Artificial Intelligence and Neuromuscular Diseases: A Narrative Review

Abstract

Neuromuscular diseases are biologically diverse, clinically heterogeneous, and often difficult to diagnose and treat, highlighting the need for computational tools that can help resolve overlapping phenotypes and support timely, mechanism-informed interventions. This narrative review synthesizes recent advances in artificial intelligence (AI) and machine learning applied to neuromuscular diseases across diagnosis, outcome modeling, biomarker development, and therapeutics. AI-based approaches may assist clinical and genetic diagnosis from phenotypic data; however, early phenotype-driven tools have seen limited clinician adoption due to modest accuracy, usability challenges, and poor workflow integration. Electrophysiological studies remain central to diagnosing neuromuscular diseases, and AI shows promise for accurate classification of electrophysiological signals. Predictive models for disease outcome and progression—particularly in amyotrophic lateral sclerosis—are under active investigation, but most remain at an early stage of development and are not yet ready for routine clinical use. Digital biomarkers derived from imaging, gait, voice, and wearable sensors are emerging, with MRI-based quantification of muscle fat replacement representing the most mature and widely accepted application to date. Efforts to apply AI to therapeutic discovery, including drug repurposing and optimization of gene-based therapies, are ongoing but have thus far yielded limited clinical translation. Persistent barriers to broader adoption include disease rarity, data scarcity, heterogeneous acquisition protocols, inconsistent terminology, limited external validation, insufficient model explainability, and lack of seamless integration into clinical workflows. Addressing these challenges is essential to moving AI tools from the laboratory into clinical practice. We conclude with a practical checklist of considerations intended to guide the development and adoption of AI tools in neuromuscular disease care.

1. Introduction

Neuromuscular diseases represent a diverse and biologically complex group of neurological disorders [1,2] that affect the motor neuron, peripheral nerve, neuromuscular junction, or muscle (

Table 1. Representative neuromuscular diseases by site of pathology.

Disease Site	Abbreviation	Comments	Prevalence
Motor neuron
Amyotrophic lateral sclerosis	ALS	Progressive degeneration of upper and lower motor neurons	Rare
Proximal spinal muscular atrophy	SMA	Childhood-onset hereditary lower motor neuron disease	Rare
Axon
Charcot–Marie–Tooth disease (axonal)	CMT2	Axonal neuropathy	Rare
Diabetic distal symmetric polyneuropathy	DSPN	Length-dependent axonal polyneuropathy	Common
Myelin
Charcot–Marie–Tooth disease (demyelinating)	CMT1A	Hereditary demyelinating neuropathy	Uncommon
Chronic inflammatory demyelinating polyneuropathy	CIDP	Immune-mediated demyelinating neuropathy	Rare
Neuromuscular junction
Myasthenia gravis	MG	Autoimmune postsynaptic neuromuscular junction disorder	Uncommon
Lambert–Eaton myasthenic syndrome	LEMS	Autoimmune presynaptic neuromuscular junction disorder	Ultra-rare
Muscle
Duchenne muscular dystrophy	DMD	X-linked recessive dystrophinopathy	Rare
Myotonic dystrophy type 1	DM1	Autosomal dominant multisystem distal myopathy with myotonia	Rare

Notes. Prevalence categories follow approximate Orphanet-style bands: Common (>1000 per 1,000,000), Uncommon (100–900 per 1,000,000), Rare (1–90 per 1,000,000), and Ultra-rare (<1 per 1,000,000). The disorders listed here are representative examples for each anatomical locus of pathology rather than an exhaustive catalog. Apparent prevalence may differ from incidence depending on disease duration and survival (for example, MG has a relatively low incidence but higher prevalence, whereas amyotrophic lateral sclerosis has a higher incidence but lower prevalence due to short survival).

). A subset of neuromuscular diseases, particularly immune-mediated conditions such as myasthenia gravis and chronic inflammatory demyelinating polyneuropathy, has benefited from advances in immunotherapy developed largely through traditional drug discovery and clinical trial methods that did not rely on artificial intelligence [3,4,5,6]. In contrast, many partially or predominantly genetic neuromuscular diseases, including amyotrophic lateral sclerosis, the muscular dystrophies, and Charcot–Marie–Tooth disease, remain difficult to diagnose promptly and to treat effectively. This narrative review focuses on these difficult-to-diagnose and difficult-to-treat neuromuscular diseases, emphasizing how artificial intelligence and machine learning can expedite diagnosis and support the discovery of new therapies.

Most neuromuscular diseases are uncommon, with prevalences below 1% in the general population. Two notable exceptions are carpal tunnel syndrome and diabetic length-dependent polyneuropathy, which affect more than 1% of adults and are among the most common neurological conditions [7,8]. Many neuromuscular diseases have a neurogenetic etiology, although autoimmune, metabolic, toxic, traumatic, and nutritional causes also contribute (Table 1).

Despite their relative rarity, neuromuscular diseases account for a substantial share of rare and genetic disorders. Orphanet [9] lists 6171 rare diseases, of which approximately 522 (8.5%) are neuromuscular. The Online Mendelian Inheritance in Man (OMIM) [10] catalogs 9326 phenotypes with confirmed or suspected genetic etiologies. Manjunath et al. [11] identified 747 neuromuscular disease–associated genes and 1240 neuromuscular phenotypes, suggesting that roughly 12% of OMIM phenotypes are neuromuscular. Because pathogenic variants in a single gene may produce multiple clinically distinct diseases or marked phenotypic variability within a single disorder, the complexity of the phenotype space is greater than gene counts alone imply. For example, pathogenic variants in the TBK1 gene may manifest as amyotrophic lateral sclerosis, frontotemporal dementia, or a mixed amyotrophic lateral sclerosis–frontotemporal dementia syndrome, whereas pathogenic variants in the PMP22 gene may cause Charcot–Marie–Tooth disease type 1A (CMT1A) or hereditary neuropathy with liability to pressure palsies (HNPP).

Charcot–Marie–Tooth disease (CMT) illustrates the diagnostic complexity that arises within this phenotype space. Although most CMT variants share common features such as distal weakness, sensory loss, hyporeflexia, and muscle atrophy, OMIM lists more than 100 phenotypically distinct CMT entities, including CMT1, CMT2, CMT4, and X-linked CMT. More than 120 genes have been implicated in hereditary neuropathies, including PMP22, MPZ, GJB1, MFN2, NEFL, SH3TC2, and GDAP1. Single genes may produce multiple phenotypes. For example, the NEFL gene, which encodes the neurofilament light chain protein, is associated with at least three CMT phenotypes (CMT1F, CMT2E, and dominant intermediate CMT). ClinVar [12] lists 786 variants in the NEFL gene, 28 of which are reported as associated with CMT, spanning missense, nonsense, frameshift, splice-site, regulatory, and structural alterations.

These layers of complexity—multiple diseases per gene, multiple genes per phenotype, and many variants per gene—make accurate diagnosis challenging [13,14,15]. A single disorder may present with variable features, while the same signs or symptoms may arise from many distinct conditions. This crowded phenotypic space contributes to diagnostic delay [16,17].

Until recently, diagnostic challenges were paralleled by a lack of effective therapies. Many neuromuscular diseases were considered largely untreatable, and the field was characterized by therapeutic pessimism. For amyotrophic lateral sclerosis alone, more than 50 compounds that showed promise in preclinical models failed in human trials [18,19,20]. This gap between expanding biological insights into disease mechanisms and repeated therapeutic failures reinforced the perception that many neuromuscular diseases were not amenable to treatment.

This landscape is changing. The emergence of targeted therapies—including antisense oligonucleotides, gene-editing therapies, and precision small molecules—is creating treatment pathways for disorders previously considered untreatable [21]. These advances increase the need for diagnoses that are as accurate as the precision of emerging therapeutic strategies [16,22,23,24,25]. Artificial intelligence is viewed as an enabling technology that can accelerate the diagnosis and treatment of rare neuromuscular diseases [16,26]. In this narrative review, we summarize contributions of AI and ML to advances in four key areas:

1.

Diagnosis.

2.

Modeling of disease outcomes.

3.

Biomarkers for disease progression.

4.

New disease therapies.

2. Methods

This narrative review summarizes applications of artificial intelligence (AI) and machine learning (ML) to diagnosis, disease outcome modeling, biomarker development, and therapeutics in neuromuscular diseases. We did not follow a PRISMA workflow and did not conduct a formal risk-of-bias assessment. The literature search was completed on 23 December 2025, in MEDLINE (via Ovid) and IEEE Xplore. For MEDLINE (Ovid), the search string (Artificial Intelligence OR Machine Learning) AND Neuromuscular Diseases, filtered to Humans and English, yielded 306 records. For IEEE Xplore, the search string neuromuscular disease AND (artificial intelligence OR machine learning) yielded 159 records. No date restrictions were applied. Records were deduplicated, and titles and abstracts were screened for relevance to neuromuscular diseases and to AI or ML applications. Inclusion criteria were as follows:

1.

Focus on neuromuscular diseases or closely related disorders affecting the motor neuron, peripheral nerve, neuromuscular junction, or muscle.

2.

Use of artificial intelligence or machine-learning methods (including deep learning).

3.

A primary application to diagnosis, outcome modeling, biomarker development, or therapeutics.

Records were excluded if they were not relevant to neuromuscular disease, did not involve AI or ML, did not have sufficient methodological detail to allow interpretation, or did not meet the database filters (non-human or non-English). Additional relevant articles were included from two sources: (1) studies identified in the bibliographies of retrieved articles and (2) articles added to support specific assertions in the text (for example, historical context or widely cited methods). After screening and incorporation of these additional sources, 106 articles were retained for the narrative review.

3. Diagnosis of Neuromuscular Diseases

Artificial intelligence and machine learning are increasingly being explored as adjuncts to the diagnosis of neuromuscular diseases [16,21,27]. These efforts are driven by two interrelated goals:

1.

To improve diagnostic accuracy in neuromuscular diseases with overlapping clinical presentations—such as muscle weakness and atrophy—where conventional diagnostic approaches may be insufficient.

2.

To establish a precise genetic diagnosis, including the responsible gene, pathogenic variant, and genomic context, in neuromuscular diseases with a suspected neurogenetic basis.

Rosenberg et al. [15] provide a comprehensive overview of the epidemiology and diagnostic challenges of neuromuscular genetic disorders, emphasizing the difficulty of moving from heterogeneous clinical phenotypes to a definitive molecular diagnosis and motivating computational, phenotype-driven decision support. Ng et al. [14] review contemporary gene-based diagnostic strategies for neuromuscular disorders, outlining when to use targeted panels versus exome or genome sequencing and highlighting persistent challenges in variant interpretation and incomplete diagnostic yield. Krenn et al. [13] provide a neuromuscular-specific overview of contemporary genetic testing strategies and diagnostic yield, further underscoring the gap between phenotype and molecular diagnosis in routine practice.

Faviez et al. [28] reviewed 68 studies of computational methods for diagnosing rare diseases from patient phenotype data. Approximately two-thirds of the included studies used machine-learning-based approaches. They highlighted tools such as Phenolyzer [29], Phenomizer [30], and PhenoTips [31] as representative systems that support phenotype-driven diagnosis by prioritizing candidate genes and/or diseases from structured phenotype descriptions.

Since that review, Phen2Gene [32] has emerged as a phenotype-to-gene prioritization tool that links Human Phenotype Ontology (HPO) profiles to candidate causal genes. More recently, Gnanaolivu et al. [33] developed a graph-based gene-prioritization method that integrates phenotype information within a network framework; although broadly applicable to rare-disease diagnosis, their evaluation did not focus specifically on neuromuscular diseases.

3.1. Electrophysiological Diagnosis

Electrophysiological testing remains a mainstay in the diagnosis of neuromuscular disorders. Electromyographic examination (EMG) helps differentiate myopathic from neuropathic changes, while nerve conduction studies (NCS) distinguish demyelinating from axonal neuropathies and can reveal characteristic patterns—such as myotonic discharges—that are highly informative diagnostically. Interpretation of these studies is operator-dependent and requires substantial expertise, and although AI and machine-learning methods are under active investigation, they have not yet entered routine clinical use for EMG and NCS interpretation [16,34,35,36,37,38].

Early work focused on handcrafted time- and frequency-domain features from needle or surface electromyography and used classical machine-learning classifiers. Using features extracted from multi-muscle electromyography recordings and time–frequency transforms such as the Hilbert–Huang transform, Torres-Castillo et al. [39] demonstrated accurate discrimination among normal controls and several neuromuscular disease categories, illustrating the feasibility of automated electromyography-based screening in heterogeneous clinical populations. More recent studies employ deep-learning architectures that operate directly on raw or minimally processed electromyography signals, reducing the need for manual feature extraction and capturing waveform characteristics that may be difficult for human readers to appreciate [37]. Across studies summarized by Piñeros-Fernández [16] and by Taha and Morren [37], models trained on electrodiagnostic data report accuracies ranging from approximately 67% to 99% on selected classification tasks, including differentiation of amyotrophic lateral sclerosis from other motor neuron and myopathic disorders and discrimination between myopathic and neuropathic patterns [16,37]. In several series, model performance matched or exceeded that of experienced electromyographers, particularly for narrowly defined binary or ternary classification problems (

Table 2. EMG evidence table.

Study	Year	Disease	Modality	Class	N	Model	Validation	Accuracy	Notes
Torres [39]	2022	ALS, myopathy, healthy	Needle EMG	3-class	25	LDA, Tree, k-NN	3-fold CV	94.4%	SD, SC, NEV
Artug [40]	2014	ALS,  myopathy, healthy	Scanning EMG	3-class	150	SVM, k-NN, MLP, RBF	70/30 split	97.8%	SIM, SC
Benazzouz [41]	2019	ALS, myopathy, healthy	Public EMG	3-class	25	RF, k-NN	10-fold CV	88.8%	SD
Boutellaa [42]	2024	ALS,  myopathy,  healthy	Public EMG	3-class	25	CNN,  LSTM	Train, validate, test	99.1%	SD, NEV
Chandra [43]	2020	ALS, healthy, myopathy	EMG	Binary	25	SVM	10-fold CV	95–97%	SD, NEV
Cooray [44]	2025	Pediatric ICU	EMG, NCS	6-class	351	DNN	5-fold CV	95.2%	SD, CI, NEV
Emon [45]	2024	healthy, myopathy, neuropathy	Public EMG	Binary	200	k-NN, SVM	5-fold CV	93.3–98.5%	SD, PDS, NEV
Kalwa [46]	2015	ALS, myopathy, healthy	Public EMG	3-class	150	DWT, k-NN	None	67%	SD, PDS, NEV, LA
Martinez [47]	2025	ALS screening	F-wave	Binary	1378	AutoML, GBM	Train, test + CV	88–89%	SC, SD, NEV
Pino [48]	2008	Simulated EMG	Needle EMG	3-class	500	LDA, NB	Simulated	92%	SIM, NEV
Samanta [49]	2020	ALS, myopathy, healthy	Needle EMG	Binary	25	ResNet50, SVM	5-fold CV	88–100%	PDS, NEV
Somani [50]	2022	ALS, myopathy, healthy	Needle EMG	3-class	177	DW-kNN	None	∼95%	SC, NEV
Xie [51]	2004	ALS, myopathy, healthy	Needle EMG	3-class	80	BP-NN	Train, test	88.6%	SC, SD, NEV
Zhang [52]	2013	ALS vs. healthy	Surface EMG	Binary	21	LDA	LOOCV	Sens. 90%, Spec. 100%	SD, SC, NEV

Abbreviations: ALS, amyotrophic lateral sclerosis; BP-NN, backpropagation neural network; CI, class imbalance; CV, cross-validation; DNN, deep neural network; DWT, discrete wavelet transform; EMG, electromyography; GBM, gradient boosting machine; k-NN, k-nearest neighbors; LA, limited annotation; LDA, linear discriminant analysis; MLP, multilayer perceptron; NB, naive Bayes; NEV, no external validation; NCS, nerve conduction studies; PDS, public dataset; RBF, radial basis function network; RF, random forest; SC, single center; SD, small dataset; SIM, simulated data.

). Beyond categorical diagnosis, AI has been used to quantify disease severity and track progression based on electromyography data. Regression models linking motor unit parameters or interference patterns to clinical scales, as well as unsupervised approaches that cluster electromyography recordings into latent disease states, have shown promise for creating more objective electrodiagnostic biomarkers [37]. However, most current electromyography-based AI systems are confined to single-center datasets. Data are often collected under tightly controlled acquisition protocols, and validation across different laboratories, machines, and operator techniques is limited. Standardization of acquisition and annotation, together with prospective evaluation in routine electrodiagnostic practice, will be necessary before these tools can be implemented reliably in clinical settings.

A review of the FDA 510(k) database found no recently cleared AI or ML electromyographic applications whose primary indication is the diagnosis of neuromuscular diseases. In 2008, the FDA cleared the NeuroMetrix ADVANCE system for nerve conduction studies and needle electromyography, including computer-assisted quantitative analysis of EMG and NCS signals (https://www.accessdata.fda.gov/cdrh_docs/pdf7/K070109.pdf, accessed on 26 December 2025). The clearance is for signal analysis and quantitation and not for disease diagnosis.

3.2. Diagnosis by Muscle Biopsy

Due to advances in genetics and laboratory testing, patients with suspected myopathy can often be diagnosed without a muscle biopsy. Kleinveld et al. [53] estimated that a muscle biopsy was informative in only about 20% of cases when performed. Nonetheless, muscle biopsies are still performed in selected, difficult-to-diagnose cases [54]. Recent work has applied AI to support disease classification and quantitative morphometry in muscle pathology.

Scodellaro et al. [26] developed an explainable AI workflow for muscle biopsy interpretation using label-free multiphoton microscopy images and a single CNN trained with either standard backpropagation (BP) or the forward-forward (FF) algorithm. They assembled 1600 images from 16 human biopsies and enforced a patient-level split (12 patients for training/validation; 4 held out for testing) with 4-fold cross-validation. For the primary classification task (Duchenne muscular dystrophy versus healthy), they reported high image-level performance (BP accuracy 97.8%; FF accuracy 90.7%). A key contribution was interpretability: class activation maps (CAMs) suggested that BP and FF attended to partially distinct histopathologic patterns; both highlighted intramuscular collagen, whereas BP additionally emphasized collagen waviness and fat and FF emphasized collagen density.

Ono et al. [55] applied AI to muscle morphometry in neuropathic and myopathic specimens using whole-slide H&E images. The dataset consisted of 49 myopathic specimens, 19 neuropathic specimens, and 20 normal controls. A fine-tuned YOLOv8 segmentation model was used to segment individual fibers and compute morphometric measures (cross-sectional area and circularity), along with graph-based spatial features (e.g., grouped atrophy). They then trained a LightGBM (tree-ensemble) classifier to label specimens as myopathic or neuropathic, achieving 85.2% accuracy (AUC 0.817), compared with 80.8% for human annotation. Three neuropathic specimens (ALS) were misclassified as myopathic, but these cases were also difficult for human annotators. The investigators note small sample size, class imbalance, and single-center specimens as key limitations.

These studies illustrate the promise of AI-assisted muscle pathology; however, limited cohort sizes and restricted external validation support interpreting the results as investigational rather than ready for routine clinical use, particularly in the context of declining use of muscle biopsy for neuromuscular disorders.

4. Modeling Disease Outcomes

Accurately measuring disease progression and forecasting outcomes are central to counseling patients, designing clinical trials, and allocating resources in neuromuscular medicine. In amyotrophic lateral sclerosis (ALS), recent AI and ML models trained on large, harmonized datasets such as PRO-ACT have reported improved predictive performance for selected tasks [56]. For example, optimized models—most notably extreme gradient boosting—achieved a sensitivity of 100%, specificity of 97.4%, accuracy of 98.0%, F1-score of 96.0%, Matthews correlation coefficient of 94.1%, and an AUC of 0.96 in distinguishing bulbar-onset from limb-onset ALS [56]. Tree-based ensemble methods iteratively correct residual errors and are well-suited to capturing nonlinear relationships in heterogeneous clinical data.

Table 3 summarizes studies that use AI and ML to model outcomes in neuromuscular disease. Many studies used the ALS PRO-ACT dataset, which includes more than 12,000 patients (https://ncri1.partners.org/PROACT, accessed on 26 December 2025). Outcomes included ALSFRS slope, fast versus slow progressors, 12-month survival, and bulbar versus limb onset. Despite the large dataset size, predictive performance was often modest as assessed by accuracy, Pearson r, AUC, or RMSE.

Blemker et al. [57] present an engineering-grade clinical progression model for facioscapulohumeral muscular dystrophy (FSHD) using a multiscale, stacked random-forest framework that links baseline whole-body mDixon MRI segmentations and clinical measures to longitudinal outcomes across imaging and function. Rather than treating each patient as a single datapoint, they exploit the structure of MRI by modeling progression at regional and muscle levels (fat fraction and lean muscle volume) and then propagating these intermediate predictions to a functional endpoint (timed up-and-go), effectively creating a hierarchical digital-twin pipeline. A key innovation is their cross-study, multi-site aggregation strategy to increase sample size and heterogeneity, paired with explainability (SHAP) to identify drivers of predicted progression. They report that measures of fat-distribution heterogeneity can carry prognostic signal beyond baseline averages. After pooling data from more than 100 patients across seven studies, they predicted regional fat fraction change with RMSE ≈ 3.5%, muscle-level fat fraction change with RMSE ≈ 2.2% (Pearson $r=0.5$), and lean muscle change with RMSE ≈ 8.3 mL, as well as change in timed up-and-go with test RMSE ≈ 0.75 s [57].

Katz et al. [58] used a random forest model to predict wheelchair status in patients with FSHD with a modest accuracy of 78%. For sporadic inclusion body myositis, Alfano et al. [59] used a small dataset and reported modest predictive performance for chair-rise ability and 2-minute walk distance. Overall, outcome modeling in neuromuscular disease has been modestly successful to date. Future work may move toward more accurate models based on digital-twin approaches for neuromuscular diseases [57,60,61].

Table 3. Outcome model evidence table.

Author	Disease	Year	Outcome	Data	N	Method	Prediction	Result
Qin [56]	ALS	2025	Bulbar vs. limb onset	PRO-ACT	NA	XGBoost	Binary	Accuracy = 98.0%
Qin [56]	ALS	2025	ALSFRS slope	PRO-ACT	NA	XGBoost, LGB, 1D CNN/DNN	Regression	RMSE RMSE ≈ 4.6–5.6
Blemker [57]	FSHD	2025	Fat fraction, lean volume	Pooled (7 studies)	>100	RF	Regression	RMSE ≈ 2.2%, RMSE ≈ 8.1 ml
Blemker [57]	FSHD	2025	TUG change	Pooled (7 studies)	>100	RF	Regression	RMSE = 0.6 s
Katz [58]	FSHD	2021	Time to wheelchair	Registry	578	RF	Binary	Accuracy = 0.78; AUC = 0.85
Alfano [59]	sIBM	2017	Chair rise ability	Internal	55	LogReg	Binary	AUC = 0.74
Alfano [59]	sIBM	2017	2-min walk distance	Internal	55	LinReg	Regression	$R^2$ = 0.53
Guo [62]	ALS	2025	12-month survival	PRO-ACT	1941	XGBoost, RP	Binary	AUC = 0.71–0.82
Jabbar [63]	ALS	2024	Fast vs. slow progressors	PRO-ACT	5030	XGBoost, BLSTM	Binary	AUC = 0.57–0.75
Turabeih [64]	ALS	2024	ALSFRS slope	PRO-ACT	2649	GBM	Regression	RMSE = 0.560
Tang [65]	ALS	2018	ALSFRS slope	PRO-ACT	∼8000	BART, RF	Regression	r = 0.43–0.55
Pancotti [66]	ALS	2022	ALSFRS slope	PRO-ACT	2921	FFNN, CNN, RNN	Regression	RMSE ≈ 0.52; $r\approx 0.45$
Al-Bdairat [67]	ALS	2025	ALSFRS slope	PRO-ACT	NA	DNN, LSTM	Regression	MSE ≈ 0.32

Note: Outcome prediction tasks are reported as binary classification or regression depending on study design. Performance metrics are reported as presented by the original authors. Abbreviations: ALS, amyotrophic lateral sclerosis; ALSFRS, ALS Functional Rating Scale; AUC/AUROC, area under the receiver operating characteristic curve; BART, Bayesian additive regression trees; BLSTM, bidirectional long short-term memory; CNN, convolutional neural network; DNN, deep neural network; FFNN, feed-forward neural network; FSHD, facioscapulohumeral muscular dystrophy; GBM, gradient boosting machine; LGB, Light GBM; LinReg, linear regression; LogReg, logistic regression; LSTM, long short-term memory; MSE, mean squared error; NA, not available; PRO-ACT, Pooled Resource Open-Access ALS Clinical Trials database; RF, random forest; RMSE, root mean squared error; RP, regularized predictors; sIBM, sporadic inclusion body myositis; TUG, timed up-and-go test; XGBoost, extreme gradient boosting.

Table 3. Outcome model evidence table.

Author	Disease	Year	Outcome	Data	N	Method	Prediction	Result
Qin [56]	ALS	2025	Bulbar vs. limb onset	PRO-ACT	NA	XGBoost	Binary	Accuracy = 98.0%
Qin [56]	ALS	2025	ALSFRS slope	PRO-ACT	NA	XGBoost, LGB, 1D CNN/DNN	Regression	RMSE RMSE ≈ 4.6–5.6
Blemker [57]	FSHD	2025	Fat fraction, lean volume	Pooled (7 studies)	>100	RF	Regression	RMSE ≈ 2.2%, RMSE ≈ 8.1 ml
Blemker [57]	FSHD	2025	TUG change	Pooled (7 studies)	>100	RF	Regression	RMSE = 0.6 s
Katz [58]	FSHD	2021	Time to wheelchair	Registry	578	RF	Binary	Accuracy = 0.78; AUC = 0.85
Alfano [59]	sIBM	2017	Chair rise ability	Internal	55	LogReg	Binary	AUC = 0.74
Alfano [59]	sIBM	2017	2-min walk distance	Internal	55	LinReg	Regression	$R^2$ = 0.53
Guo [62]	ALS	2025	12-month survival	PRO-ACT	1941	XGBoost, RP	Binary	AUC = 0.71–0.82
Jabbar [63]	ALS	2024	Fast vs. slow progressors	PRO-ACT	5030	XGBoost, BLSTM	Binary	AUC = 0.57–0.75
Turabeih [64]	ALS	2024	ALSFRS slope	PRO-ACT	2649	GBM	Regression	RMSE = 0.560
Tang [65]	ALS	2018	ALSFRS slope	PRO-ACT	∼8000	BART, RF	Regression	r = 0.43–0.55
Pancotti [66]	ALS	2022	ALSFRS slope	PRO-ACT	2921	FFNN, CNN, RNN	Regression	RMSE ≈ 0.52; $r\approx 0.45$
Al-Bdairat [67]	ALS	2025	ALSFRS slope	PRO-ACT	NA	DNN, LSTM	Regression	MSE ≈ 0.32

Note: Outcome prediction tasks are reported as binary classification or regression depending on study design. Performance metrics are reported as presented by the original authors. Abbreviations: ALS, amyotrophic lateral sclerosis; ALSFRS, ALS Functional Rating Scale; AUC/AUROC, area under the receiver operating characteristic curve; BART, Bayesian additive regression trees; BLSTM, bidirectional long short-term memory; CNN, convolutional neural network; DNN, deep neural network; FFNN, feed-forward neural network; FSHD, facioscapulohumeral muscular dystrophy; GBM, gradient boosting machine; LGB, Light GBM; LinReg, linear regression; LogReg, logistic regression; LSTM, long short-term memory; MSE, mean squared error; NA, not available; PRO-ACT, Pooled Resource Open-Access ALS Clinical Trials database; RF, random forest; RMSE, root mean squared error; RP, regularized predictors; sIBM, sporadic inclusion body myositis; TUG, timed up-and-go test; XGBoost, extreme gradient boosting.

5. Biomarkers for Disease Progression

Biomarkers play a central role in neuromuscular disease by supporting both initial diagnosis and the longitudinal assessment of disease severity and progression [68,69]. In this section, we focus on biomarkers actively evaluated for monitoring disease burden and progression in neuromuscular disorders. Imaging modalities—particularly magnetic resonance imaging (MRI) and ultrasound—have emerged as leading candidates for quantitative and semi-quantitative biomarkers of muscle involvement and disease progression [70,71,72,73,74]. Additional candidate biomarkers include measures derived from gait analysis, voice recordings, wearable actigraphy, and structured or semi-quantitative neurological examinations, many of which are well-suited to AI-enabled analysis.

5.1. MRI

Machine learning applied to muscle MRI has shown promise for differentiating phenotypically similar neuromuscular diseases and for quantifying disease severity and progression. In an early diagnostic application, Chen et al. [75] applied radiomics-based machine learning to thigh Dixon MRI to distinguish Duchenne muscular dystrophy (DMD) from Becker muscular dystrophy (BMD). In a cohort of 62 patients, models using logistic regression, k-nearest neighbors, support vector machines, multilayer perceptrons, and random forests achieved classification accuracies of 81.2–90.6%, compared with 69.4% for expert radiologists. Specificity improved substantially (71.0–86.0% vs. 19.0%), while sensitivity remained high (85.6–95.0% vs. 95.1%), with F1-scores ranging from 85.2 to 92.6%. These results suggest that MRI radiomics combined with machine learning can improve early differentiation between closely related dystrophinopathies.

Beyond classification, the primary clinical value of MRI in neuromuscular disease lies in assessing disease severity and progression by quantifying muscle fat replacement (

Table 4. MRI evidence table.

Author	Year	Disease	Outcome	MRI	Task	N	Method	Result
Chen [75]	2025	DMD, BMD	Disease class	3T, T2 Dixon	Classify	62	U-Net, GLCM, GLDM, GLSZM	Accuracy = 81.2–90.6%
Fantacci [77]	2016	NMD	Estimated non-muscle	1.5T, T1	Quantify	26	Rule-based segmentation, histogram modeling	Spearman $\rho =0.97$
Gadermayr [78]	2019	Myopathy	Muscle segmentation	1.5T, T1	Segment	41	FCNN	Dice = 0.91
Huysmans [76]	2025	LGMD, BMD, DM1, CMT1A, HC	Fat fraction, disease class	1.5T/3T Dixon	Multi-class	156	U-Net CNN, RF, SHAP	5-class accuracy = 89%; AUC ≈ 0.96–0.99
Marfisi [79]	2019	Mixed NMD	Fat infiltration, disease severity	T1	Quantify	46	Fuzzy c-means clustering	No ground truth comparison
O’Donnell [80]	2024	CMT1A	Muscle segmentation	T1, Dixon	Segment	27	MuscleSense (U-Net, CNN)	No ground truth comparison

Abbreviations: BMD, Beckermuscular dystrophy; CMT1A, Charcot–Marie–Tooth disease type 1A; CNN, convolutional neural network; Dice, Dice similarity coefficient; DM1, myotonic dystrophy type 1; DMD, Duchenne muscular dystrophy; FCM, fuzzy c-means; FCNN, fully convolutional neural network; GLCM/GLDM/GLSZM, gray-level radiomicmatrices; HC, healthy controls; LGMD, limb-girdlemuscular dystrophy; MRI, magnetic resonance imaging; NMD, neuromuscular disease; RF, randomforest; SHAP, SHapley Additive exPlanations.

). AI and ML methods have been widely applied to automate muscle segmentation, estimate fat fraction, and derive quantitative imaging biomarkers. Across studies, AI-assisted MRI analyses achieve high performance for fat-fraction estimation and disease stratification, with reported AUC values ranging from 0.96 to 0.99 [76].

De Wel et al. [81] provided biological validation by using muscle biopsy as ground truth for AI-assisted MRI analysis in limb–girdle muscular dystrophy R12 (LGMDR12). They studied 27 adults with LGMDR12 and 27 matched controls, with a biopsy subset of 16 patients and 15 controls. Using a custom convolutional neural network for semi-automated three-dimensional segmentation of multi-stack thigh Dixon MRI, they computed proton density fat fraction (PDFF) for individual muscles along their full length. MRI-derived PDFF showed strong concordance with biopsy-measured fat fraction in affected muscles (e.g., semimembranosus $r=0.85$, vastus lateralis $r=0.68$), supporting MRI PDFF—enabled by AI-based segmentation—as a quantitative biomarker of neuromuscular disease severity and progression. AI-assisted muscle MRI quantification has also begun to enter regulated clinical practice. Springbok’s MuscleView 2.0 is an FDA 510(k)-cleared AI software (K251682, accessed on 20 January 2026 at https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfPMN/pmn.cfm?ID=K251682) that automatically segments lower-extremity muscles and bones on MRI and provides quantitative metrics such as muscle volume and intramuscular fat percentage, with clearance for image segmentation and quantification rather than a stand-alone diagnostic indication.

5.2. Ultrasound

The role of ultrasound in the diagnosis and management of neuromuscular diseases remains unsettled. Van Alfen et al. [82] emphasize that muscle ultrasound is clinically useful for diagnosing neuromuscular disease, highlight quantitative muscle ultrasound (QMUS) as the most sensitive approach, and note that software- and hardware-dependence limit standardization and widespread adoption, motivating newer approaches (including deep learning) to address current limitations. Hobson-Webb [70] reviews emerging technologies aimed at addressing key limitations of neuromuscular ultrasound—especially operator dependency and the lack of a viable contrast agent—including ultra-high-frequency ultrasound (potential fascicle-level imaging), shear-wave elastography (stiffness as a surrogate of tissue health), photoacoustic imaging (functional/contrast-like information such as inflammation), and AI as a route to improved standardization and diagnostic performance; however, she notes that these methods are not yet ready for widespread routine use. Vicino et al. [71] conclude that muscle ultrasound has a role partly overlapping with—and complementary to—MRI, and that current evidence supports its use as a diagnostic and follow-up marker across myopathies (including skeletal, facial, and respiratory muscles). They specifically flag shear-wave elastography and AI decision-support algorithms as active areas of development.

Marzola et al. [83] used deep-learning segmentation models (U-Net, U-Net++, FPN, Attention U-Net, and an ensemble) to automatically segment muscle cross-sectional area (CSA) in transverse ultrasound images. They then quantified echogenicity as the mean grayscale within the CSA and computed an adjusted z-score (using a regression-predicted grayscale value based on age, sex, weight, length, and BMI). Muscles were labeled normal or abnormal using the clinical rule z-score $>2$, so the model’s classification reflects agreement with this echogenicity-based threshold rather than disease-specific diagnosis. In a large retrospective dataset (3917 images from 1283 subjects; tibialis anterior, gastrocnemius medialis, and biceps brachii), they reported high normal/abnormal classification accuracy across splits: 98.1% (train), 99.1% (validation), and 97.0% (test).

The use of AI to enhance interpretation of muscle ultrasound remains a developing technology in neuromuscular disease. As of 2025, FDA-cleared AI ultrasound tools in the musculoskeletal domain support image acquisition and quantification but do not carry labeled indications for diagnosing neuromuscular muscle diseases such as muscular dystrophies or myopathies.

5.3. Other Biomarkers of Disease Status

Beyond imaging-based biomarkers, wearable sensors, smartphone-based assessments, and home-based devices enable continuous, high-frequency measurement of neuromuscular function in real-world settings. These technologies can capture ecologically valid signals—including gait, balance, voice, facial expression, and limb movement—that may be more sensitive to early change than episodic clinic-based assessments. In several rare neuromuscular diseases, AI models applied to these data streams have yielded candidate digital biomarkers that correlate with, and in some cases outperform, traditional clinical rating scales.

Gait analysis provides a prominent example. In dystrophinopathies and spinal muscular atrophy, inertial measurement units (IMUs) placed on the lower limbs or trunk, as well as computer-vision-based pose estimation from video, have been used to extract stride length, cadence, joint angles, and symmetry indices [84,85]. Supervised learning models trained on these features can distinguish affected individuals from controls, differentiate ambulatory stages, and detect subtle longitudinal changes that may precede clinically apparent decline. Because these measures can be collected at home, they can provide a more granular view of disease dynamics than clinic-based six-minute walk tests or timed functional assessments.

Voice and speech represent similarly rich sources of digital biomarkers, particularly in amyotrophic lateral sclerosis and bulbar-onset motor neuron disease. Deep-learning models operating on acoustic features (e.g., jitter, shimmer, formant trajectories, spectral envelopes) or directly on raw audio waveforms have been used to classify ALS versus controls, estimate disease severity, and track progression over time [86,87,88]. In some studies, voice-based models detect early bulbar involvement before substantial changes in ALS Functional Rating Scale–Revised (ALSFRS-R) scores, suggesting potential utility for early monitoring in genetically at-risk individuals.

Actigraphy and multi-sensor wearable platforms have been used to quantify overall activity levels, sleep–wake patterns, and limb-use asymmetries in hereditary neuropathies and inflammatory myopathies. Machine learning applied to these longitudinal time series can segment daily behavior into activity states, detect deviations associated with exacerbations or treatment response, and generate composite progression indices that correlate with clinician-rated scales. Given the density and noise inherent in these data, AI methods are particularly valuable for feature extraction, anomaly detection, and robust within-subject modeling [89,90,91,92,93,94].

Conte et al. [95] proposed NeuroExam as a digital platform for neuromuscular diseases that transforms subjective bedside strength and functional assessments into standardized, sensor-based quantitative measures. By reducing inter-observer variability in neurological examination, such approaches can yield more reliable longitudinal trajectories of weakness and hypotonia, which may be particularly important for slowly progressive congenital and inherited myopathies. The authors further suggest that these exam-derived features can be integrated with genetic and pathophysiological data to construct composite biomarkers of disease severity and progression.

For rare neuromuscular diseases, digital biomarkers offer several advantages: reduced reliance on frequent in-person visits to specialized centers, increased statistical power through dense longitudinal sampling, and support for decentralized or hybrid clinical trial designs. However, important challenges remain, including device heterogeneity, variable patient adherence, uncertainty regarding minimal clinically important differences for digital endpoints, and limited regulatory experience with AI-derived biomarkers. Carefully designed validation studies that link digital measures to established clinical and functional outcomes will be essential before these tools can be adopted as primary or surrogate endpoints in neuromuscular disease trials [86,96].

6. Therapeutics for Neuromuscular Diseases

Although still early, artificial intelligence and machine learning are being applied to identify drug repurposing candidates for neuromuscular disorders and to support the development and optimization of gene- and RNA-based therapies (Table 5).

6.1. Computational Approaches to Drug Repurposing for Neuromuscular Diseases

Yu et al. [97] developed a network-medicine approach to identify candidates for drug repurposing in ALS. They started from ALS-associated SNPs identified by genome-wide association studies (GWAS) and mapped these variants onto brain multi-omics quantitative trait loci (x-QTLs), including expression (eQTL), protein (pQTL), splicing (sQTL), methylation (meQTL), and histone acetylation (haQTL) QTLs, to link risk loci to putative target genes. Using a graph neural network applied to a large human protein–protein interaction (PPI) network, they prioritized 105 ALS-associated genes and defined an ALS “disease module” in the interactome. Overlaying known drug–target profiles onto this module via network-proximity analysis, they nominated several repurposing candidates, including diazoxide, gefitinib, paliperidone, and dimethyltryptamine (DMT). Although these agents have supportive preclinical and mechanistic evidence, they remain hypothesis-generating and, to date, lack dedicated ALS clinical trials; thus, the strategy is promising but investigational.

In a different AI approach to identifying drug targets for ALS, Pun et al. [98] used PandaOmics, an AI-enabled target-discovery platform, to move from ALS multi-omics data to druggable targets suitable for repurposing. They integrated RNA-seq datasets from post-mortem CNS tissue and iPSC-derived motor neurons with more than 20 AI and bioinformatics scoring models combining differential expression, pathway context, text mining, clinical and funding evidence, and expert (KOL) metrics to rank genes for ALS relevance. By applying druggability filters (e.g., known small-molecule ligands, non-essential genes), they partitioned ranked genes into “novel” biology targets and “high-confidence, druggable” targets that can be linked to existing compounds, thereby yielding repurposing candidates without de novo chemistry. They functionally validated a subset of targets in a C9orf72-mediated ALS Drosophila model, reporting that knockdown of several AI-prioritized genes rescued neurodegeneration. Overall, they identified 28 potential therapeutic targets: 17 high-confidence and 11 novel targets.

Sunildutt et al. [99] used a transcriptomics-driven Connectivity Map plus docking strategy to identify drug candidates for sporadic ALS. They started from a microarray dataset of lumbar spinal cord gray matter, identified differentially expressed genes (180 upregulated, 213 downregulated), and built a protein–protein interaction network (STRING) with pathway enrichment (DAVID, KEGG, Reactome) to define key ALS-related genes and pathways. The up- and downregulated gene sets were then submitted to the Clue Connectivity Map to identify small molecules whose expression signatures were negatively correlated with the ALS signature, on the premise that such drugs could reverse disease-associated transcriptional changes. They selected the top 50 negatively scored drugs and narrowed these to 10 candidates via literature review. Finally, they selected 10 key ALS genes and performed molecular docking in AutoDock Vina (1.2.0) between these proteins and the 10 drugs, ranking compounds by binding affinity and yielding nine repurposing candidates (cefaclor, diphenidol, flubendazole, fluticasone, lestaurtinib, nadolol, phenamil, temozolomide, tolterodine). They emphasized lestaurtinib (high affinity to multiple targets) and NOS3 (reported to interact with all shortlisted drugs).

Gerring et al. [100] used a genetics-led enrichment strategy combining common and rare ALS risk variants to prioritize drug classes for repurposing. Using ALS GWAS and exome-wide association data (Project MinE; https://projectmine.com/), they converted variant-level associations into gene-level scores using positional tests (e.g., MAGMA, mBAT-combo) and expression-based approaches (e.g., TWAS, SMR) in brain and blood, further weighting results by colocalization metrics (including HEIDI) to prioritize genes with stronger evidence that genetically regulated expression influences ALS risk. In parallel, they used rare-variant burden tests on exonic whole-genome sequencing data to identify genes enriched for disruptive or damaging variants. They then applied GSEA-style enrichment to test whether Anatomical Therapeutic Chemical (ATC) drug classes (from DrugBank) were overrepresented among top-ranked ALS risk genes. Their analysis highlighted B-Raf/RAF–MAPK pathway inhibition (e.g., BRAF inhibitors such as vemurafenib, dabrafenib, and encorafenib) from the common-variant pipeline and vitamin B-related targets (e.g., cobalamin and niacin classes) from the rare-variant pipeline. They annotated candidates with blood–brain barrier penetration and existing trial data and examined signals in ALS iPSC-derived motor-neuron transcriptomes, reporting patterns suggesting relevance of BRAF inhibitors in C9orf72 lines and vitamin B signaling in SOD1 lines. Hoolachan et al. [101] used a prednisolone-anchored, transcriptomics-based repositioning strategy aimed at SMN-independent muscle therapies in spinal muscular atrophy (SMA). They treated severe SMA and control mice with prednisolone, performed bulk RNA-seq on skeletal muscle, and identified genes and pathways shifted toward normal by prednisolone. Using these “prednisolone-restored” signatures, they queried drug–gene expression resources to identify approved non-glucocorticoid compounds predicted to mimic prednisolone’s transcriptomic effects, yielding 580 candidates and a prioritized set of 20 commercially available drugs. Two orally available agents, metformin and oxandrolone, were then tested in SMA cellular systems and the Smn2B^−/− mouse model; although both showed some benefit in C. elegans SMA models and oxandrolone conferred modest survival improvement in mice, neither reproduced prednisolone’s robust effects, and high-dose metformin worsened outcomes. These results highlight both the promise and limitations of transcriptomic similarity-based repurposing approaches for SMA muscle pathology.

6.2. Artificial Intelligence-Enabled Gene Therapies for Neuromuscular Diseases

In an editorial, Erdougan [102] suggested that genome editing (especially CRISPR/Cas9) offers treatment potential for many genetic disorders, explicitly listing Duchenne muscular dystrophy, spinal muscular atrophy, and ALS among neuromuscular indications, while emphasizing persistent challenges such as off-target effects, delivery, and variable efficiency. He argued that AI/ML models (e.g., DeepCRISPR, CRISTA, DeepHF) can improve sgRNA design, predict off-target and on-target activity, support selection among base, prime, and epigenome editing strategies, and optimize expression and distribution of editing components, thereby enhancing precision and safety of genome-editing therapies.

Koutsoni et al. [103] developed a machine-learning framework to predict CRISPR–Cas9 off-target editing specifically for Duchenne muscular dystrophy guide RNAs. They curated a DMD-focused dataset of gRNA/off-target pairs with measured indel percentages from experimental studies and patents, extracted sequence-based features, and trained multiple regression models (XGBoost, decision tree, SVR, random forest) with nested cross-validation. A decision-tree regressor achieved the best Spearman correlation between predicted and observed indels, and their tool matched or outperformed generic off-target predictors on this DMD dataset, supporting disease-specific models as a strategy to design safer CRISPR therapies in DMD. In a review article, Kim et al. [104] similarly argue that AI will make CRISPR systems more precise, safer, and more efficient.

In DMD, antisense oligonucleotides (ASOs) can partially correct protein malfunction by splice switching in dystrophin pre-mRNA to restore an in-frame transcript and allow production of a shorter, Becker-like but functional dystrophin. In a review article, Leckie et al. [105] argue that ML models trained on large ASO datasets can predict potency, toxicity, and off-target risk. The review highlights tools such as eSkip-Finder and related ML frameworks that learn sequence and structural features associated with effective exon skipping or RNA-degradation ASOs, with emphasis on disorders such as Duchenne muscular dystrophy and SMA, where splicing correction or knockdown of toxic transcripts is central. Kang et al. [106] introduced ASOptimizer, a deep-learning platform that optimizes both ASO sequence and chemical modification patterns to improve stability, target engagement, and RNase H activity. Given a user-defined ASO sequence against a disease target (such as dystrophin or SMN2), the model explores a large combinatorial space of modification sites, encodes these as molecular graphs, and ranks modification patterns by a learned performance score, returning top candidates through a web interface.

Table 5. Therapeutics evidence table.

Author	Year	Domain	Disease	Method	Maturity
Yu [97]	2024	Drug repurposing	ALS	Graph neural network on PPI interactome + network proximity (disease module)	Early
Pun [98]	2022	Drug repurposing	ALS	PandaOmics AI target discovery (multi-model scoring: omics + text-mining + evidence integration)	Early
Sunildutt [99]	2024	Drug repurposing	ALS	Connectivity Map signature reversal + molecular docking (AutoDock Vina)	Early
Gerring [100]	2025	Drug repurposing	ALS	Genetics-led gene prioritization + ATC drug-class enrichment (MAGMA/mBAT; TWAS/SMR; colocalization)	Early
Hoolachan [101]	2024	Drug repurposing	SMA	Transcriptomic signature matching/perturbagen similarity (prednisolone-anchored repositioning)	Early
Koutsoni [103]	2022	Gene editing (CRISPR)	DMD	Supervised ML regression for off-target prediction (XGBoost/DT/SVR/RF; sequence features)	Early
Kang [106]	2025	ASO optimization	DMD	Deep learning for ASO design (sequence + chemical-modification optimization; graph-based ranking)	Early

Note: Method summarizes the primary computational approach used for therapeutic discovery or optimization; some studies use statistical genetics, enrichment, docking, or other bioinformatics workflows rather than machine learning per se. Abbreviations: ALS, amyotrophic lateral sclerosis; ASO, antisense oligonucleotide; ATC, Anatomical Therapeutic Chemical classification; CRISPR, clustered regularly interspaced short palindromic repeats; CV, cross-validation; DMD, Duchenne muscular dystrophy; DL, deep learning; DT, decision tree; eQTL/pQTL/sQTL/meQTL/haQTL, expression/protein/splicing/methylation/histone acetylation quantitative trait loci; GNN, graph neural network; GWAS, genome-wide association study; MAGMA, Multi-marker Analysis of GenoMic Annotation; mBAT, multivariate set-based association test; ML, machine learning; PPI, protein–protein interaction; RF, random forest; SMA, spinal muscular atrophy; SMR, summary data-based Mendelian randomization; SVR, support vector regression; TWAS, transcriptome-wide association study; WGS, whole-genome sequencing. Maturity is the authors’ assessment of how ready the method is for routine clinical use.

Table 5. Therapeutics evidence table.

Author	Year	Domain	Disease	Method	Maturity
Yu [97]	2024	Drug repurposing	ALS	Graph neural network on PPI interactome + network proximity (disease module)	Early
Pun [98]	2022	Drug repurposing	ALS	PandaOmics AI target discovery (multi-model scoring: omics + text-mining + evidence integration)	Early
Sunildutt [99]	2024	Drug repurposing	ALS	Connectivity Map signature reversal + molecular docking (AutoDock Vina)	Early
Gerring [100]	2025	Drug repurposing	ALS	Genetics-led gene prioritization + ATC drug-class enrichment (MAGMA/mBAT; TWAS/SMR; colocalization)	Early
Hoolachan [101]	2024	Drug repurposing	SMA	Transcriptomic signature matching/perturbagen similarity (prednisolone-anchored repositioning)	Early
Koutsoni [103]	2022	Gene editing (CRISPR)	DMD	Supervised ML regression for off-target prediction (XGBoost/DT/SVR/RF; sequence features)	Early
Kang [106]	2025	ASO optimization	DMD	Deep learning for ASO design (sequence + chemical-modification optimization; graph-based ranking)	Early

Note: Method summarizes the primary computational approach used for therapeutic discovery or optimization; some studies use statistical genetics, enrichment, docking, or other bioinformatics workflows rather than machine learning per se. Abbreviations: ALS, amyotrophic lateral sclerosis; ASO, antisense oligonucleotide; ATC, Anatomical Therapeutic Chemical classification; CRISPR, clustered regularly interspaced short palindromic repeats; CV, cross-validation; DMD, Duchenne muscular dystrophy; DL, deep learning; DT, decision tree; eQTL/pQTL/sQTL/meQTL/haQTL, expression/protein/splicing/methylation/histone acetylation quantitative trait loci; GNN, graph neural network; GWAS, genome-wide association study; MAGMA, Multi-marker Analysis of GenoMic Annotation; mBAT, multivariate set-based association test; ML, machine learning; PPI, protein–protein interaction; RF, random forest; SMA, spinal muscular atrophy; SMR, summary data-based Mendelian randomization; SVR, support vector regression; TWAS, transcriptome-wide association study; WGS, whole-genome sequencing. Maturity is the authors’ assessment of how ready the method is for routine clinical use.

7. Discussion

This narrative review found sustained research activity across four domains—diagnosis, outcome modeling, biomarker development, and therapeutics—yet limited routine clinical adoption. With the notable exception of AI-enabled muscle MRI quantification (including FDA 510(k)-cleared tools such as MuscleView 2.0, which are authorized for image segmentation and quantification rather than stand-alone diagnosis), most neuromuscular AI systems remain investigational and are not integrated into everyday clinical workflows.

7.1. Why Clinical Adoption Remains Limited

Across domains, translational barriers recur and can be grouped into five practical bottlenecks:

1.

Small, fragmented datasets with heterogeneous acquisition:

Most neuromuscular conditions are rare, data are distributed across institutions, and protocols vary (imaging, EMG, phenotyping, and longitudinal follow-up). This limits statistical power, impairs generalizability, and makes high accuracy results difficult to interpret without careful validation.

2.

Validation gaps that limit trust:

Many studies rely on retrospective single-center cohorts, internal cross-validation, and incomplete reporting of leakage safeguards, calibration, and uncertainty. External validation, prospective evaluation, and deployment studies remain uncommon.

3.

Modest or context-dependent performance gains:

In several settings—phenotype-driven diagnosis, gene prioritization, and outcome prediction—reported gains are often incremental or benchmark-dependent. By contrast, muscle MRI segmentation and quantification provides a consistent advantage by improving scalability and reproducibility of quantitative measures.

4.

Interoperability and terminology inconsistency:

Inconsistent use of standardized ontologies and machine-readable codes limits reproducibility, multi-site pooling, and downstream computation.

5.

Workflow friction and clinician acceptance:

Standalone tools that require extra data entry, manual phenotype encoding, or separate logins impose overhead. Limited transparency and unclear failure modes further slow adoption, particularly for high-stake decisions.

These bottlenecks suggest that progress will depend less on incremental model tweaks and more on building deployable systems: harmonized multi-center datasets, rigorous model validation, interoperable data representations, and seamless workflow integrations.

7.2. A Practical Roadmap for the Next Generation of Tools

To reduce redundancy, we consolidate design and reporting recommendations garnered from this review into a single actionable checklist (Box 1). We suggest a checklist as a resource for what should be reported and evaluated in future neuromuscular AI studies.

Box 1. Checklist for neuromuscular AI tools (design, evaluation, implementation).

• Clinical use-case clarity: intended user, decision supported, downstream action, and why AI is needed.
• Value proposition: clinically meaningful improvement over standard practice (accuracy, speed, scalability, or cost).
• Training dataset transparency: cohort source, N, missingness, reference standard, class imbalance and mitigation.
• Validation strength: internal vs. external validation, leakage safeguards, patient-level splits where relevant.
• Bias and fairness: subgroup performance (age/sex/race/ethnicity/disease subtype/severity) and bias assessment.
• Explainability and plausibility: interpretable outputs (e.g., SHAP/saliency) plus sanity checks (ablation/counterfactuals).
• Calibration and uncertainty: calibration assessment, confidence reporting, and abstention/defer strategies.
• Workflow integration: where it runs (EHR/PACS/EMG system), clicks/time burden, and operational fit.
• Reproducibility: sufficient methodological detail for replication and benchmarking.
• Standards and interoperability: use of standardized terminologies and comparable metrics.
• Post-deployment monitoring: drift detection, retraining triggers, and adverse-event handling.

7.3. Maturity Assessment

Across application domains, we qualitatively rated maturity on a translational spectrum—from early-stage, through advanced preclinical validation, to deployed clinical use—rather than as a formal regulatory designation, and we summarize this for therapeutics in the “Maturity” field of Table 5. Application maturity is heterogeneous. Phenotypedriven decision-support tools remain difficult to operationalize at scale because they depend on structured phenotyping and are rarely workflow-native, and most remain at an earlystage or advanced preclinical phase with little evidence of routine clinical use. Outcome models (especially in ALS) benefit from shared datasets such as PRO-ACT, but performance remains moderate for individualized prognostication when evaluated rigorously, so these tools are best regarded as early-stage. Therapeutics-oriented pipelines (network medicine, genetics-led prioritization, and multi-omics target discovery) are promising but have not yet produced broadly practice-changing neuromuscular therapies and remain early-stage in maturity. In contrast, AI-enabled muscle MRI segmentation and quantification are among the most mature application areas, with stronger technical validation and early regulatory progress; unlike EMG- or ultrasound-based tools, some MRI-based applications have obtained FDA 510(k) clearance for segmentation and quantification.

7.4. Review Limitations

This work is a narrative review rather than a systematic review. Although structured searches were performed with explicit databases, search strings, and record counts, we did not conduct a PRISMA workflow or formal risk-of-bias assessment. Study reporting quality was heterogeneous, and key methodological details (patient-level splits, missing-data handling, model calibration, uncertainty estimation and handling, and external validation) were inconsistently reported, limiting cross-study comparability. In addition, digital biomarker studies often do not connect candidate signals to trial endpoints or minimal clinically important differences, and regulatory qualification remains uncommon in neuromuscular disease. Finally, we did not systematically evaluate commercial deployment or post-market performance; readiness is inferred primarily from validation strategy and workflow considerations.

8. Conclusions

Neuromuscular diseases highlight both the need and the challenge for AI-enabled precision medicine: phenotypic complexity, genetic heterogeneity, and small cohorts make diagnosis and prognostication difficult, and they slow translation of promising algorithms into practice. Current applications of AI span diagnosis, outcome modeling, biomarkers, and therapeutics, but most systems remain investigational. Translational progress is clearest in quantitative muscle MRI, where AI improves scalability and reproducibility and has begun to enter regulated clinical practice.

Broader adoption will require multi-institutional datasets, standardized acquisition, rigorous external and prospective validation, interoperable phenotype representations, and integration into clinical workflows to reduce clinician burden. To support this transition from research prototypes to routine clinical use, we provide a practical checklist (Box 1) to guide the development of neuromuscular AI tools.