Artificial Intelligence in Neuromuscular Diseases: Opportunities for a Data-Scarce Field Authors: Shang Ma, Sushan Luo, Huahua Zhong Published: 2026 ABSTRACT Neuromuscular diseases (NMDs) encompass over 800 distinct entities affecting approximately one in 1000 individuals worldwide, with progressive muscle weakness, atrophy, and motor impairment as primary clinical manifestations. The rarity of most NMDs creates fundamental challenges for artificial intelligence (AI) and machine learning (ML) applications that typically require large-scale datasets. In this narrative review we synthesize the literature published between 2018 and 2025 on AI applications across the NMD spectrum, organized by clinical application domain. We examine how AI has advanced diagnostic capabilities through genetic variant interpretation, muscle magnetic resonance imaging analysis, electromyography-based classification, and computational pathology. In disease monitoring and prognosis, wearable-derived digital biomarkers have achieved regulatory qualification (US Food and Drug Administration [FDA] and European Medicines Agency [EMA]) as clinical trial endpoints for Duchenne muscular dystrophy, while AI-driven survival models for amyotrophic lateral sclerosis (ALS) have been validated across 14 European centers. Proteomic and multi-omics analyses using ML have identified diagnostic panels for ALS. However, most reported models were developed and internally validated on single-center datasets, and few have undergone external or prospective validation or clinical implementation. Despite these achievements, research intensity varies dramatically across NMD subtypes, with ALS and Duchenne muscular dystrophy dominating while myotonic dystrophy, congenital myopathies, and metabolic myopathies remain virtually unexplored. Critical gaps persist in computational pathology, multi-center validation, and clinical translation. In this review, we discuss how federated learning, international collaborative networks (TREAT-NMD, Solve-RD, EURO-NMD), and foundation models can address these challenges, and propose directions for future AI-enhanced clinical studies in this data-scarce field.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s40120-026-01017-8. Key Summary Points Why carry out this study? Neuromuscular diseases (NMDs) include more than 800 heterogeneous disorders affecting approximately one in 1000 individuals worldwide, yet most NMD subtypes remain limited by small datasets that constrain the development and validation of artificial intelligence (AI) and machine learning (ML) toolsAlthough AI applications in rare diseases have expanded rapidly, no comprehensive review has systematically evaluated AI across the NMD clinical pathway while addressing the unique challenges of data scarcity, model validation, and clinical translationThe aim of this narrative review was to synthesize recent evidence on AI applications across the NMD spectrum, identify current achievements and unmet needs, and evaluate emerging strategies—including federated learning, foundation models, and international collaborative networks—that may facilitate AI implementation in rare NMDs What was learned from the study? AI has demonstrated substantial potential in NMD diagnosis, disease monitoring, prognosis prediction, and biomarker discovery, achieving high performance in genetic variant interpretation, muscle magnetic resonance imaging analysis, electromyography classification, computational pathology, and multi-omics analysesWearable-derived digital biomarkers have achieved US Food and Drug Administration (FDA) and European Medicines Agency (EMA) qualification as clinical trial endpoints for Duchenne muscular dystrophy, while AI-based prognostic models for amyotrophic lateral sclerosis (ALS) have undergone multi-center validation across European clinical networksResearch activity remains concentrated in ALS and Duchenne muscular dystrophy, whereas several NMD subtypes—including myotonic dystrophy, congenital myopathies, metabolic myopathies, and Lambert–Eaton myasthenic syndrome—remain underrepresented in AI researchMajor barriers to clinical implementation include limited data availability, insufficient external validation, data-sharing restrictions, and regulatory challenges. Emerging approaches such as federated learning, international data-sharing initiatives, and foundation models may help address these limitations and accelerate future AI-enabled precision care in NMDs Why carry out this study? Neuromuscular diseases (NMDs) include more than 800 heterogeneous disorders affecting approximately one in 1000 individuals worldwide, yet most NMD subtypes remain limited by small datasets that constrain the development and validation of artificial intelligence (AI) and machine learning (ML) toolsAlthough AI applications in rare diseases have expanded rapidly, no comprehensive review has systematically evaluated AI across the NMD clinical pathway while addressing the unique challenges of data scarcity, model validation, and clinical translationThe aim of this narrative review was to synthesize recent evidence on AI applications across the NMD spectrum, identify current achievements and unmet needs, and evaluate emerging strategies—including federated learning, foundation models, and international collaborative networks—that may facilitate AI implementation in rare NMDs What was learned from the study? AI has demonstrated substantial potential in NMD diagnosis, disease monitoring, prognosis prediction, and biomarker discovery, achieving high performance in genetic variant interpretation, muscle magnetic resonance imaging analysis, electromyography classification, computational pathology, and multi-omics analysesWearable-derived digital biomarkers have achieved US Food and Drug Administration (FDA) and European Medicines Agency (EMA) qualification as clinical trial endpoints for Duchenne muscular dystrophy, while AI-based prognostic models for amyotrophic lateral sclerosis (ALS) have undergone multi-center validation across European clinical networksResearch activity remains concentrated in ALS and Duchenne muscular dystrophy, whereas several NMD subtypes—including myotonic dystrophy, congenital myopathies, metabolic myopathies, and Lambert–Eaton myasthenic syndrome—remain underrepresented in AI researchMajor barriers to clinical implementation include limited data availability, insufficient external validation, data-sharing restrictions, and regulatory challenges. Emerging approaches such as federated learning, international data-sharing initiatives, and foundation models may help address these limitations and accelerate future AI-enabled precision care in NMDs Introduction Neuromuscular diseases (NMDs) constitute a large, heterogeneous group of diseases affecting the motor unit comprising the anterior horn cells, peripheral nerves, neuromuscular junctions, and skeletal muscles. Progressive muscle weakness represents the primary clinical manifestation, typically accompanied by sensory disturbances, fatigue, and muscle atrophy, with variable rates of disease progression [1]. Most NMDs show prevalence rates of 1–10 per 100,000 general population. Disease-specific estimates highlight the epidemiological profiles: Duchenne muscular dystrophy (DMD) has a global prevalence of 7.1 per 100,000 males [2], amyotrophic lateral sclerosis (ALS) ranges from 1.57 to 11.80 per 100,000 persons [3], and myasthenia gravis (MG) affects 1.5 to 17.9 per 100,000 persons [4, 5]. Although individual subtypes are rare, hereditary NMDs collectively reach an estimated prevalence of 111.9 per 100,000 individuals [6]. These examples also span distinct pathophysiological categories—monogenic muscle disease (DMD), multifactorial motor neuron degeneration (ALS), and antibody-mediated neuromuscular junction dysfunction (MG)—and therefore require different diagnostic, monitoring, and biomarker strategies, a distinction we keep in view when interpreting the AI studies discussed in this review. The vast majority of NMDs qualify as rare diseases under both US (< 200,000 affected) and European Union (EU) (< 5 per 10,000) definitions [7, 8]. Patients with rare NMDs face prolonged diagnostic odysseys. A large-scale European survey reported a mean diagnostic delay of 4.7 years from symptom onset, with 56% of patients requiring > 6 months after initial medical contact. The authors of one study reported that patients consulting more than eight healthcare professionals had a markedly elevated risk of diagnostic delay (odds ratio [OR] 5.15) and that misdiagnosis affected 73% of patients (OR 2.48 for prolonged delay) [7]. Late-onset facioscapulohumeral muscular dystrophy (FSHD) exemplifies this challenge, with mean diagnostic delays of 6.7 years and near-universal misdiagnosis in a 13-patient cohort [9]. These diagnostic delays directly translate into delayed therapeutic interventions, highlighting the urgent need for decision-support tools. Artificial intelligence (AI) and machine learning (ML) could help address these challenges, but their application to NMDs runs into a basic problem: the algorithms that work best in common diseases depend on large datasets that rare diseases cannot provide. Privacy risks and small sample sizes represent significant barriers, with only 11.8% of ML studies on rare diseases validated on external datasets [10]. Traveling model approaches combined with federated learning demonstrate feasibility even with single-sample contributions per site, as shown across 83 international centers for neurological disease classification [11]. Integrative strategies, such as harmonizing 1221 skeletal muscle transcriptomes across heterogeneous sources, have revealed disease progression spectra undetectable in individual cohorts [12]. In this narrative review, we examine AI applications across the NMD spectrum, organized by clinical domain. We assess strategies for overcoming data scarcity, evaluate diagnostic and monitoring applications, identify gaps in coverage across disease subtypes, and propose directions for future studies. An overview of the major applications of AI across the NMD clinical pathway is presented in Fig. 1.Fig. 1Overview of artificial intelligence (AI) applications across the clinical pathway in neuromuscular diseases (NMDs). Top: Four methodological strategies address the inherent data scarcity of rare diseases. Middle (shaded): AI tools organized by clinical domain, from genetic diagnosis through biomarker discovery, with representative performance metrics. Bottom: research intensity across NMD subtypes, highlighting the concentration in amyotrophic lateral sclerosis (ALS) and Duchenne muscular dystrophy (DMD) and the near-absence of AI work in most other conditions. The numbers in each box of the "bottom layer" represent the number of publications on AI intelligence for each subtype of neuromuscular diseases between 2018 and 2025. For abbreviations of NMD subtypes not defined in caption, see Table 1. AUC Area under the ROC curve, EMG electromyography, FDA/EMA US Food and Drug Administration/European Medicines Agency, MRI magnetic resonance imaging Overview of artificial intelligence (AI) applications across the clinical pathway in neuromuscular diseases (NMDs). Top: Four methodological strategies address the inherent data scarcity of rare diseases. Middle (shaded): AI tools organized by clinical domain, from genetic diagnosis through biomarker discovery, with representative performance metrics. Bottom: research intensity across NMD subtypes, highlighting the concentration in amyotrophic lateral sclerosis (ALS) and Duchenne muscular dystrophy (DMD) and the near-absence of AI work in most other conditions. The numbers in each box of the "bottom layer" represent the number of publications on AI intelligence for each subtype of neuromuscular diseases between 2018 and 2025. For abbreviations of NMD subtypes not defined in caption, see Table 1. AUC Area under the ROC curve, EMG electromyography, FDA/EMA US Food and Drug Administration/European Medicines Agency, MRI magnetic resonance imaging Methods The PubMed, Web of Science, Scopus, and Google Scholar databases were searched for articles published between January 2018 and June 2025, with emphasis on the period 2022–2025. Search terms combined NMD-related keywords (“neuromuscular disease,” “muscular dystrophy,” “amyotrophic lateral sclerosis,” “myasthenia gravis,” “spinal muscular atrophy,” “myopathy,” “neuropathy”) with AI-related terms (“artificial intelligence,” “machine learning,” “deep learning,” “neural network,” “transfer learning,” “federated learning,” “natural language processing”). Eligible studies included original research, systematic reviews, and technical reports applying AI/ML to NMD diagnosis, prognosis, or treatment response prediction. We prioritized peer-reviewed studies that described a defined AI/ML method with quantitative performance and that analyzed human NMD cohorts (or directly relevant disease models); we excluded conference abstracts without an accessible full text, studies lacking an explicit AI/ML component, and reports in which NMDs could not be separately assessed. Records were screened by one author (SM) and verified by the senior authors (SL and HZ), with disagreements resolved by consensus. As a narrative rather than systematic review, we did not perform a formal risk-of-bias appraisal or quantitative meta-analysis; instead, study limitations are discussed qualitatively throughout. Throughout the entire review, we use “AI” as an umbrella term spanning classical ML, deep learning (DL), natural language processing (NLP), and large language models (LLMs)/foundation models, as distinct from conventional statistical prediction models, and we denote the specific model class where it bears on interpretation. Additional references were identified through manual screening of reference lists. In total, 80 references were incorporated into this narrative review. Ethical Approval This article is based on previously conducted studies and does not contain any new studies with human participants or animals performed by any of the authors. Methodological Strategies for Overcoming Data Scarcity Applying AI to rare diseases forces researchers to confront a mismatch: most ML methods rely on thousands of labeled samples, while many NMD subtypes affect fewer than 100 patients at any single center. Four strategies (transfer learning; few-shot learning; federated learning; synthetic data generation) have proven effective, although the maturity and clinical readiness of these strategies differ considerably. Transfer learning is currently the most practical approach. The principle is straightforward—train on abundant data from a related domain, then fine-tune on the small NMD-specific dataset. Hu et al. showed that pre-training on 1.6 million electrocardiograms improved the detection of rare cardiac diseases from an area under the ROC curve (AUC) of 0.49–0.70 to AUC 0.77–0.88 [13]. In the NMD field, convolutional neural network (CNN) models pre-trained on general imaging tasks and fine-tuned on 2536 thigh magnetic resonance imaging (MRI) scans from 432 patients reached 91% accuracy for dystrophinopathy classification, outperforming experienced radiologists (84% accuracy) [14]. The rapid growth of biomedical foundation models (BiomedCLIP with 15 million image-text pairs [15]; PathChat with over 100 million histology patches [16]) means that researchers in the field of NMDs can increasingly benefit from models that have already learned general visual features, even when disease-specific training data remain scarce. Few-shot learning takes transfer learning further by explicitly designing algorithms that learn from very few examples. SHEPHERD, an AI deep learning tool, demonstrated this on 465 patients in the Undiagnosed Diseases Network research study, using a knowledge graph of over 100,000 nodes and 1 million edges to diagnose rare diseases from phenotype descriptions alone [17]. Prototypical networks have achieved 88.9% accuracy with as few as ten labeled examples per class [18]. These methods are particularly relevant for NMDs with extremely small patient populations, although they have not yet been tested in NMD-specific clinical workflows. Federated learning solves a different problem: how to train models across multiple hospitals without centralizing sensitive patient data. In a three-hospital candidemia study, federated learning improved the AUC by 9% and true positive rate by 24% over locally trained models [19]. For NMDs, the EURO-NMD registry has built FAIR (Findable, Accessible, Interoperable, and Reusable)-compliant federated infrastructure across 82 European neuromuscular centers [20], providing the technical backbone for future multi-center AI studies. The practical challenge remains that federated learning requires standardized data formats and governance agreements, which most NMD registries still lack. Synthetic data generation using Generative Adversarial Networks (GANs) and diffusion models can artificially expand small datasets. GAN-augmented training improved the external validation AUC from 83.6% to 93.3% in one MS lesion detection study [21], and text-guided diffusion models have shown promise for generating anatomically plausible training images [22]. However, the biological validity of synthetic muscle imaging data remains unproven, and this approach carries a risk of amplifying biases present in the original small sample. AI in NMD Diagnosis Genetic Variant Interpretation The AI model AlphaMissense classifies 89% of all 71 million possible human missense variants with 90% precision (benchmarked against the ClinVar archive classifications), leveraging the AlphaFold AI system architecture without requiring extensive disease-specific training sets [23]. Among the earliest ML approaches, CADD integrates more than 60 genomic annotations to score the deleteriousness of single-nucleotide variants and indels across the genome, with its CADD-Splice extension adding DL-derived splice predictions [24, 25]. The SpliceAI DL-based tool predicts splice junctions from pre-messenger RNA (mRNA) sequences, estimating that 9–11% of pathogenic mutations in rare genetic disorders are caused by cryptic splicing [26], while the transformer-based Spliformer DL tool further improves speed and accuracy [27]. A separate line of work uses phenotype information to narrow the search space. The Exomiser software application, when incorporating human phenotype ontology (HPO) terms, places the correct diagnostic variant in its top-10 list in 88.2% of cases—compared to just 3–27% without phenotypic input [28, 29]. Other tools take different angles on the same problem: Phen2Gene prioritizes genes from HPO terms [30]; AMELIE mines the primary literature to rank candidate genes [31]; and Face2Gene’s DeepGestalt recognizes over 200 syndromes from facial photographs with 91% top-10 accuracy, later extended to over 1000 disorders by GestaltMatcher [32]. At the level of clinical deployment, a Baylor Genetics model ranked 96.7% of diagnostic variants in the top ten across 180 cases, reaching 98.4% in trio analyses [33]. Together, these tools suggest that AI-assisted variant interpretation is approaching practical utility for clinical genetics laboratories. Medical Imaging Muscle MRI Magnetic resonance imaging-based AI in NMDs has advanced along two parallel tracks: automated segmentation and disease classification. Regarding segmentation, the convolutional neural network HRNet achieved a mean (± standard deviation [SD]) Dice score of 0.91 ± 0.08 across 67 patients with NMD and identified all individual muscles in the dataset, with the important caveat that performance dropped when fatty infiltration exceeded 20% [34], precisely the stage at which quantitative assessment matters most clinically. AI-accelerated T2 mapping has reduced scan times by over 50% while preserving diagnostic agreement (r = 0.99, intraclass correlation [ICC] 0.992) [35], a practical advance for patients who often struggle with prolonged scanning sessions. Regarding classification, the most clinically compelling result comes from differentiating DMD from Becker muscular dystrophy (BMD) in children aged 36–60 months, where ML algorithms achieved 81–91% accuracy compared to 19% specificity for expert radiologists [36]. This is a diagnostic window in which early distinction directly affects treatment decisions. A Swin Transformer reached 96% accuracy classifying healthy controls, BMD, and autosomal recessive limb girdle muscular dystrophy (LGMD2) in a three-class setting [37]. The AI-based Myo-Guide platform took the broadest approach, pooling 2961 MRI scans across 20 NMD types from 26 centers. Its top-3 accuracy of 84.7% (cross-validation) and 75.0% on 14 unseen cases exceeded that of the best expert clinician (64.3%) [38], though the small head-to-head sample warrants cautious interpretation. Muscle ultrasound Ultrasound AI is less mature, partly because standardized imaging protocols are less established. The largest segmentation study used 3917 images from 1283 subjects (precision 0.88, recall 0.92) [39], and the DeepACSA tool provides automated cross-sectional area measurement with an ICC of 0.97–0.99 [40]. More clinically relevant is ML-driven Heckmatt grading for FSHD: 25,005 images from 290 participants, with AUC of 0.95 for normal and 0.97 for abnormal classification, externally validated [41]. Synthetic ultrasound generation via CycleGAN [42] may help address the scarcity of labeled data, although clinical validation is lacking. Across imaging modalities, segmentation is more mature than diagnostic classification, and MRI-based tools are currently closer to clinical readiness than ultrasound—reflecting larger, multi-center datasets (e.g., Myo-Guide) and, in a few cases, external validation—whereas ultrasound AI remains constrained by less standardized image acquisition. Electrophysiology and Gait Analysis For electromyography (EMG) classification using ML reports, an accuracy ranging from 67% to 99.5% for distinguishing normal, ALS, and myopathy signals has been reported[43], although this wide range reflects heterogeneous study designs rather than a single robust benchmark. Gait analysis using AdaBoost classifiers reached 99.17% accuracy for identifying neurodegenerative conditions, including ALS [44], and VGG-19 networks applied to DMD ultrasound achieved 98.53% accuracy for ambulatory function classification [45]. These numbers are impressive but should be interpreted with caution: most studies used single-center data with limited external validation. Computational Pathology While digital pathology has been widely adopted in oncology, it remains almost untouched in neuromuscular disease (our landscape search identified only isolated proof-of-concept studies and no publicly available whole-slide imaging dataset for skeletal muscle; see Electronic Supplementary Material), a surprising gap given that muscle biopsy remains the diagnostic gold standard for inflammatory myopathies, metabolic myopathies, and congenital myopathies. Biopsy interpretation depends on a small number of specialized pathologists concentrated in academic centers, and this creates a genuine diagnostic inequity for patients without access to these centers. Kabeya et al. showed what is possible: using 4041 cropped images from 1400 hematoxylin/eosin-stained slides, their CNN classified 11 muscle diseases with an AUC of 0.996, outperforming nine physicians in a constrained evaluation [46]. However, this study also illustrates the field’s limitations. Publicly available whole-slide imaging datasets do not exist for skeletal muscle, forcing all models to work with small patch-level crops rather than in the whole-slide context that pathologists actually use for diagnosis. AI in Disease Monitoring and Prognosis Disease Trajectory and Survival Prediction Amyotrophic lateral sclerosis survival prediction has received far more attention than any other prognostic application in NMDs, for one simple reason: the PRO-ACT database provides over 10,700 patient records from 23 clinical trials (an unusually rich resource for a rare disease). DREAM Challenge algorithms trained on this database achieved C-indices of 0.80–0.81 [47, 48]. The ENCALS model took a different approach, using just eight clinical predictors validated across 14 European centers with 11,475 patients to achieve a C-index of 0.78 [49]. What makes ENCALS stand out is not its accuracy but its validation: it remains the only ALS survival model with both external validation and low risk of bias. Neuroimaging adds a complementary perspective; in one study, ML models identified brain atrophy patterns in the corpus callosum and thalamus as mortality predictors in 178 patients [50]. This availability of abundant data for ALS contrasts with most other NMDs, which lack both robust databases and validated models, underscoring how much infrastructure matters for AI progress. Wearable Digital Biomarkers Wearables offer a practical solution to the small-cohort problem: continuous data collection from each patient generates rich datasets even from small groups. The KineDMD study equipped patients with DMD with a 17-sensor bodysuit, producing movement-based behavioral fingerprints that outperformed standard clinical scales [51]. This led to the FDA/EMA qualification of the Stride Velocity 95th Centile (SV95C; Sysnav Healthcare, Vernon, France) as the first wearable-derived digital primary endpoint for DMD trials (ICC 0.970), reducing required sample sizes by increasing outcome sensitivity [52]. In ALS, wrist-worn accelerometers measuring upper limb activity correlate with ALSFRS-RSE (Amyotrophic Lateral Sclerosis Functional Rating Scale-Revised Self-Entry) scores [53], although regulatory qualification has not yet been pursued. Treatment Response Prediction Machine learning can also identify treatment response predictors from very small cohorts. Cerebrospinal fluid (CSF) levels of the microRNAs miR-206 and miR-133a-3p was found to predict nusinersen response in patients with spinal muscular atrophy (SMA) [54], a finding with direct cost implications given that annual therapy costs exceed $100,000. In another study, random forest models predicted SMA scoliosis with a receiver operator characteristic–area under the curve (ROC-AUC) of 0.85 across 86 patients [55], and electrical impedance myography with ML reached 93% accuracy for SMA subtype classification [56]. For myasthenia gravis, a systematic review of 11 studies found an AUC of 0.765–0.944 for crisis and intensive care unit (ICU) admission prediction [57]. AI In Biomarker Discovery Proteomics High-throughput proteomics paired with ML can extract diagnostic signatures from the high-dimensional, small-sample datasets typical of NMD research. Chia et al. used Olink Explore 3072 plasma proteomics in 183 patients with ALS versus 309 controls, achieving an AUC of 98.3% for diagnosis—replicated in an independent cohort—and showing that disease processes affect skeletal muscle years before symptom onset [58]. In DMD, profiling of 153 patients using the SomaScan 7 K proteomics assay identified 36 proteins predicting clinical milestones, including loss of ambulation [59]. In idiopathic inflammatory myopathies (IIM)—a group of acquired, immune-mediated NMDs—a LASSO-derived 5-protein panel yielded an AUC of 0.824 for detecting anti-MDA5 + dermatomyositis [60]. These proteomic approaches are notable because they work with the small cohort sizes inherent to rare diseases while still achieving clinically relevant discrimination. Transcriptomics and Multi-omics Integration These single-cell and multi-omics studies rely on ML methods—dimensionality reduction, unsupervised clustering, and latent-factor models—to resolve cellular heterogeneity, even when AI is not their headline framing. Single-cell technologies are beginning to reveal the cellular complexity underlying NMD pathology. In a DMD primate model, single-cell RNA sequencing (scRNA-seq) uncovered defects across fibro-adipogenic progenitors, muscle stem cells, and immune cells simultaneously [61], a breadth of pathology that bulk transcriptomics would have missed. Coulis et al. went a step further by combining scRNA-seq with spatial transcriptomics, pinpointing Galectin-3+ macrophages as key drivers of muscle fibrosis via SPP1/osteopontin signaling [62]. Perhaps the most instructive example is the MOFA-based (Multi-Omics Factor Analysis, an unsupervised ML method) multi-omics analysis of 203 patients with IIM, which integrated transcriptomic, proteomic, and metabolomic data from the same muscle samples to distinguish dermatomyositis from immune-mediated necrotizing myopathy and antisynthetase syndrome [63]. This study was validated in an independent cohort of 44 patients, a rarity in the field, and demonstrates that combining modalities can compensate for small individual datasets. Natural Language Processing and Clinical Phenotyping Much of the clinical data relevant to NMDs is locked in unstructured text (clinic notes, radiology reports, discharge summaries) and is therefore invisible to most ML pipelines. Natural language processing (NLP) tools are beginning to unlock this resource. Schrader et al. developed a claims-based algorithm that identifies patients with DMD from administrative data with 91.6% positive predictive value [64], which could be useful for cohort assembly and epidemiological studies. PhenoBrain, a BERT (bidirectional encoder representations from transformers)-based pipeline, outperformed both 50 specialist physicians and GPT-4 (generative pre-trained transformer 4) on rare disease diagnostic benchmarks (top-3 recall 0.513), and when combined with specialist input, recall rose to 0.768 [65]. General-purpose LLMs are also being tested: GPT-4o correctly ranked the diagnosis first in 19.8% of 4917 standardized rare-disease vignettes across ten languages [66]. These numbers remain modest: an 80% miss rate at rank-1 would not satisfy most clinicians, but they establish a baseline that is likely to improve as models are fine-tuned on domain-specific data and prospectively evaluated before clinical use. Disease-Specific AI Research Landscape Artificial intelligence research activity varies widely across NMD subtypes, with a heavy concentration in a few well-resourced conditions and near-complete absence in others. Broadly, three tiers emerge: conditions with a substantial AI literature (ALS, inflammatory myopathies, MG, and DMD/BMD), conditions with sparse but emerging activity (SMA, FSHD, LGMD, myotonic dystrophy, and Charcot–Marie–Tooth disease [CMT]), and conditions where disease-specific AI is essentially absent (metabolic and congenital myopathies and Lambert–Eaton myasthenic syndrome). This gradient tracks data availability more than clinical need and cuts across the genetic/acquired divide, since the well-studied group includes both inherited (DMD) and acquired (IIM, MG) diseases; the uneven depth of the disease-specific discussion below reflects the same gradient, as conditions with little or no AI literature can only be described briefly. Table 1 summarizes the current landscape. Search strategy and additional details are provided in the Electronic Supplementary Material.Table 1Artificial intelligence research landscape across neuromuscular disease subtypes between 2018 and 2025DiseasePublications (n)Key AI applicationsNotable gapsALS547Survival prediction, proteomics, speech analysis, wearables, brain imagingMost studied NMD; extensive PRO-ACT database availableIIM151IIM subtype classification, MRI analysis, multi-omics biomarkersLimited treatment response modelingMG100Crisis prediction (AUC 0.765–0.944), ICU risk stratificationNo wearable digital biomarkersDMD/BMD86MRI and ultrasound imaging, SV95C digital endpoint, cardiac monitoringLimited computational pathologySMA44Treatment response to nusinersen, scoliosis prediction, impedance myographyNo multi-omics integrationFSHD33Ultrasound Heckmatt grading, MRI segmentationLimited wearable studiesMetabolic myopathies19General applications onlyNear-absence of disease-specific AIMyotonic dystrophy16Cardiac MRI, omics approachesNo disease-specific AI diagnostic or monitoring toolsLGMD10MRI differential diagnosis (BMD versus LGMD2)Limited given extreme heterogeneityCongenital myopathies9General applications onlyNo specific ML models despite well-characterized pathological featuresCMT7Deep learning MRI segmentationWearables and gait analysis unexploredLambert–Eaton myasthenic syndrome1NoneCompletely unexploredAI Artificial intelligence, ALS amyotrophic lateral sclerosis, AUC area under the receiver operating characteristic curve, BMD Becker muscular dystrophy, CMT Charcot–Marie–Tooth disease, DMD Duchenne muscular dystrophy, FSHD facioscapulohumeral muscular dystrophy, ICU intensive care unit, IIM idiopathic inflammatory myopathies, LGMD limb-girdle muscular dystrophy, MG myasthenia gravis, ML machine learning, MRI magnetic resonance imaging, NMD neuromuscular disease, PRO-ACT Pooled Resource Open-Access ALS Clinical Trials, SMA spinal muscular atrophy, SV95C stride velocity 95th centile Table 1Artificial intelligence research landscape across neuromuscular disease subtypes between 2018 and 2025DiseasePublications (n)Key AI applicationsNotable gapsALS547Survival prediction, proteomics, speech analysis, wearables, brain imagingMost studied NMD; extensive PRO-ACT database availableIIM151IIM subtype classification, MRI analysis, multi-omics biomarkersLimited treatment response modelingMG100Crisis prediction (AUC 0.765–0.944), ICU risk stratificationNo wearable digital biomarkersDMD/BMD86MRI and ultrasound imaging, SV95C digital endpoint, cardiac monitoringLimited computational pathologySMA44Treatment response to nusinersen, scoliosis prediction, impedance myographyNo multi-omics integrationFSHD33Ultrasound Heckmatt grading, MRI segmentationLimited wearable studiesMetabolic myopathies19General applications onlyNear-absence of disease-specific AIMyotonic dystrophy16Cardiac MRI, omics approachesNo disease-specific AI diagnostic or monitoring toolsLGMD10MRI differential diagnosis (BMD versus LGMD2)Limited given extreme heterogeneityCongenital myopathies9General applications onlyNo specific ML models despite well-characterized pathological featuresCMT7Deep learning MRI segmentationWearables and gait analysis unexploredLambert–Eaton myasthenic syndrome1NoneCompletely unexploredAI Artificial intelligence, ALS amyotrophic lateral sclerosis, AUC area under the receiver operating characteristic curve, BMD Becker muscular dystrophy, CMT Charcot–Marie–Tooth disease, DMD Duchenne muscular dystrophy, FSHD facioscapulohumeral muscular dystrophy, ICU intensive care unit, IIM idiopathic inflammatory myopathies, LGMD limb-girdle muscular dystrophy, MG myasthenia gravis, ML machine learning, MRI magnetic resonance imaging, NMD neuromuscular disease, PRO-ACT Pooled Resource Open-Access ALS Clinical Trials, SMA spinal muscular atrophy, SV95C stride velocity 95th centile Artificial intelligence research landscape across neuromuscular disease subtypes between 2018 and 2025 AI Artificial intelligence, ALS amyotrophic lateral sclerosis, AUC area under the receiver operating characteristic curve, BMD Becker muscular dystrophy, CMT Charcot–Marie–Tooth disease, DMD Duchenne muscular dystrophy, FSHD facioscapulohumeral muscular dystrophy, ICU intensive care unit, IIM idiopathic inflammatory myopathies, LGMD limb-girdle muscular dystrophy, MG myasthenia gravis, ML machine learning, MRI magnetic resonance imaging, NMD neuromuscular disease, PRO-ACT Pooled Resource Open-Access ALS Clinical Trials, SMA spinal muscular atrophy, SV95C stride velocity 95th centile The concentration is not surprising: ALS benefits from the PRO-ACT database with over 10,700 patient records, and DMD profits from well-organized registries and strong advocacy organizations. Within ALS, voice analysis has reached clinical-grade performance (R2 = 0.92 for dysarthria assessment from smartphone recordings [67]); within DMD, quantitative ultrasound scatteromics achieved an AUC of 0.97–0.98 for ambulatory staging [68]. However, these are diseases that already had data infrastructure. The real question is whether AI methods can be made to work for conditions without such advantages. Myotonic dystrophy, with its cardiac, cognitive, and muscular involvement, would seem a natural fit for integrative AI approaches, yet it has no disease-specific algorithms. Congenital myopathies have distinctive histopathological features readily amenable to image classification, yet no one has built a model. Metabolic myopathies (Pompe disease, McArdle disease) and Lambert-Eaton syndrome are entirely absent from the AI literature. These differences in research intensity also reflect genuine differences in what each disease requires from AI, and the same tool is not equally useful across conditions. In DMD, the genetic diagnosis is straightforward and rarely the clinical bottleneck; the more pressing need is to capture and interpret natural history data for both monitoring and correlating motor function with therapeutic response, which is why wearable digital biomarkers and quantitative imaging have dominated this space. In ALS, by contrast, diagnosis and classification remain difficult because of marked clinical heterogeneity and persistent diagnostic delay, so detailed and standardized clinical phenotyping is a prerequisite rather than an optional refinement, and the most productive applications have accordingly been survival prediction and patient stratification. Inflammatory myopathies occupy an intermediate position, in which serological and histological subtyping is comparatively well developed whereas treatment response prediction is not. Recognizing these disease-specific requirements is essential when interpreting the studies summarized in Table 1, because a model that addresses a genuine clinical bottleneck in one disease may be redundant in another. Discussion The studies reviewed here suggest that data scarcity, long considered to be the defining obstacle for AI in rare diseases, may be less prohibitive than currently assumed. Foundation models pre-trained on millions of general-domain samples can be fine-tuned with remarkably small NMD-specific datasets [13, 17], and few-shot learning frameworks like SHEPHERD bypass the need for large annotated cohorts altogether [17]. What matters, in our view, is not the absolute size of the dataset but how creatively the available data are leveraged. The Myo-Guide platform illustrates this well: by pooling 2961 MRI scans from 26 centers across 20 disease types, it outperformed individual expert clinicians despite each contributing center having only a modest number of cases [38]. That said, we must be cautious about the headline performance figures reported in this literature. The accuracy rates of 91–99% seen across imaging and EMG studies are almost exclusively derived from internal validation on single-center datasets [69, 70]. External validation, when attempted, often reveals substantial performance drops. Only 11.8% of rare disease ML studies have tested their models on external data [10], and fewer still have undergone prospective clinical evaluation. The ENCALS survival model [49] stands out precisely because it was validated across 14 independent centers—a standard that should be the norm rather than the exception. It is therefore useful to read these applications through a translational-maturity lens: most published NMD models remain as exploratory algorithm development validated only internally; a smaller number have achieved external or multi-center validation (for example, the ENCALS survival model and ML Heckmatt ultrasound grading); and very few have reached prospective validation, clinical deployment, or regulatory qualification—the SV95C wearable endpoint being the principal exception. Current LLM/NLP applications should also be interpreted within this translational maturity framework, as most remain at the exploratory stage despite promising retrospective performance. High headline metrics (accuracy, AUC, Dice, or C-index) should not be equated with clinical usefulness, which additionally requires prospective validation, workflow integration, interpretability, and demonstrated benefit over current care. Taken together, these observations indicate that the major challenge facing AI in neuromuscular diseases is no longer algorithm development itself but rather the lack of rigorous validation, standardized evaluation frameworks, and high-quality multi-center datasets. Without addressing these issues, excellent algorithmic performance is unlikely to translate into meaningful clinical benefit. As documented in Table 1, the distribution of AI research across NMD subtypes is strikingly uneven, reflecting a self-reinforcing cycle. Diseases with existing data infrastructure (ALS with PRO-ACT, DMD with organized registries) attract more ML researchers, who then generate more data, further widening the gap. Breaking this cycle will require deliberate investment in data collection for neglected conditions, the transfer of learning from well-studied diseases to rare ones, and federated approaches that can pool small datasets across sites. Clinical trial design is another area where AI could have an outsized impact. Rare NMD trials routinely fail to reach enrollment targets because the patient population is simply too small. The regulatory qualification of SV95C as a wearable-derived digital endpoint for DMD [52] was a milestone because it addressed this problem directly: by increasing the sensitivity of outcome measures, it reduced the number of patients needed per trial. Extending similar digital biomarker approaches to ALS, MG, and SMA could be the single most practical near-term contribution of AI to the NMD field. A related barrier is that European rare disease databases remain largely siloed. Fewer than 17% of database custodians are willing to share data for ML purposes [71], citing governance, harmonization, and data quality concerns [71, 72]. Federated learning offers a practical workaround, enabling models to train across decentralized sites without raw data leaving each hospital [73], but its adoption in NMD research remains limited. On the regulatory front, the US Food and Drug Administration (FDA) has now authorized over 1000 AI/ML-enabled medical devices [74], and the European Medicines Agency (EMA) issued complementary guidance in September 2024 [75]. These frameworks provide a viable path to clinical deployment, although none have been applied specifically to NMD diagnostic tools to date. Beyond willingness to share, clinical deployment in rare diseases must navigate privacy regulation (notably the EU General Data Protection Regulation and the US Health Insurance Portability and Accountability Act), heterogeneous national governance, and the heightened re-identification risk that small cohorts carry; federated and synthetic data approaches mitigate, but do not fully resolve, these constraints. Equally fundamental is standardized clinical phenotyping, which we consider to be a prerequisite for the successful development and clinical implementation of AI in NMDs. Without standardized and structured phenotyping across registries, AI models inevitably learn from inconsistent labels, thereby limiting their reproducibility, generalizability, and clinical applicability. Therefore, coordinated and harmonized clinical data collection should be regarded as a fundamental prerequisite rather than a downstream consideration in AI research. In our opinion, fragmented data infrastructure has now become a greater obstacle than algorithmic innovation itself. Continued development of increasingly sophisticated AI models will provide only limited benefit unless data accessibility and interoperability improve simultaneously. How Future Studies Could be Improved by AI We see five priorities for the field going forward. These map onto four questions that we believe should guide the field: Which applications are closest to clinical implementation?Which applications are technically promising but clinically immature?What minimum validation standards should precede clinical adoption?Which NMD subfields most need targeted investment? Which applications are closest to clinical implementation? Which applications are closest to clinical implementation? Which applications are technically promising but clinically immature? Which applications are technically promising but clinically immature? What minimum validation standards should precede clinical adoption? What minimum validation standards should precede clinical adoption? Which NMD subfields most need targeted investment? Which NMD subfields most need targeted investment? Based on the limitations identified above, we believe that future research should prioritize not only technological innovation but also improvements in study design, data quality, and clinical validation. The most immediate need is to integrate existing registries into federated AI-ready networks. The infrastructure already exists: the TREAT-NMD global registry network spans 65 registries worldwide, the Solve-RD research project links over 300 clinicians across 51 sites re-analyzing more than 19,000 unsolved cases, and the DARWIN EU network of registries covers 160 million patient records across 16 European countries [76, 77]. What is missing is the technical layer that allows federated model training on top of these registries without requiring data centralization. The EU Horizon-funded CoMPaSS-NMD project illustrates this direction, applying AI-driven clustering to harmonized clinical, genetic, histological, and imaging data from five European countries to refine diagnosis and patient stratification in hereditary NMDs [78]. Large language models deserve attention as diagnostic aids, but with realistic expectations. ChatGPT-4o achieved a 13.3% diagnostic rate on Undiagnosed Diseases Network cases, approximately 2.4-fold the 5.6% rate achieved by traditional clinical review [79]. While these preliminary findings are encouraging, they also indicate that current LLM/NLP systems remain insufficiently accurate for routine clinical use, missing nearly 87% of these diagnostically challenging cases. Retrieval-augmented generation and multilingual pipelines may further improve diagnostic performance [66, 80]; however, rigorous prospective clinical evaluation remains essential before any clinical deployment. Within the translational maturity framework discussed above, current LLM/NLP applications should therefore be regarded as exploratory technologies that have demonstrated proof-of-concept performance but have not yet reached the level of prospective validation, clinical implementation, or regulatory maturity required for routine clinical practice. Wearable digital biomarkers should be extended beyond DMD. The SV95C qualification showed that continuous motion data can reduce required sample sizes for clinical trials by up to 50% [52]. Upper limb accelerometry for ALS, facial expression tracking for MG, and grip myotonia quantification for myotonic dystrophy are all technically feasible but lack the regulatory validation work that made SV95C successful. Computational pathology for muscle biopsies is, in our assessment, one of the largest missed opportunities. Muscle biopsy interpretation still depends on a small number of expert pathologists, and diagnostic criteria for conditions like perifascicular atrophy in dermatomyositis, rimmed vacuoles in inclusion body myositis, and fiber-type grouping in denervation are well-defined and amenable to automated detection. The bottleneck is the absence of publicly available, annotated whole-slide imaging datasets. As one step in this direction, our group has developed an open-access, non-commercial computational pathology toolkit (MyoToolkit; http://myotoolkit.huashanmuscle.com/) for automated muscle biopsy analysis, including fiber segmentation and morphometric quantification [81, 82]. Despite its considerable potential, progress in this area is currently constrained more by data availability than by methodological limitations, underscoring the urgent need for publicly accessible, expertly annotated pathology datasets. Finally, multi-omics integration should move from proof-of-concept to routine practice. The MOFA factor analysis model-based analysis of 203 patients with IIM [63] demonstrated that combining transcriptomic, proteomic, and metabolomic data reveals disease subtypes invisible to any single modality. As sequencing costs continue to fall, there is little justification for single-omics studies of diseases where tissue samples are already being collected. Although the current evidence remains heterogeneous and the depth of mechanistic understanding differs substantially among diseases, continued advances in high-throughput multi-omics technologies, AI-assisted analytical approaches, standardized multi-center studies, and international collaborative initiatives may progressively improve study design, data interpretation, and biomarker discovery in the coming years. Therefore, the current limitations should be viewed not only as challenges but also as opportunities for future research. In sum, data scarcity has not prevented meaningful AI progress in NMDs, but it has resulted in that progress being slow, uneven, and difficult to validate. The path forward requires less algorithmic novelty and more clinical infrastructure: shared datasets, federated networks, standardized endpoints, and prospective validation studies. Many of the building blocks are now in place, but most AI applications in NMDs still require prospective validation and clinical implementation before their benefit can be established. What the field needs now is the collaborative framework to put them to use. Supplementary Information Below is the link to the electronic supplementary material.Supplementary file1 (PDF 206 KB) Supplementary file1 (PDF 206 KB)