Approximately 10–25 million EEG exams are performed annually worldwide [1]. For epilepsy diagnosis, advances in computer systems, data storage, and processing, combined with the growing demand for personalized and precision medicine, have also contributed to a substantial increase in the number of EEG recordings conducted in both hospital and home settings [2], [3]. In particular, the demand for advanced patient care has driven the need for high-resolution EEG data, offering deeper insights into patient-specific neural activity and treatment responses. A notable example is intracranial EEG in patients with complex epileptogenic zones, where hundreds of electrode contacts generate a vast number of channels that require meticulous review. Furthermore, long-term EEG monitoring, which was once confined to specialized hospital units, is now increasingly performed in home environments thanks to ambulatory [4], subcutaneous [5] or next-generation wearable EEG devices [6]. These new devices allow continuous monitoring over extended periods, providing clinicians with valuable insights into neurological conditions such as epilepsy. These prolonged EEG recordings improve the chances of finding epileptic activity, yielding higher diagnostic efficiency [7]. Nevertheless, the exponential growth in EEG data volume places a considerable burden on healthcare service providers and medical infrastructure.

In clinical practice, Interictal Epileptiform Discharges (IEDs) are the primary epilepsy biomarker in modern EEG analysis [8], as they are present in nearly all epilepsy cases, regardless of etiology [9]. IEDs encompass spikes, polyspikes, and sharp waves, either alone or followed by a slow wave. Their detection is essential for characterizing epilepsy types and identifying pathological brain networks [10]. Moreover, IEDs occur frequently — up to several times per minute — and follow circadian or multidien rhythms, correlating with seizure likelihood. This makes them a promising candidate for personalized seizure risk assessment [11]. In addition, although the role of IEDs in presurgical evaluation remains limited, interictal abnormalities have often been shown to correlate with areas of surgical resection and postoperative seizure outcomes [12], [13]. For instance, in cases of temporal lobe epilepsy with unilateral hippocampal atrophy and concordant clinical and neuropsychological findings, such correlations may, in some cases, obviate the need for mandatory ictal recordings during presurgical workup [14].

In the context of IEDs, visual assessment remains the gold standard for clinical EEG interpretation. However, the manual review of IEDs is highly time-consuming, and requires specialized expertise, leading to an increased workload for neurologists and technicians, especially in long-term EEG recordings. For example, a routine 30-minute EEG recording typically requires between 5 mins and 1 h (median: 13 mins) for visual assessment and reporting by an epilepsy specialist [15]. As a result, reviewing a full day of EEG often demands several hours of intense effort. The time required varies depending on multiple factors, including the presence of abnormalities and the amount of artifacts in the recording. But reader fatigue and distraction are common and likely affect performance. Also, despite well-established classification criteria — such as those defined by the International Federation of Clinical Neurophysiology (IFCN) [16] —, the manual review of IEDs is subjective. This subjectivity contributes to a high misdiagnosis rate and low inter- and intra-observer agreement [17]. Several factors can lead to the high variability:•

the IED is too small and therefore overlooked;

•

the IEDs have their origin in brain areas that are deep and cannot be easily detected from scalp electrodes;

•

their morphology varies significantly, with differences in amplitude, duration, and the slope of both the spike and slow-wave components [16].

Additionally, various factors — such as electrode placement, patient age, state (awake vs. asleep), and medication effects — can influence IED characteristics, further complicating their identification [18].

To address these challenges, automated detection methods have been proposed for IED identification, offering greater consistency and scalability. Research in automated IED detection began 45 years ago and has since evolved from traditional machine learning approaches based on time- and wavelet-domain features to more recent deep learning techniques. Today, after decades of progress, these AI-driven tools are increasingly being utilized to support clinicians by identifying abnormalities in their daily practice and minimizing false-positive rates. This review explores classic and modern AI approaches to IED detection, assessing their performance and limitations. Given the extensive literature on the topic, we focus on representative methods illustrating key concepts in IED detection.