Authors: (including presenting author): :
Sit LMK (1), Fong KNK (2), Ho CM (1)
Affiliation: :
(1) Occupational Therapy Department, Tai Po Hospital, (2) Department of Rehabilitation Science, The Hong Kong Polytechnic University
Keyword 3: :
Machine-learning algorithm
Keyword 4: :
Prediction model
Introduction: :
Predicting functional recovery and length of stay (LOS) in stroke patients is important for individualized treatment planning, decision on discharge disposition and appropriate resource allocation. Existing prediction models are often not locally validated and rarely use data-driven machine learning approaches on large clinical datasets.
Objectives: :
The study aimed to examine predictive attributes for functional recovery and LOS among patients receiving in-patient stroke rehabilitation, and to develop and validate clinical prediction models using both conventional regression analysis and a decision tree algorithm.
Methodology: :
A retrospective study was conducted using data from the Occupational Therapy Department Stroke Registry of Tai Po Hospital, covering 5133 patients referred for stroke rehabilitation between January 2008 and December 2021. Six domains of variables were analyzed: demographic, socioeconomic, medical information, physical measures, cognitive measures, and functional measures. Primary outcomes were ADL functional gain (Modified Barthel Index, MBI admission–discharge difference) and LOS of patients, with secondary outcomes of discharge disposition and post-stroke mortality. Multivariable linear and multinomial logistic regression were used to identify significant predictors, and classification and regression trees (CART) were constructed in R (rpart) using a 70/30 training-test split to build and validate decision rules for functional outcome and LOS.
Result & Outcome: :
Regression analysis indicated that baseline and discharge MBI scores, communication ability, first-ever stroke, and renal dysfunction were significant predictors of functional outcomes with an overall model fit of R² = 0.999, p < .001. Factors such as patients’ age, caregiver availability, days since stroke onset, lesion site and region, limb tonicity, mobility level, cognitive function, and improvements in MBI scores significantly predicted the LOS for patients (R² = 0.269, p < .001). Patients’ discharge to residential care was significantly predicted by factors like caregiver availability (OR = 1.740, p = .007), upper limb hypertonicity (OR = 2.998, p < .001) and discharge ADL independence (OR = 0.165, p < .001), whereas no variable demonstrated statistically significant association with post-stroke mortality. The pruned CART models yielded cross-validated R² values of 0.834 for ADL gain and 0.255 for LOS. The models also identified the discharge MBI score and baseline balance as primary splits for functional gain and LOS respectively. These findings highlight the potential of utilizing large registry data and CART for personalized rehabilitation planning within a local Chinese population.
