Artificial Intelligence for Sleep Bruxism Detection: A Technical Pipeline Synthesis

Sleep bruxism is the rhythmic or nonrhythmic masticatory muscle activity that occurs during sleep. It is conventionally diagnosed by a rule-based algorithm. Candidate muscle bursts are flagged against a fixed electromyographic (EMG) amplitude threshold, classified by duration into phasic, tonic, or mixed episodes, and counted into an hourly rate compared against a second threshold. These thresholds generalize poorly across individuals and nights. This has motivated a growing body of artificial intelligence (AI) research aimed at replacing one or more stages of this classical algorithm with a learned decision function. This paper synthesizes that literature as a technical pipeline across 28 reviewed AI-based sleep bruxism studies, identified through Web of Science and Google Scholar searches for "sleep bruxism" combined with artificial intelligence terms, conducted in 2026, and screened for a learned, trained classification approach reported in English. These accuracy figures should be read alongside several complicating factors. Ten studies draw on the same subset of two patients from a public polysomnography database. Several studies framed as sleep bruxism detection were validated only on awake, simulated grinding. In addition, the ground-truth labels these classifiers are trained against remain contested within the field's own consensus literature. The paper's contribution is therefore twofold. It assembles the pipeline synthesis itself, spanning ground truth, recording modality, preprocessing, feature extraction, and architecture, as a technical reference not previously brought together in this form. It also uses that synthesis to set a concrete research agenda, including validation benchmarks across nights and across subjects, a shift from binary toward continuous severity targets, and an extension of detection into real-time treatment devices and clinical prognosis.