Predicting amyotrophic lateral sclerosis stage based on multi-parameter ultrasound: development and validation of an interpretable machine learning model.
Authors
Affiliations (6)
Affiliations (6)
- The Fourth School of Clinical Medicine, Zhejiang Chinese Medical University, Hangzhou First People's Hospital, Hangzhou, China.
- Department of Ultrasonography, Affiliated Hangzhou First People's Hospital Chengbei Campus, School of Medicine, Westlake University, 50 Jingxin Street, Hangzhou City, 310000, China.
- Division of Health Sciences, Hangzhou Normal University, Hangzhou, China.
- Department of Ultrasound, Chinese and Western Hospital of Zhejiang Province (Hangzhou Red Cross Hospital), Hangzhou, China.
- Department of Ultrasonography, Affiliated Hangzhou First People's Hospital Chengbei Campus, School of Medicine, Westlake University, 50 Jingxin Street, Hangzhou City, 310000, China. [email protected].
- Department of Ultrasonography, Affiliated Hangzhou First People's Hospital Chengbei Campus, School of Medicine, Westlake University, 50 Jingxin Street, Hangzhou City, 310000, China. [email protected].
Abstract
Amyotrophic lateral sclerosis (ALS) lacks sensitive, objective staging tools to guide clinical management and trials. Existing methods have limited granularity and rely on subjective assessment, while biomarker and imaging approaches can be invasive or impractical for serial use. Ultrasound is a safe, portable imaging modality that can detect neuromuscular changes, but it has not yet been applied to ALS staging. We developed and validated an interpretable ultrasound model for clinical staging and risk stratification in ALS. We enrolled 300 ALS patients, classified as early-stage (King's stages 1-2; n = 148) or late-stage (3-4; n = 152). Each patient underwent ultrasound of key muscle groups, including the diaphragm (excursion and thickening), geniohyoid (shear-wave velocity), and peripheral skeletal muscles (thickness and cross-sectional area). Six machine learning models were trained to predict early vs late stage from these ultrasound metrics combined with clinical factors. Performance was evaluated on a test set using area under the curve (AUC), F1 score and Brier score. Feature importance was analyzed with SHapley Additive exPlanation (SHAP) values. In the test set, the random forest achieved an AUC of 0.843, an F1 score of 0.727, and a Brier score of 0.177, with sensitivity 0.80 and specificity 0.68. SHAP analysis identified diaphragm excursion during deep breathing (DEDB) as the top predictor, followed by masseter muscle thickness (MMT) and geniohyoid shear-wave velocity (GHSWVmean). Higher DEDB, MMT and GHSWVmean values predicted earlier stage, whereas lower peripheral muscle thickness and older age indicated late-stage disease. Multiparameter ultrasound combined with machine learning offers a non-invasive, bedside tool for ALS staging. The model's accuracy and interpretability enable objective tracking of disease progression and may support timely interventions and patient stratification in clinical practice and trials. Leveraging widely accessible ultrasound technology, this approach is feasible for routine ALS care and research.