Personalised stroke rehabilitation: an AI pipeline for exercise programmes using a co-designed decision support tablet application

Rajkumar, Thrisha, Koerner, Sarah S., Pinto, Anika, Shakya, Regan, Pope, James, Galvez Trigo, Maria ORCID: https://orcid.org/0000-0001-6492-0955, Al-Nuaimi, Ali, Loizou, Michael, O'Hara, Kenton and Kumar, Praveen 2026. Personalised stroke rehabilitation: an AI pipeline for exercise programmes using a co-designed decision support tablet application. Presented at: 19th International Conference on Health Informatics, HEALTHINF 2026, Marbella, Spain, 2-4 March 2026. SciTePress,

PDF - Accepted Post-Print Version Download (1MB)

Abstract

Stroke rehabilitation requires personalised and continuously adapted exercise programmes, resulting in significant therapist involvement and is often impractical for patients recovering at home in community settings. This motivates the need for assistive tools and decision support systems to enhance efficiency and rehabilitation progress. This position paper presents an integrated pipeline combining a therapist-informed tablet application with artificial intelligence (AI) models to support therapists in decision-making. Co-designed with stroke therapists, HCI researchers, AI experts, and PwS, the application captures baseline and weekly reassessment data, including BBS, TUG, pain, perceived difficulty, and FITT prescriptions, across 4–6 week cycles to determine whether to progress, sustain, or regress exercises. To facilitate early model development, we created a clinically informed synthetic dataset (n = 336 sessions across 5 PwS profiles over 12 weeks) that simulates functional progression and therapist decision-making patterns. This dataset reflects key features identified through workshops with clinicians and PwS, capturing essential assessment metrics such as stroke characteristics, functional scores, therapist goals, patient feedback, exercise difficulty, repetitions, duration, body area, FITT parameters, and exercise recommendations. We trained and evaluated models to predict weekly progression decisions. Logistic regression achieved a weighted F1-score of 51.6%, while a multilayer perceptron reached 79.3% and a decision tree 90.2%. Clinical data will be collected in the next stage of the project (5–8 PwS, 4–6 weeks) and integrated with the synthetic dataset using real–synthetic fusion. Future work includes exploring generative AI for richer rehabilitation trajectories and validating models through an ongoing clinical study. Overall, this work advocates for AI-augmented tools in rehabilitation informatics to support scalable, patient-centred stroke care in the community.

Item Type:	Conference or Workshop Item - published (Paper)
Status:	In Press
Schools:	Schools > Computer Science & Informatics
Publisher:	SciTePress
Date of First Compliant Deposit:	27 January 2026
Date of Acceptance:	19 December 2025
Last Modified:	28 Jan 2026 11:00
URI:	https://orca.cardiff.ac.uk/id/eprint/184245

Actions (repository staff only)

Edit Item