AI and Neurological Rare & Neuroimmune Diseases ── ALS ── 2026-11-01
ALS: treatment, trials and AI
- Abstract
BACKGROUND: Motor neuron disease (MND) is a devastating and fatal neurodegenerative disease with limited treatment options. Diagnosis is hampered by the lack of definitive and reliable investigations. Dysarthria is a common early symptom, but current methods for assessing speech are crude, subjective and burdensome. Speech processing and artificial intelligence (AI) can offer objective, reliable and scalable assessments with low-burden and patient-centred digital technologies. OBJECTIVES: Describe existing work in speech processing methods to assess dysarthria in MND. Develop a patient-centred and user-friendly speech recording app for people living with MND. Collect and curate a longitudinal speech dataset with detailed demographic, language background and clinical annotations. Assess the usability and acceptability of the proposed assessment methods. Develop deep learning models for dysarthria severity assessment, and clinically evaluate their validity, reliability and utility in diagnosis and monitoring. Assess potential biases and confounding factors in a digital speech biomarker for MND. METHODS: A systematic literature review of speech processing and digital speech biomarkers in MND was conducted. A speech recording app, co-produced with people living with MND and their caregivers, was deployed in observational studies recruiting MND and neurologically healthy control groups. Assessments included six speech tasks and digital outcome measures with concurrent clinical assessments, including the revised amyotrophic lateral sclerosis functional rating scale (ALSFRS-R) and cognitive tests. The usability and acceptability of the app assessments were examined through participant feedback. 20 deep learning speech embeddings were benchmarked; the best were used to train a novel model via scalar ordinal triplet learning. The clinical validity and reliability of the novel model were evaluated in a held-out test set, and externally validated in German (DZNE Rostock) and Taiwanese (VoiceBank-2023) datasets, with results pooled using random effects models. Confounding bias was assessed using a silence-based ablation study and regression-based statistical adjustment for demographic and recording-related variables. Clinical correlations were evaluated against functional, cognitive and quality of life measures. Progression modelling used linear mixed-effects models to estimate individual trajectories, identify heterogeneous progression sub-groups, and assess the predictive value of baseline assessments for future rates of decline. RESULTS: A total of 298 participants (115 with MND, 183 controls) contributed to the speech corpus across 540 assessments, with 91 (31%) contributing longitudinally for up to 3 years. Participants were representative of the broader MND population. The speech assessment procedure received positive feedback, showing good to excellent usability and acceptability. Deep learning speech embeddings outperformed conventional acoustic features in classifying diagnosis. The proposed model was evaluated across internal and external datasets (254 speakers: 136 with MND, 118 controls). Classification performance for MND diagnosis and dysarthria detection was excellent, with pooled AUCs of 0.905 (95% CI [0.706, 0.974]) and 0.937 (95% CI [0.859, 0.973]) respectively, and was significantly associated with the ALSFRS-R speech subscore (β = 0.552, 95% CI [0.274, 0.830]). Reliability was excellent (internal consistency α = 0.989; test-retest ICC = 0.829). The model generalised well to external datasets of different languages after calibration with only 10 control speakers, outperforming conventional supervised models on all measures and showing superior reliability to the ALSFRS-R speech subscore. Confounding bias assessment demonstrated robustness against demographic and recording-related variables. Model predictions correlated with ALSFRS-R bulbar (β = 1.510, p < 0.001) and total (β = 2.314, p = 0.005) scores; and EQ-5D-5L quality of life domains (p < 0.01), but not with non-bulbar ALSFRS-R domains or cognitive function. Bulbar onset was associated with worse baseline speech function (Cohen’s d = -1.011, p = 0.011) and faster progression (Cohen’s d = -0.931, p = 0.026) than limb onset. Progression modelling identified stable fast and slow progressing subgroups (H = 33.94, p < 0.001), with baseline assessments predicting subsequent dysarthria decline (Cohen’s f² = 2.118, permutation test p < 0.001). CONCLUSION: Patient-centred digital speech assessments are scalable, robust and well-accepted by people with MND. The proposed deep learning model outperformed supervised methods by leveraging weakly labelled data to learn dysarthria severity relationships, yielding superior clinical validity and reliability compared to existing scales. The model generalises across languages after minimal calibration and confounding analysis confirmed it captures disease-relevant speech signals rather than spurious correlations. Clinical correlations and progression modelling further support the model’s construct validity in monitoring within-individual speech decline. These findings establish our proposed model as a state-of-the-art digital speech biomarker for MND, with applications in early detection, prognostication, progression monitoring, and clinical trials. This work represents a significant step towards clinical deployment. Future directions include prospective and multi-centre validation, multilingual expansion, integration into care pathways, and extension to other neurodegenerative disorders.
Journal IF: unknown (could not be matched)Reference: [No authors listed]. Deep learning digital speech biomarkers in motor neuron disease. University of Edinburgh. [Epub ahead of print]. doi:10.7488/era/7947.Checked: Abstract only - Abstract
Background: Electromyogram (EMG) is considered for the diagnosis of amyotrophic lateral sclerosis (ALS). We investigated how frequently an initial electrophysiological evaluation is performed in ALS patients.
Objective: We aimed to identify ALS cases using the feature selection algorithm (FSA). A pattern of feature involvement was expected in the early phase of ALS based on the typical consistent involvement of a few particular muscles in the late stages.
Methods: We used a clinical and EMG dataset of 36 patients (21 ALS patients and 15 healthy) for analysis. The present approach used a Random Forest (RF)-based machine learning (ML) algorithm for the prediction of ALS patients. The goal of this study was to determine how feature selection techniques can enhance machine learning algorithms that can be used in real-time applications.
Findings: The experimental findings demonstrate the superior performance of the proposed model. This work demonstrates how machine learning models can offer important predictive insights in ALS, improving knowledge of disease dynamics and aiding patient care. According to the comparison results, the RF model utilizing the feature selection algorithm demonstrated a positive improvement, and the accuracy of the RF model with and without the feature selection technique was found to be 87.50% and 75%, respectively.
Conclusions: The algorithm improved the model accuracy from 75% to 87.50%, resulting in an absolute increase of 12.5 percentage points in classification accuracy. This study found that fibrillation, positive sharp waves (PSW), and fasciculation features were high-yielding in EMG of ALS cases.
Journal IF-equivalent: 4.0 (OpenAlex 2-year mean citedness, value as of 2026-10-16, retrieved 2026-10-17; not the official Clarivate IF)Reference: Dutt A, Tiwari A, Tripathi S, Tiwari PK, Kumar M, Dwivedi A, et al. Enhancing Machine Learning Performance for ALS Classification Using Feature Selection Technique. Diagnostics. 2026 Oct 29;16(21):3573. doi:10.3390/diagnostics16213573.Checked: Abstract only