Longevity & AgingArtículo de investigaciónAcceso abierto

AI Is Reshaping Neuromuscular Disease Care, but Most Tools Remain Unvalidated

A narrative review of AI in neuromuscular disease finds strong diagnostic and monitoring signals, yet few tools have external validation or clinical use.

jueves, 19 de noviembre de 2026 2 visualizaciones
Publicado en Neurol Ther
Clinician reviewing a glowing thigh muscle MRI scan with overlaid AI segmentation, wearable sensor bodysuit on a monitor nearby

Resumen

This narrative review covers 2018–2025 literature on AI in neuromuscular diseases (NMDs), a group of over 800 mostly rare disorders. AI tools show high performance in variant interpretation, muscle MRI, EMG, pathology, proteomics and prognosis. A wearable endpoint (SV95C) has FDA/EMA qualification for Duchenne trials, and the ENCALS ALS survival model was validated across 14 European centers. However, most models were built and tested on single-center data, few have external or prospective validation, and research concentrates in ALS and Duchenne while conditions like myotonic dystrophy and congenital myopathies are nearly untouched. Federated learning, international networks and foundation models are proposed to close the gaps. This is a review, not new data, with no formal bias appraisal.

0:00--:--

Resumen detallado

Neuromuscular diseases affect roughly one in 1000 people and include more than 800 disorders, most of them rare. Diagnosis is slow (mean delay 4.7 years in a European survey), and rarity starves the machine learning methods that normally need large datasets. This review asks where AI is helping and what stands in the way.

The authors ran a narrative search (PubMed, Web of Science, Scopus, Google Scholar; January 2018 to June 2025) and incorporated 80 references. Screening was by one author with verification by two others. There was no formal risk-of-bias appraisal and no meta-analysis. The review covers strategies for scarce data (transfer learning, few-shot learning, federated learning, synthetic data), then diagnosis, monitoring and prognosis, and biomarker discovery.

Reported results are mostly performance metrics from the original studies. In imaging, a pre-trained CNN reached 91% accuracy for dystrophinopathy on thigh MRI versus 84% for radiologists, and the Myo-Guide platform (2961 scans, 20 NMD types) reached 75% top-3 accuracy on 14 unseen cases versus 64.3% for the best clinician. A pathology CNN classified 11 muscle diseases with AUC 0.996 on cropped images. In ALS, ENCALS (eight predictors, 14 centers, 11,475 patients) gave a C-index of 0.78 and is the only ALS survival model with external validation and low bias risk. Plasma proteomics reached AUC 0.983 for ALS, replicated in an independent cohort. The wearable SV95C is the first FDA/EMA-qualified digital primary endpoint for Duchenne trials. Language models fared modestly: GPT-4o ranked the correct rare disease first in only 19.8% of vignettes.

Implications: AI is nearest to practical use in variant interpretation, MRI and wearable endpoints, and trial design. Research is lopsided, with 547 ALS publications and 86 for DMD/BMD versus 1 for Lambert–Eaton and 9 for congenital myopathies. Federated infrastructure such as EURO-NMD (82 centers) and foundation models may help.

Caveats are substantial. Most models are single-center and internally validated, and only 11.8% of rare-disease ML studies use external datasets. Small head-to-head comparisons, segmentation errors when fat infiltration exceeds 20%, no public muscle whole-slide datasets, and unproven synthetic data all limit confidence.

Hallazgos clave

  • Wearable stride velocity (SV95C) is the first FDA/EMA-qualified digital primary endpoint for Duchenne trials, with an ICC of 0.970.
  • ENCALS ALS survival model (C-index 0.78) is externally validated across 14 European centers, the only one also with low bias risk.
  • Muscle MRI AI beat clinicians in small comparisons (Myo-Guide 75.0% vs 64.3% on 14 unseen cases), so caution is warranted.
  • Plasma proteomics ML gave AUC 0.983 for ALS (183 patients vs 309 controls), replicated in an independent cohort.
  • Research is skewed: ALS has 547 AI publications, while Lambert–Eaton has 1 and congenital myopathies 9; most models lack external validation.

Metodología

Narrative review of AI/ML studies in NMDs from 2018 to June 2025, searching four databases with 80 references incorporated. Single-author screening with verification, and no formal risk-of-bias assessment or meta-analysis. Performance figures are those reported by the original studies and are not pooled.

Limitaciones del estudio

Most underlying studies were single-center, internally validated and small, with few external or prospective tests. The review is narrative with a single screener and no bias appraisal, so selection bias is possible. Coverage is uneven across diseases, and the paper is about neuromuscular disease, not aging mechanisms or lifespan.

¿Te ha gustado este resumen?

Recibe la última investigación sobre longevidad en tu bandeja de entrada cada semana.

Introduce tu correo electrónico para suscribirte: