The stakes are higher than they might first appear. Modern epilepsy care increasingly leans on artificial intelligence: machine learning models watch hours of video for telltale movements, and large language models are being explored to extract structured information from unstructured electronic medical records. But an AI system is only as good as its labels, and a European expert working group, the European Reference Network for Rare and Complex Epilepsies, has cautioned that while large language models can extract and interpret unstructured data from medical records, further validation and clinical implementation are needed to ensure accuracy, reliability and integration into routine practice. A fixed, numerical seizure taxonomy, curated by expert humans, is precisely the kind of foundation those experts say is needed for training and validating clinical algorithms.

The ILAE 2025 position paper simplified and clarified the 2017 system, streamlining taxonomic rules, sharpening the definition of consciousness and formalizing phenomena-based descriptors that capture what a seizure actually looks like. For clinicians, the streamlined model is easier to adopt across different care settings. For software developers, however, it created a new problem: how do you represent a hierarchical, nuanced clinical classification inside a medical record, a relational database, or a stream of data flowing between devices, without every vendor inventing its own incompatible encoding? Knight and colleagues answered with a technique familiar to low-level programmers but rarely seen in clinical informatics: bitfields.

The idea is elegantly simple. Each independent clinical concept — whether a seizure is focal or generalized, whether consciousness was preserved or impaired, whether the event was epileptic at all — is assigned its own single bit inside an unsigned integer. Because every possible seizure type is just a combination of these constituent features, the whole core ontology of the ILAE 2025 system fits comfortably within a 32-bit unsigned integer, a data type supported natively by nearly every programming language and relational database in existence. The expanded system of roughly 100 predefined semiological descriptors, together with somatotopic modifiers indicating body part and direction, pushes the total to 128 bits, which the authors show still leaves a few bits free for future growth.

The choice of constraints was dictated by an unexpected bottleneck: JavaScript. While languages such as C, C++, Rust and Java offer native integers up to 128 bits wide, JavaScript — arguably the most consequential language for modern applications — treats its numbers as double-precision floating point values and limits bitwise operations to 32 bits. If the specification must work everywhere, it must work in JavaScript. The authors therefore designed the core classification to fit within 32 bits, reserving wider big-integer representations for the descriptor space. Even then, JavaScript’s highest bit requires special handling to avoid accidentally introducing negative signed values, a pitfall the team baked into their library as an automatic safeguard.

For the reference implementation, the team chose TypeScript, described as JavaScript with syntax for types. The language’s build-time type checking proved ideal for numeric modeling: enumerated subsets can be defined as types so the compiler guarantees only valid bit positions are used, and derived types give self-describing names to the large, unmemorable numbers that arise when bitfields grow. Crucially, TypeScript code still runs as JavaScript at runtime, in web browsers and Node.js servers alike, making it a natural lingua franca for a specification intended to cross software boundaries. The full reference implementation, with test cases and examples, is released under the permissive MIT license in a public repository.

The encoding is not merely a storage trick; it enables surprisingly powerful querying. Bitwise AND operations reveal whether one value’s traits are present in another, while OR combines traits. Searching a database of thousands of recorded events for all focal seizures, all tonic-clonic events including their subtypes, or all events with a postictal descriptor becomes a single arithmetic comparison per row — fast, indexable and portable. The authors catalog more than two dozen search patterns, and they demonstrate the system against real clinical vignettes from the ILAE practical guide, including a 12-year-old boy’s focal-to-bilateral tonic-clonic seizure whose gyratory movements, head version and eye deviation are each captured as ordered, timestamped descriptors ORed onto the base classification.

Migration from the old standard was a central concern. The authors tested the claim that ILAE 2017 databases convert easily, mapping all 63 coded seizure types proposed by the SCORE consortium into the new system. The result is reassuring but nuanced: every 2017 type maps to a unique, reversible 2025 value, yet because the 2025 model eliminated many subtypes — delegating them to descriptors instead — the mapping often requires the combination of classification plus descriptor to remain lossless. Ambiguous motor and nonmotor cases need extra information from free-text notes or manual review before migration. The project is validated by a 326-test conformance suite, 665 unit tests in continuous integration, and independent reimplementations in Python and JavaScript that prove cross-language interoperability.

For clinical workflows, the researchers found that classification is inherently additive: as evidence accumulates — from a caregiver’s description to a video-EEG recording in an epilepsy monitoring unit, the diagnostic gold standard — a label migrates through successive numerical identities. They show how an unknown-origin seizure with an attached myoclonic descriptor becomes a generalized myoclonic seizure as EEG evidence arrives, and argue that medical software should retain those descriptors and edit histories even after a seizure is reclassified as generalized, because the temporal details carry clinical value that the strict standard does not require. A validate() function checks stored values against clinical rules, flagging impossible combinations such as a seizure labeled both focal and generalized.

The team is candid that the project, at version 0.0.4, remains exploratory and has not been endorsed by the ILAE or any standards organization. Its intent is to spark community discussion and, ultimately, an open consensus standard adopted across software solutions, with semantic versioning protecting early adopters from silent encoding changes. Future work includes conformance tests for more languages, benchmarks testing how large language models might transform unstructured clinical text into the standardized numeric labels, and mappings to established interoperability ecosystems including HL7 FHIR, SNOMED CT, the OMOP Common Data Model and the OBO Epilepsy Ontology. Preliminary mappings already exist, from hexadecimal strings stored as FHIR Observation values to PostgreSQL bit-string columns queried with native bitwise operators — demonstrated end-to-end in an example web application where a seizure label travels from a browser’s big integer, through JSON, a Rust web server and back into the database without losing a single bit. If the community adopts it, a single well-chosen integer could become the common tongue through which clinicians, databases and AI systems finally describe seizures the same way.

Subject of Research: A numerical bitfield encoding for implementing the ILAE 2025 seizure classification taxonomy in medical software

Article Title: Implementing the ILAE 2025 seizure taxonomy in medical software

Article References: Knight, A., Jousea, S., & Peltola, J. (2026). Implementing the ILAE 2025 seizure taxonomy in medical software. Frontiers in Computational Neuroscience, 20, Article 1877135. https://doi.org/10.3389/fncom.2026.1877135

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1877135

Keywords: ILAE 2025, seizure classification, epilepsy, medical software, bitfield encoding, interoperability, TypeScript, clinical informatics, seizure descriptors, artificial intelligence, electronic medical records, health data standards