Background: Physical activity is low after stroke. High intensity treadmill training and self-management strategies positively affect physical activity and walking outcomes, but it is unclear how best to tailor interventions. Identifying participant and intervention-related predictors may enable improved tailored intervention design.
Aim: To identify participant and intervention-related predictors of physical activity and walking outcomes after stroke using a machine learning approach
Methods: This secondary analysis of data from a randomised controlled trial (ACTRN12613000744752) included participants within two months of stroke who could walk. Participants received a self-management program embedded in high-intensity treadmill training (3 sessions per week, 8 weeks). Participant characteristics included demographics, baseline step count, and self-efficacy for walking and exercise. Intervention characteristics included treadmill performance (speed, distance, and rate of perceived exertion) and self-management strategies. Primary outcomes were daily step count, walking speed and distance at baseline and 26 weeks. Machine learning algorithms identified clusters, and multiple regression identified predictors of physical activity and walking outcomes.
Results: Fifty-six participants, 80% male, mean 28 (SD 15) days post-stroke, completed the intervention. Three clusters were identified. Cluster 1 (n=20, age 58+13) had the highest baseline step count (6724+3017). Cluster 2 (n=15, age 70+9) had a baseline step count of 4667+1733. Cluster 3 (n=21, 61+11) had lowest baseline step count (2922+1538). Predictors for physical activity and walking outcomes varied among the clusters, with R² values ranging from 0.50 to 0.87.
Conclusion: Tailoring interventions to individual participant characteristics is required when targeting post-stroke physical activity and walking.