Domain-specific functional outcome prediction in stroke rehabilitation: A multicenter artificial intelligence study

 Tell me PRECISELY HOW THIS GETS SURVIVORS RECOVERED!

Or just shut your fucking yaps and let some people with brains solve stroke! Somehow you are so blitheringly stupid you don't know predictions are completely fucking useless! WOW! IMPRESSIVE STUPIDITY!

Domain-specific functional outcome prediction in stroke rehabilitation: A multicenter artificial intelligence study

Author links open overlay panel, , , , , , , , , , 

Highlights

• •
  AI-based models predicted domain-specific functional outcomes after stroke, including ambulation, cognition, and ADLs.
• •
  Alignment-based regularization improved cross-institutional generalizability across four multicenter rehabilitation cohorts.
• •
  Domain-specific prediction models achieved AUROCs up to 0.897 in internal and 0.924 in external validation.
• •
  A prototype web-based decision support tool provides patient-specific recovery trajectories at 3 and 6 months after stroke.

Abstract

Background

Stroke is a leading cause of long-term disability worldwide, yet existing clinical decision support tools rely on global disability metrics, such as the modified Rankin Scale, which do not adequately reflect patient-centered rehabilitation or recovery goals.

Objective

We aimed to develop, validate, and implement artificial intelligence (AI)–based clinical support models for predicting domain-specific functional outcomes after stroke across multiple rehabilitation institutions and timepoints.

Methods

We utilized prospective and retrospective multicenter data collected from patients with stroke across four rehabilitation institutions in South Korea (2017–2024). Prognostic models were developed for three functional domains—ambulation, cognitive function, and activities of daily living—under two temporal scenarios: acute-to-subacute and acute-to-early-chronic prediction. Alignment-based regularization was applied to improve cross-institutional generalizability.

Results

The proposed framework achieved strong predictive performance in internal validation (AUROC up to 0.897 for ambulation and 0.864 for cognition) and favorable external validation performance, with AUROCs up to 0.924 (ADLs) and 0.892 (cognition), despite institution-specific differences in cohort size and variable availability across centers. The implemented web-based clinical decision support system provides real-time prediction of individualized recovery trajectory with intuitive visualization designed to support clinician–patient communication.

Conclusions

Our findings demonstrate the predictive feasibility of AI-based modeling for domain-specific stroke prognosis and present a prototype implementation illustrating the potential clinical applicability of the proposed framework across heterogeneous institutional settings. The use of routinely collected clinical variables and the preliminary cross-institutional validation results support further investigation of real-world rehabilitation practices, pending prospective clinical evaluation.