Prediction of 6-month poststroke spasticity in patients with acute first-ever stroke by machine learning

Abstract

Abstract Objective Poststroke spasticity (PSS) reduces arm function and leads to low levels of independence. This study suggested applying machine learning (ML) from routinely available data to support the clinical management of PSS. Design 172 patients with acute first-ever stroke were included in this prospective cohort study. Twenty clinical information and rehabilitation assessments were obtained to train various ML algorithms for predicting 6-month PSS defined by a modified Ashworth scale (MAS) score ≥ 1. Factors significantly relevant were also defined. Results The study results indicated that multivariate adaptive regression spline (area under the curve (AUC) value: 0.916; 95% confidence interval (CI): 0.906-0.923), adaptive boosting (AUC: 0.962; 95% CI: 0.952-0.973), random forest (RF) (AUC: 0.975; 95% CI: 0.968-0.981), support vector machine (SVM) (AUC: 0.980; 95% CI: 0.970-0.989) outperformed the traditional logistic model (AUC: 0.897; 95% CI: 0.884-0.910) (P < 0.05). Among all of the algorithms, the RF and SVM models outperformed the others (P < 0.05). FMA score, days in hospital, age, stroke location, and paretic side were the most important features. Conclusion These findings suggest that ML algorithms can help augment clinical decision-making processes for the assessment of PSS occurrence, which may enhance the efficacy of management for patients with PSS in the future.