**Background:** Epilepsy is a complex neurological disorder affecting ~50 million people worldwide, with a lifetime prevalence of ~1%. Despite the availability of effective treatments, a significant treatment gap persists, particularly in low- and middle-income countries (LMICs), where >75% of people lack access to medication. The development of antiepileptic drugs (AEDs) has historically relied on serendipity, with key discoveries such as potassium bromide (1857), phenobarbital (1912), and valproic acid (1962) arising from unexpected observations. Over time, preclinical models and clinical trial methodologies have evolved, and more recently, artificial intelligence (AI) and machine learning (ML) have emerged as promising tools to accelerate drug discovery and address the needs of the ~30% of patients with drug-resistant epilepsy.
**Methods:** This is a narrative review that synthesizes historical and contemporary literature on AED development. The authors describe key preclinical models—maximal electroshock stimulation (MES, introduced in 1937), pentylenetetrazol (PTZ)-induced test (validated in 1944), and kindling models (first described in 1969)—and their respective roles in identifying antiseizure activity. For clinical trials, a PubMed search was conducted using individual AED names as MeSH terms combined with 'Clinical Trial' [Publication Type], with article selection performed independently by two clinical neurologists and discrepancies resolved by a third clinical pharmacologist. The review also discusses the application of AI/ML techniques in drug development, including quantitative structure–activity relationships (QSAR), target identification, and pharmacology repurposing.
**Key Results:** The review highlights that early AEDs (e.g., phenytoin, carbamazepine, valproic acid) were approved based on uncontrolled observational studies with small sample sizes (e.g., phenytoin: 200 subjects, average follow-up 4.3 months) and without placebo controls. From the 1930s to the 1950s, clinical trials increased from 15% to 56% of published studies, but methodological rigor was low: only 110 of 250 studies had a formal protocol, only three used a placebo, and 107 were uncontrolled or did not report bias. The Kefauver–Harris Drug Amendment (1962) mandated 'adequate and well controlled investigations,' leading to the modern Phase 1–3 trial structure. Second-generation AEDs (e.g., lamotrigine, felbamate, topiramate) emerged from the 1990s onward, characterized by randomized, double-blind, placebo-controlled designs. Despite these advances, only a small proportion of patients with refractory epilepsy achieve seizure control with newer agents. Preclinical models have limitations: MES and PTZ tests fail to detect efficacy of some clinically effective AEDs (e.g., levetiracetam is ineffective in both but works in kindled models). The Anticonvulsant Screening Program (ASP, now Epilepsy Therapy Screening Program, ETSP) has screened >32,000 compounds by 2018, contributing to the identification of felbamate, topiramate, and lacosamide. AI/ML techniques have been applied to target identification, lead compound discovery, and pharmacology repurposing, identifying potential AEDs such as doxycycline, metformin, nifedipine, and pyrantel tartrate.
**Clinical Implications:** The review underscores that despite over a century of AED development, a substantial proportion of patients—especially in LMICs—remain untreated or receive suboptimal therapy (e.g., phenobarbital, which has neurotoxic side effects). The high rate of drug-resistant epilepsy (25–40% of focal epilepsy patients) highlights the need for etiology-oriented therapies and precision medicine. AI and ML offer opportunities to integrate multi-omics data, improve target identification, and accelerate drug repurposing, potentially bridging the gap between translational research and clinical practice. However, challenges remain in reproducibility, external validation, and the need for universal standards. The authors recommend that future AED development should prioritize accessible, cost-effective treatments for LMICs, continue rigorous RCTs supplemented by phase IV pragmatic studies, and leverage AI to address the heterogeneity of epilepsy and the unmet needs of pharmacoresistant patients.