A recent effort — known as Omni-iEEG — brings together pre-surgical brain recordings from eight epilepsy centers, covering 302 patients and 178 hours of data, trying to benefit epilepsy sufferers numbering over 50 million people worldwide. [Photo provided to chinadaily.com.cn]

As artificial intelligence (AI) becomes increasingly common in healthcare research, a Chinese medical doctor working on machine learning applications is pioneering research to benefit epilepsy sufferers, which number over 50 million people worldwide.

"Infrastructure determines whether innovation can translate," said Dr Yipeng Zhang, a researcher working on machine learning applications in epilepsy — a neurological condition characterized by abnormal or excessive brain activity that results in seizures. "If we want AI to assist in surgical decisions, we need frameworks that allow results to be compared across hospitals."

About one-third of epilepsy sufferers worldwide experience seizures that cannot be controlled by medication. For many of these patients, surgical removal of seizure-generating brain tissue provides the best chance of long-term relief. Identifying that tissue relies heavily on intracranial electroencephalography (iEEG), a high-resolution recording of brain activity from implanted electrodes.

Over the past decade, researchers have developed AI systems to assist in analyzing iEEG recordings — particularly in detecting high-frequency oscillations (HFOs) — and signal patterns associated with seizure-generating regions. Many studies report promising performance.

However, most AI systems in this space are trained on data from a single hospital or research center. Differences in recording protocols, labeling conventions, and clinical definitions make it difficult to compare results across institutions or determine whether findings generalize.

"The field has focused heavily on improving AI accuracy," Zhang said, noting that "without shared evaluation standards, it's hard to know whether systems will perform reliably outside the original study setting."

Zhang's earlier work focused on refining pathological HFO detection and contributed to the development of PyHFO, a research tool used by independent groups studying seizure-related brain activity. He said improving individual systems is only part of the challenge.

A recent effort, known as Omni-iEEG, brings together pre-surgical brain recordings from eight epilepsy centers, covering 302 patients and 178 hours of data. The dataset aligns clinical metadata under common standards and defines benchmark tasks that link AI system outputs to post-operative seizure outcomes.

Rather than evaluating whether an algorithm can detect abnormal signals alone, the framework assesses whether the brain regions identified by AI correspond to better surgical results.

Regulatory agencies have increasingly emphasized reproducibility and cross-site validation in medical AI. Experts say multi-center benchmarks may become essential before such systems can be integrated into routine surgical planning.

As AI intelligence continues to expand in clinical research, some experts suggest the next phase of progress may depend less on new algorithms and more on shared standards that enable reliable validation.

For epilepsy surgery, where decisions are irreversible and precision is measured in millimeters, that shift could have significant implications.