Objective: To construct an item response theory (IRT) model for functional assessment of non-acute stroke patients based on the international classification of functioning, disability and health (ICF) for functional assessment, verification of its reliability and validity, and exploration of the overall benefit index and interrelationships of functional items through graphical models (GMs).

Methods: A cross-sectional study was conducted, enrolling non-acute stroke patients admitted by the First Rehabilitation Hospital of Shanghai from July 2023 to December 2023. Assessments included the Mini-Mental State Examination (MMSE), Functional Independence Measure (FIM), Stroke Social Network Scale (SSNS), and ICF Rehabilitation Set (ICF-RS). Based on the Mokken assumption, nonparametric IRT (npIRT) was applied to the ICF-RS items through Mokken Scale Analysis (MSA), by which items suitable for a parametric IRT model (pIRM) were identified, and person-item maps were generated to compare patient's abilities with the difficulty levels of functional items. Based on Bayesian learning, the benefit values of the items and the network parameters of the functional items were estimated using the Ising model, and a maximum spanning tree was generated to visualize the distribution of functional items and the potential intervention pathways.

Results: Adhering to the maximum heterogeneity sampling, a total of 100 non-acute stroke patients (males = 70, females = 30) were included, yielding a Rasch scale containing 22 ICF items with a latent class reliability coefficient (LCRC) of 0.959. The dichotomized Rasch model scores significantly correlated strongly with FIM scores (P < 0.001, r= 0.89). The most difficult items included "b455 Exercise tolerance", "d660 Assisting others" and "d640 Doing housework". The easiest item was "d550 Eating". By integrating IRT and GM, the item "d420 Moving oneself" was identified as a functional item that, although within the patient's ability range, still presents impairments, and its improvement may serve as a "bridge" for alleviating other functional impairments. The Ising model indicated that enhancing the ability of patients in "d510 Washing oneself" would yield the greatest overall functional benefit.

Conclusion: Based on pIRM, a Rasch model with good reliability and validity can be extracted for functional assessment of patients with non-acute phase of stroke, and the application of the IRT model and Ising network analysis can customize personalized rehabilitation goals for patients.