Neurology · Medical Literature
Artificial intelligence tools for the assessment and management of dysphagia: a scoping review
Abstract
Objectives To explore the existing artificial intelligence (AI) tools used in the assessment and management of dysphagia. Design Scoping review. Data sources MEDLINE (Ovid), Scopus, CINAHL (EBSCOhost), Cochrane Library, Joanna Briggs Institute (JBI) Evidence Synthesis, ProQuest and Google Scholar were searched for studies published between January 2000 and May 2025. Eligibility criteria Studies involving adults diagnosed with dysphagia that applied AI techniques for screening, diagnosis, prognosis or management were included. Non-human studies, paediatric-only studies, non-clinical simulations and studies without direct clinical application were excluded. Methods The review was conducted in accordance with the JBI methodology and reported following the Preferred Reporting Items for Systematic Review and Meta-Analysis extension for Scoping Reviews (PRISMA-ScR) guidelines. Two reviewers independently screened titles and abstracts, assessed full-text articles for eligibility and extracted data using a standardised data charting form. Data were synthesised using narrative synthesis. Results A total of 4863 records were identified, of which 27 studies met the inclusion criteria, representing 8842 participants. Most studies (21/27, 77.8%) published between 2013 and 2025 and conducted in the USA (9/27, 33.3%), South Korea (6/27, 22.2%), Japan (3/27, 11.1%) and China (3/27, 11.1%). The majority of studies (21/27, 77.8%) employed machine learning or deep learning approaches applied to videofluoroscopic swallowing studies, fibreoptic endoscopic evaluation of swallowing, wearable sensor data, acoustic signals or structured clinical datasets. Most studies are prospective (8/27, 29.6%) and retrospective (7/27, 25.9%) investigations; case series (3/27, 11.1%). Only one randomised controlled trial (1/27, 3.7%) evaluated AI-based dysphagia intervention. AI applications primarily focused on screening and assessment tasks. Conclusion The current literature demonstrates a growin
Clinical relevance
Dysphagia affects a large adult population, particularly those with neurological conditions such as stroke, and carries significant risks of aspiration, pneumonia, malnutrition, and mortality. Current gold-standard assessments—videofluoroscopic swallowing study (VFSS) and fibreoptic endoscopic evaluation of swallowing (FEES)—are resource-intensive, subjective, and limited in accessibility, creating a clinical gap that AI-driven tools aim to address. This scoping review is timely given the rapid proliferation of machine learning and deep learning applications in dysphagia care, providing neurologists and rehabilitation specialists with the first comprehensive map of the available evidence.
Key findings
1. From 4,863 records screened, only 27 studies (n = 8,842 participants) met inclusion criteria, reflecting the nascent state of the field and limited high-quality evidence base. 2. The majority of included studies (21/27, 77.8%) employed machine learning or deep learning approaches, applied across modalities including VFSS, FEES, wearable sensor data, acoustic signals, and structured clinical datasets. 3. Geographic concentration was marked: the USA accounted for 33.3% of studies (9/27), South Korea 22.2% (6/27), Japan 11.1% (3/27), and China 11.1% (3/27), limiting generalizability to other healthcare settings. 4. Study design quality was low overall: most studies were prospective (8/27, 29.6%) or retrospective (7/27, 25.9%) observational investigations, with case series representing 11.1% (3/27); only one randomised controlled trial (1/27, 3.7%) evaluated an AI-based dysphagia intervention. 5. AI applications were predominantly focused on screening and assessment tasks rather than therapeutic management or prognostication, indicating a significant gap in intervention-level evidence. 6. No pooled effect sizes, confidence intervals, p-values, or NNT/NNH were calculable from this scoping review, as narrative synthesis was employed; quantitative meta-analysis was not performed. NOTE: Specific performance metrics (accuracy, AUC, sensitivity/specificity) were not reported in the provided abstract and have not been fabricated.
Practice implications
Clinicians should regard AI-assisted dysphagia tools—particularly those applied to VFSS and acoustic/wearable sensor analysis—as investigational adjuncts rather than validated replacements for standard instrumental assessment, given that only one RCT exists in the entire literature to date. Neurologists and speech-language pathologists managing post-stroke or neurodegenerative dysphagia may cautiously monitor emerging machine learning screening tools, but should require demonstration of prospective, multi-site validation and regulatory approval before clinical adoption. Future procurement and research investment should prioritise AI tools designed for therapeutic management and prognosis, domains currently underrepresented in the literature, and should include diverse patient populations beyond those studied predominantly in the USA, South Korea, Japan, and China.