Multimodal machine learning to predict response to ultrasound-guided botulinum and vibration therapy in muscle spasticity a clinical and imaging correlation study

TL;DR

This study uses machine learning to predict how well ultrasound-guided botulinum toxin and vibration therapy work together to treat muscle spasticity after stroke.

Contribution

The novel contribution is a multimodal machine learning model that combines clinical and imaging data to predict treatment response in spasticity management.

Findings

Combined botulinum toxin and vibration therapy improved spasticity and functional outcomes more than botulinum toxin alone.

A neural network model achieved the highest predictive accuracy (AUC = 0.87) for treatment response.

Ultrasound biomarkers like muscle thickness and echo intensity were important features in the model but had limited direct correlation with clinical outcomes.

Abstract

Muscle spasticity remains a challenging motor complication following stroke. Although botulinum toxin (BTX) injection is widely accepted for the management of focal spasticity, therapeutic responses vary considerably among individuals. Vibration therapy has been proposed as a complementary modality; however, predictive models that integrate clinical and imaging features to anticipate treatment response remain limited.

To evaluate the clinical efficacy of ultrasound-guided botulinum toxin combined with vibration therapy in post-stroke spasticity management and to develop a multimodal machine learning model for individualized outcome prediction.

A total of 200 participants were randomized to receive either BTX alone or BTX combined with vibration therapy. Clinical assessments—including the Modified Ashworth Scale (MAS), Fugl–Meyer Assessment (FMA), and Barthel Index (BI)—were conducted at baseline and 12 weeks post-treatment. Ultrasound-derived biomarkers, including muscle thickness, echo intensity, and blood flow score, were collected. Multiple machine learning models (random forest, support vector machine, XGBoost, and a feedforward neural network) were trained using combined clinical and imaging features to predict treatment response.

Compared with BTX alone, the combined treatment group demonstrated significantly greater reductions in MAS (mean ΔMAS: 1.48 vs. 1.12, p < 0.01) and greater improvements in FMA and BI scores (p < 0.05). Among the evaluated models, the neural network achieved the highest predictive performance (AUC = 0.87). Muscle thickness and echo intensity emerged as influential features in the prediction models; however, their direct associations with clinical outcomes were limited. Correlation analysis revealed generally weak associations between ultrasound-derived biomarkers and changes in clinical outcomes, indicating limited direct correlations at the individual level.

Ultrasound-guided BTX combined with vibration therapy provides greater improvements in spasticity and functional outcomes compared with BTX alone. Multimodal machine learning models demonstrate potential for predicting individual treatment response, supporting the adjunctive role of ultrasound-derived biomarkers in personalized spasticity management.

Figures