Identifying determinants of readmission and death post-stroke using explainable machine learning

Background: Stroke remains a global health challenge with high rates of mortality and rehospitalization placing significant demands on healthcare systems. Identifying factors that determine outcomes of post-hospitalization improves resource allocation. Traditional statistical prediction models are suboptimal for the analysis of complex, multi-dimensional datasets. The objective of our study is to define the extended list of clinical and non-clinical predictors, which we believe can be achieved using Explainable Machine Learning (XML) models as an expansion of conventional methods. Methods: We evaluated 11 established XML models that represent key ML methodologies to predict 90-day outcomes, namely mortality and rehospitalization among stroke survivors. The study population are 1,300 post-stroke individuals enrolled in the Transitions of Care Stroke Disparities Study (TCSD-S) (NIH/NIMH, NCT03452813) between June 2018 – October 2022. The care after transition data is sourced from participating comprehensive stroke centers and from the Florida Stroke Registry. The analysis incorporated clinical (e.g., age, stroke severity, comorbidities) and non-clinical factors including Social Drivers of Health (SDOH). A combined ranking approach, using Weighted Importance Scores and Frequency Counts, identified significant predictors across models. Results: The resulting list of selected predictors included both established clinical factors and non-clinical factors, which enhanced prediction accuracy. Out of 38 identified predictors, 20 are non-clinical variables reflecting the importance of SDOH, environmental factors, and behavioral modifications beyond traditional clinical predictors of death/readmission. A secondary analysis restricted to ischemic stroke patients (n = 1,038) yielded virtually identical predictive performance, indicating robustness of the model within this subgroup. Conclusions: Integrating SDOH, environmental factors, and behavioral modifications alongside traditional clinical predictors enhances the predictive accuracy of post-stroke outcome models. This underscores the critical role of addressing socioeconomic disparities during post-stroke transitions of care. Moreover, XML models’ ability to identify predictors spanning clinical and non-clinical domains suggests their potential to guide recovery. The resulting predictors are crucial for post-hospital care and hold strong potential for identifying individuals at risk of stroke, making them potentially significant across pre-stroke and hospitalization stages.