (inverted U-shaped) relationship (Figure 2B): risk peaked at moderate elevations (approximately 0.10 SD) but paradoxically declined at extremely high values (> 4 SD). This phenomenon suggests complex biological mechanisms, potentially related to seizure suppression by severe metabolic acidosis or competing risks such as mortality. Furthermore, Spearman correlation analysis (Figure 3A) showed a significant positive correlation between standardized CAR and LAR (r = 0.385, p < 0.001), indicating that while the two are interrelated, they represent distinct pathophysiological pathways, supporting their joint inclusion in prediction models.

To validate the “Two-Hit” hypothesis, we grouped CAR and LAR by quartiles and plotted a heatmap of epilepsy incidence (Figure 3B). The results revealed a clear diagonal gradient in risk distribution: patients in the “Double Low” group (CAR-Q1 and LAR-Q1) had the lowest epilepsy incidence (2.4%), whereas those in the “Double High” group (CAR-Q4 and LAR-Q4) had the highest incidence (13.5%). Notably, the risk in the concurrent high-value group (13.5%) significantly exceeded the risk associated with the elevation of single markers (CAR-Q4 alone: ~5.5%; LAR-Q4 alone: ~4.3%). This strongly confirms that immune-inflammatory and hypoxia-metabolic stresses possess a significant synergistic effect in the epileptogenic process.

To evaluate the robustness of CAR’s predictive value, we performed stratified subgroup analyses by age, gender, stroke severity (NIHSS), hypertension, and cortical involvement (Figure 4). The results indicated that the predictive value of CAR remained highly consistent across all subgroups (interaction p-values > 0.05). Clinically significant is the finding that even in subgroups with minor stroke (NIHSS < 4) and no cortical involvement, CAR remained a significant independent predictor. This suggests that the contribution of systemic immuno-nutritional dysregulation to epilepsy risk is independent of structural brain injury severity, implying that this marker holds important screening value even in traditionally low-risk populations.

We further evaluated the improvement in prediction accuracy by adding CAR and LAR to traditional clinical models (Table 3; Figure 5). The Baseline Model, comprising demographic characteristics, NIHSS, and lesion location, yielded an AUC of 0.856 (95% CI 0.844–0.868). Adding single biomarkers improved model performance (Baseline + CAR: AUC 0.873; Baseline + LAR: AUC 0.884). The “Two-Hit Model,” integrating baseline features with both biomarkers, demonstrated the optimal discriminatory ability, achieving an AUC of 0.888 (95% CI 0.877–0.899) (Figure 5A). Beyond AUC enhancement, reclassification metrics showed substantial improvement in risk stratification: the Two-Hit Model yielded a NRI of 0.820 (95% CI 0.755–0.880, p < 0.001) and an IDI of 0.139 (95% CI 0.125–0.153, p < 0.001). DCA (Figure 5B) confirmed that the Two-Hit Model provided higher net benefit than the Baseline Model across a wide range of probability thresholds, and the calibration curve (Figure 5C) showed good agreement between predicted and observed probabilities (Brier score = 0.032).

To rigorously assess the robustness of the model, we performed internal validation using 1,000 bootstrap resamples. The model demonstrated high stability with minimal overfitting. As shown in Supplementary Table S5, the optimism-corrected AUC was 0.886, which was nearly identical to the apparent AUC (Optimism = 0.0004). Furthermore, the model exhibited excellent calibration, yielding an optimism-corrected Brier score of 0.033 and a calibration slope of 0.998. The calibration plot (Supplementary Figure S2) visually confirmed the strong agreement between predicted probabilities and observed outcomes across all risk deciles.

To facilitate the clinical translation and practical application of our findings, we constructed an interactive, open-access web-based risk calculator based on the final multivariate “Two-Hit” model (Supplementary material). This digital tool integrates easily accessible clinical variables (age, gender, NIHSS score, cortical involvement) and the two core biomarkers (CAR and LAR). Clinicians need only input the patient’s specific parameters, and the calculator instantly outputs the personalized probability of developing PSE within the next year (e.g., the example patient in Figure 6 has a predicted probability of 12.3%, classified as intermediate risk). Based on preset thresholds, the tool stratifies patients into risk categories (low/intermediate/high). Compared to traditional static nomograms, this dynamic tool avoids manual reading errors and enables rapid, precise bedside risk assessment to assist clinicians in formulating individualized neuroprotective and monitoring strategies.

The core innovation of this study lies in utilizing CAR and LAR to quantify the systemic “Two-Hit” effect. We explicitly interpret these acute-phase biomarkers not merely as statistical predictors, but as accessible systemic surrogates for the central pathological processes of “epileptogenic priming”—specifically, BBB disruption, neuroimmune activation, and cerebral metabolic dysregulation.

The First Hit (Immune-Inflammatory Cascade): We observed a robust J-shaped association between CAR and PSE risk. To address the temporal latency between acute inflammation and delayed seizures, we frame CAR not as a direct real-time driver of chronic epilepsy, but as a “risk-associated marker” reflecting the magnitude of the initial “epileptogenic priming.” Mechanistically, this acute-to-chronic bridge is mediated by early blood–brain barrier (BBB) dysfunction. Systemic inflammation (surged CRP) correlates strongly with circulating pro-inflammatory mediators such as IL-1β, TNF-α, and HMGB1, which can directly compromise endothelial tight junctions (13, 26). This acute leakage allows serum albumin to extravasate into the brain parenchyma. Crucially, unlike a passive bystander, this extravasated albumin binds to TGF-β receptors on perivascular astrocytes, triggering a persistent signaling cascade. This pathway leads to the downregulation of inwardly rectifying potassium channels (Kir4.1) and glutamate transporters (GLT-1), resulting in impaired potassium buffering and aberrant synaptogenesis that unfolds over weeks to months (10, 27–30). Therefore, although CAR is measured acutely, it captures the severity of the initial neuroimmune activation and BBB insult that sets the biological trajectory for delayed epileptogenesis. The sharp rise in risk when CAR exceeds the population mean (0 SD) implies that the epileptogenic network may only be fully activated when the inflammatory burden breaches the body’s compensatory limit.

The “bell-shaped” relationship between LAR and PSE provides novel insights into cerebral metabolic distress. Lactate is not merely a byproduct of anaerobic glycolysis but a critical modulator of neuronal excitability (14, 15). In the context of stroke, elevated lactate signals mitochondrial dysfunction and an “energy crisis” in the ischemic penumbra. The LAR index quantifies this metabolic stress relative to the host’s homeostatic reserve. Moderate elevation of LAR (approx. 0.10 SD) likely reflects a “metabolic hyper-excitatory state”: hypoxia-induced acidosis promotes glutamate release while inhibiting GABAergic transmission, thereby lowering the seizure threshold (31). However, the declining risk trend at extremely high LAR levels (>4 SD) suggests complex mechanisms. Biologically, severe acidosis might inhibit NMDA receptor activity (“paradoxical anticonvulsant” effect) or be accompanied by multiple organ failure, introducing a competing risk of death (32–34). Thus, LAR functions as a surrogate for the cerebral metabolic dysregulation that precedes network reorganization. Clinicians should maintain high vigilance for patients in the ‘moderate metabolic disturbance’ zone (LAR ~0.10 SD), where hypoxic metabolism is sufficient to induce excitotoxicity but has not yet reached the acidosis threshold for neural suppression.

Our heatmap analysis visually confirmed the “Two-Hit” hypothesis. Patients in the concurrent high-value group (High CAR + High LAR) had an epilepsy incidence (13.5%) nearly triple that of the single-marker elevation groups. This supports a pathological model where an inflammatory response and hypoxic hit, superimposed on a foundation of nutritional depletion (low albumin) constituting a susceptible brain microenvironment, produce cumulative damage that ultimately facilitates the formation of epileptogenic foci.

Conventional wisdom suggests that PSE prediction relies primarily on neuroimaging anatomical features such as infarct location (35–37). However, subgroup analysis in this study revealed a clinically significant finding: CAR and LAR maintained significant predictive efficacy even in patients with minor stroke (NIHSS score < 4) and no cortical involvement (i.e., deep or subcortical infarcts). This result revises the traditional perception that “minor stroke or deep infarction implies low epilepsy risk,” indicating that systemic immuno-metabolic microenvironmental dysregulation can drive epileptogenesis independently of the extent of structural brain injury. For these “cryptic” patients with mild radiographic injury but hematological markers suggesting “high metabolic risk,” the model constructed in this study provides a crucial supplement for risk identification.

To promote the effective translation of statistical models into clinical application, we developed the complex multivariate regression model into an interactive web-based risk calculator. Compared to traditional static nomograms, this digital tool demonstrates significant clinical advantages: it enables immediate bedside assessment, provides precise individualized probability estimates, and effectively eliminates errors associated with manual chart reading. Using this tool, clinicians can rapidly identify patients without obvious cortical structural damage but with extremely high metabolic risk, thereby tailoring individualized monitoring plans (e.g., long-term EEG monitoring) or preventive neuroprotective strategies. The NRI of 0.820 further confirms the tool’s core value in optimizing risk stratification—specifically, accurately correcting misclassifications by traditional anatomical models and identifying potential high-risk patients within the low-risk population.

Furthermore, the findings of this study are not limited to risk stratification but also provide a new theoretical perspective for early intervention in PSE. Previous neurocritical care management has often focused on intracranial pressure and glycemic control, without fully appreciating the central role of basal nutritional status in neuro-immune regulation. Given that CAR and LAR reveal the critical position of “immuno-nutritional” imbalance in the epileptogenic process, targeted nutritional support strategies show potential value in preventing PSE. For instance, early initiation of enteral nutrition to maintain intestinal barrier integrity, or appropriate supplementation of human serum albumin in patients with severe hypoalbuminemia to restore the body’s antioxidant defense capacity, may help mitigate the pathological cascade of the “Two-Hit.” Although this hypothesis requires further verification through prospective randomized controlled trials, our model emphasizes the importance of integrating “nutritional resuscitation” into neuroprotective strategies, aligning closely with the “Immunonutrition” therapeutic concept currently advocated in critical care medicine (38, 39).