Factors associated with clinical meaningful recovery after upper limb task-oriented training in people with stroke: a cohort study

<b>Introduction:</b>This study investigated upper extremity (UE) recovery predictors in post-stroke patients undergoing task-oriented training (TOT) rehabilitation. The data were collected at the Don Gnocchi Foundation hospitals in Italy between 2011 and 2015. Ethical approval was obtained, and all participants provided informed consent.<b>Methods:</b><b>Participants:</b> 64 participants were recruited.<b>Intervention:</b> Participants received 25 sessions of TOT (45 minutes each, five days/week) as an adjunct to standard therapy. TOT focused on functional tasks with real-life objects and was progressively adapted to individual needs.<b>Outcome Measures:</b><b>Fugl-Meyer Assessment - Upper Extremity (FMA-UE):</b> To assess impairment (ICF body function domain).<b>15-item Action Research Arm Test (ARAT-15):</b> To assess activity/performance (ICF activity domain).<b>Quick version of the Disability of the Arm, Shoulder, and Hand questionnaire - 9 items (Q-DASH-9):</b> To assess participation restrictions (ICF participation domain).<b>Predictor Variables:</b> Age, sex (i.e., male, female), dominance of the affected side (affected bodyside; i.e., dominant; non-dominant), chronicity (i.e., chronic; subacute), injury typology (i.e., ischemic; hemorrhagic), injury localization (i.e., cortical; sub-cortical), and baseline scores on the FMA-UE, ARAT-15, and Q-DASH-9.<b>Data Analysis:</b>Descriptive statistics were used to summarize data.Wilcoxon signed-rank test was used to compare pre- and post-intervention scores.Effect sizes were calculated using matched-pairs rank-biserial correlation.Participants were classified as "Responders" or "Non-responders" based on achieving minimally clinically important differences (MCID) in outcome measures.Stepwise binary logistic regression models were developed to identify predictors of responder status. A bidirectional approach was employed, starting with an empty model. Predictors were added (forward selection) or removed (backward elimination) one at a time based on whether their inclusion improved the Akaike information criterion (AIC).Model accuracy was assessed using McFadden’s pseudo-R2, Scaled Brier Score, Receiver Operating Characteristic curve (AUC), and Hosmer–Lemeshow test.Subgroup sensitivity analysis was performed to assess model robustness.<br>