Differentiating Laryngeal Dystonia and Muscle Tension Dysphonia Using Artificial Intelligence: A Narrative Review of Current Evidence and Future Directions.

IntroductionLaryngeal dystonia (LD) (also referred to as spasmodic dysphonia), and muscle tension dysphonia (MTD), are voice disorders that present with some overlapping features. This may result in both diagnostic uncertainty and delayed treatment. Because of the different pathological origins of these disorders, as well as appropriate treatments, accurate differentiation is essential. Current clinical methods often rely on subjective perceptual metrics and inconsistent physiologic markers. This narrative review synthesizes current evidence on artificial intelligence (AI) and machine learning (ML) approaches for voice-disorder analysis and evaluates their potential to support the differential diagnoses of LD and MTD.MethodsPubMed, Google Scholar, EMBASE, and Scopus were searched using Boolean combinations of terms related to LD, MTD, voice disorders, AI/ML, acoustic features, and clinical differentiation, yielding 807 results and 522 articles following the removal of duplicates. After screening and full-text review, 23 studies met the inclusion criteria.DiscussionAI and ML models demonstrate high accuracy for general dysphonia detection with traditional feature-based classifiers and deep learning systems achieving >90% performance across several datasets. Notably, this high accuracy does not equate to LD-MTD differentiation. However, existing datasets minimally represent LD and contain virtually no MTD representation, limiting the clinical applicability. LD-specific physiologic signatures (task-dependent spasms and reduced cortical inhibition) are recognized in clinical literature but rarely incorporated into computational models. Hyperfunctional patterns characteristic of MTD, including increased effort and spectral noise, remain underrepresented in AI research.ConclusionThough AI shows promise for objective voice-disorder assessment, current methodologies and datasets are insufficient for reliable LD-MTD differentiation. Progress will require etiologically specific, expert-labeled datasets as well as multimodal acoustic, physiologic, and imaging inputs.