A multivariable modeling framework using decision trees and visual mapping can help guide patient allocation across post-stroke rehabilitation settings, according to an evaluation conducted on synthetic clinical data. Early and timely admission to post-stroke rehabilitation remains a key determinant of functional recovery. To address the allocation challenge, researchers structured and tested a decision-support framework designed to direct stroke patients to appropriate subacute care tiers.
To develop the framework, investigators constructed a heterogeneous synthetic dataset representing 700 stroke patients. The synthetic cohort incorporated 49 demographic, clinical, and functional variables derived from multiple statistical distributions reported in the literature. An expert panel of clinicians independently reviewed and annotated the dataset, assigning each case to one of seven distinct rehabilitation programs that ranged from home-based services to high-intensity hospital-based care to establish the target outcome.
Feature importance analysis identified several key variables as the most influential predictors of clinical allocation. These primary predictors included cognitive function status, National Institutes of Health Stroke Scale scores, patient age, the presence of early seizures, and alcohol consumption.
The researchers trained six supervised machine learning classifiers alongside shallow Classification and Regression Trees of varying depths to predict program assignments. The evaluation protocol implemented 200 repetitions of stratified ten-fold cross-validation. Performance metrics were calculated from pooled out-of-fold predictions and reported as means and standard deviations to assess predictive reliability against expert benchmarks.
In internal evaluations against the clinician-assigned labels, the best-performing decision tree model achieved a mean accuracy of 0.847 and a balanced accuracy of 0.393. An XGBoost model reached a mean accuracy of 0.832 and a balanced accuracy of 0.381. Both machine learning approaches exceeded the institutional benchmark, which registered an accuracy of 0.753 and a balanced accuracy of 0.325 across the multi-class allocation task.
The study also integrated Generative Topographic Mapping to explore patient relationships within the different prediction models. This visualization technique helped identify clusters of patients with shared clinical characteristics and phenotypes. Furthermore, it enabled direct visual comparisons of program recommendations between models, helping researchers identify specific patient characteristics that drove divergent decisions between baseline methods and decision trees.
The researchers concluded that developing and testing multivariable prediction models using an openly shared, expert-annotated synthetic dataset is feasible. They noted that the framework offers an adaptable foundation for creating bespoke clinical decision-support systems. Because the models operate transparently, the authors emphasized that the methodology could support individualized, context-sensitive rehabilitation planning tailored to varying institutional resources and clinical practices.
Subject of Research: Machine learning decision trees and mapping for stroke rehabilitation allocation
Article Title: Data-driven clinical decision trees and generative topographic mapping to guide stroke rehabilitation program allocation
Article References: Komaris, D. S., Anagnostou, P., Nikolopoulos, S., Tsaopoulos, D., Guillen-Sola, A., Duarte, E., Bellfield, R. A. A., Ortega-Martorell, S., Olier, I., & Kompatsiaris, I. (2026). Data-driven clinical decision trees and generative topographic mapping to guide stroke rehabilitation program allocation. PLOS One, 21(11), e0363251. https://doi.org/10.1371/journal.pone.0363251
Image Credits: AI Generated
DOI: 10.1371/journal.pone.0363251
Keywords: stroke rehabilitation, clinical decision support, classification and regression trees, CART models, generative topographic mapping, XGBoost, synthetic patient dataset, multivariable prediction, NIHSS score, cognitive function status, rehabilitation program allocation, stratified cross-validation, subacute rehabilitation, clinical machine learning, patient phenotype clustering, early seizures
Cite Scienmag News
Cassandra Pierce. (November 21, 2026). Decision Trees and Generative Mapping Model Allocation for Stroke Rehabilitation. Scienmag. https://scienmag.com/decision-trees-and-generative-mapping-model-allocation-for-stroke-rehabilitation/
Cassandra Pierce. "Decision Trees and Generative Mapping Model Allocation for Stroke Rehabilitation." Scienmag, 21 November 2026, https://scienmag.com/decision-trees-and-generative-mapping-model-allocation-for-stroke-rehabilitation/. Accessed 21 November 2026.
Cassandra Pierce. "Decision Trees and Generative Mapping Model Allocation for Stroke Rehabilitation." Scienmag. November 21, 2026. https://scienmag.com/decision-trees-and-generative-mapping-model-allocation-for-stroke-rehabilitation/








