https://orcid.org/0000-0001-7573-0671
Figures
Abstract
Background: Frailty is common after stroke and linked to poor outcomes, but many measures are clinician-rated, time-consuming, and not suited to patient-reported data. To address these issues, we developed and validated a frailty score from the Swedish Stroke Register (Riksstroke) three-month follow-up questionnaire.
Methods: We analyzed responses from 19,470 stroke survivors to nine patient-reported items covering function, mood, fatigue, pain and general health, in the 2021–2022 Riksstroke questionnaire. Dimensionality was assessed with Mokken Scale Analysis and exploratory factor analysis. Item response theory (IRT) was used for score computation. Competing graded response IRT models (unidimensional, correlated-factor, bifactor) were compared, and measurement fairness was examined using differential item functioning (DIF) across age, sex, and education. Prognostic validity was tested with Kaplan–Meier curves and Cox regression for all-cause mortality.
Results: From the Mokken Scale Analysis, all items met scalability criteria. Factor analysis suggested two correlated interpretable facets (Physical Functioning; Well-being/Mental Health). A bifactor IRT model provided the best fit to the data, comprising a general frailty dimension while addressing the strong correlation between the facets. DIF was minimal for sex and education, with modest age-related effects. Higher frailty scores were associated with increased mortality in adjusted Cox models and Kaplan–Meier curves. Tools for computing frailty scores are available at https://github.com/joakimwallmark/frailty-irt-scores.
Conclusions: A robust, fair, and prognostically meaningful frailty score can be derived from patient-reported items in Riksstroke. More broadly, the study demonstrates how routinely collected patient-reported outcome measures can be leveraged to build scalable frailty scores, offering efficient cost-effective tools for monitoring outcome and guiding quality improvement in stroke care.
Citation: Wallmark J, Wiberg M, Eriksson M (2026) Developing and validating a frailty score based on patient-reported outcome 3 months after stroke: A Riksstroke-based study. PLoS One 21(2): e0343249. https://doi.org/10.1371/journal.pone.0343249
Editor: Yoshitaka Ishibashi, Japanese Red Cross Medical Center, JAPAN
Received: October 7, 2025; Accepted: February 3, 2026; Published: February 20, 2026
Copyright: © 2026 Wallmark et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The data analyzed in the study consist of third-party data from Riksstroke and Statistics Sweden and their use has been approved by the Swedish Ethical Review Authority (document number 2023-07750-01). Because of the sensitive nature of the data and according to Swedish legislation (https://etikprovningsmyndigheten.se/for-forskare/vad-sager-lagen/) data cannot be made available for use beyond what has been approved by the ethical review authority. Therefore, the data cannot be made publicly available. Data may be made available from Riksstroke (contact via riksstroke@regionvasterbotten.se) or Statistics Sweden (contact via mikrodata@scb.se) upon reasonable request by researchers who meet the criteria for access to confidential data according to Swedish laws and regulations.
Funding: The study was funded by the Swedish Research Council, https://www.vr.se/english.html, (grant number 2024-02846 to ME, and grant number 2022-02046 to MW). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: No authors have competing interests.
Introduction
Stroke is a leading cause of long-term disability and premature mortality worldwide [1]. In Sweden, the national stroke register (Riksstroke) issues a follow-up questionnaire three months after stroke to monitor patient recovery and to inform quality improvement at the local, regional, and national level [2]. A central challenge is how to summarize this patient-reported information into a single, interpretable measure that reflects overall health and can be used for risk stratification.
Frailty provides one such framework. Frailty is commonly defined as a clinical syndrome of reduced physiological reserve and resilience, in which seemingly minor stressors can cause disproportionate deterioration in health [3,4]. Conceptually, it reflects the erosion of the body’s in-built reserves across multiple organ systems, making individuals more vulnerable to illness, complications, and mortality. In stroke, frailty is highly prevalent: systematic reviews estimate that about one quarter of patients presenting with acute stroke are already frail, and nearly half are frail or prefrail when broader definitions are applied [3,5,6]. Frailty is clinically important because it predicts worse recovery, reduced effectiveness of acute treatments, longer hospital stays, institutionalization, and higher mortality [5].
Frailty is not the same as chronological age or disability. A patient can be frail without being severely disabled, and vice versa [5]. Importantly, frailty measures have sometimes outperformed disability scales in predicting long-term outcomes [4]. This has led to repeated calls for standardized, explicit assessment of frailty in stroke research and clinical care, rather than assuming that age or disability adequately capture it [4,6].
Several approaches to measuring frailty are used in clinical and research settings. Phenotype-based tools such as the Fried criteria [7] or FRAIL scale [8] define frailty by meeting a set of physical criteria (e.g., weakness, slowness, exhaustion, low activity, weight loss). Cumulative-deficit indices, such as the Frailty Index (FI) [9], the electronic Frailty Index (eFI) [10], and the Hospital Frailty Risk Score (HFRS) [11], quantify frailty as the proportion of accumulated health deficits. In clinical practice, judgment-based tools such as the Clinical Frailty Scale (CFS) [12,13] are also widely used, providing a 9-point clinician-rated summary of overall fitness or frailty based on comorbidity, function, and cognition. Although these measures have strong prognostic validity for outcomes such as mortality and institutionalization, they are clinician-assessed, time-consuming, and not tailored to large-scale patient-reported follow-up data [14–16].
The Riksstroke follow-up collects 20 patient-reported items [17,18]. The English translation of the questionnaire is currently available at https://www.riksstroke.org/wp-content/uploads/2021/03/3-manaders-uppfoljning-2021_-engelska-1.pdf. In this study, we consider nine of these items that cover function, mood, fatigue, pain, and general health—domains that are central to frailty. Methodologically, frailty can be viewed as a latent construct underlying these observed indicators. Psychometric models, such as Item Response Theory (IRT) [19], provide a principled way to evaluate dimensionality, estimate fair and comparable scores across subgroups, and validate them against external outcomes. Previous work on patient-reported outcome measures (PROMs) in stroke has applied IRT for scale development (but not specifically for frailty) [20] and shown the prognostic value of PROMs for future health events [21].
In contrast to clinician-rated frailty tools (e.g., HFRS, eFI), this study develops and validates an IRT-derived frailty score directly from patient-reported outcome items in Riksstroke. We (1) assess dimensionality using Mokken Scale Analysis and exploratory factor analysis, (2) compare unidimensional, correlated-factor, and bifactor graded IRT models, (3) examine and test measurement fairness across age, sex, and education using differential item functioning (DIF) analyses, and (4) evaluate external validity by linking the score to all-cause mortality.
Material and methods
Riksstroke
This study is based on data from Riksstroke, the national Swedish stroke register, which includes a structured follow–up questionnaire administered three months after stroke. The questionnaire includes patient–reported items that cover domains such as activities of daily living, mobility, mood, fatigue, pain, and general health. For the present analyses, we focused on a subset of nine items (English versions presented in Table 1) that map most directly onto established frailty domains. Items regarding transportation and cognitive symptoms (memory/concentration) were excluded a priori. This was done to minimize environmental confounding (transportation) and to focus the latent trait on physical resilience and general well-being, adhering closer to the physical frailty phenotype rather than a cumulative deficit index.
A short name (in bold) has been assigned to each item. Response options have been rearranged in the order from most to least favorable outcome
This study was approved by the Swedish Ethical Review Authority (reference number 2023-07750-01). Data were pseudonymized before it was submitted for the study. It was accessed for research December 5, 2024. Patients are informed about registration in the quality registry Riksstroke that the registry aims to support high and consistent quality of care for stroke patients throughout Sweden, and that data may be used for research purposes. In accordance with the Personal Data Act (Swedish law No. SFS 1998:204), no informed consent is needed to collect data from medical charts or inpatient records for quality registries. However, patients are informed of their rights to deny participation (opt-out consent).
The source population comprised all patients with ischemic or hemorrhagic stroke (International Classification of Diseases 10th Revision, ICD-10 codes I61, I63, and I64) during 2021–2022 who responded to the three–month follow–up questionnaire. We included all respondents with available answers to the nine frailty–related items. Questionnaires completed without the patient’s involvement (i.e., filled in solely by a relative or healthcare professional) were excluded, as patient input is essential for measurement invariance analysis [22]. To preserve the ordinal structure of responses, “don’t know” responses were coded as missing. As a pre–specified data–quality criterion, records with more than three missing responses across the nine items were excluded. The resulting analytic sample, along with descriptive characteristics, is reported in the Results section.
Statistical methods
Item responses were treated as ordered as displayed in Table 1; we denote the response to item j as , and code “don’t know” as missing.
Dimensionality assessment.
To create a frailty score, the items must first be shown to measure a single, coherent underlying construct. We therefore assessed dimensionality with two complementary methods: Mokken Scale Analysis (MSA) [23,24] and an Exploratory Factor Analysis (EFA).
MSA was employed to assess the unidimensionality and hierarchical ordering of the items. MSA evaluates if items form a scale where individuals endorsing a more “difficult” or “extreme” item are also likely to endorse all “easier” or less “extreme” items. This involved an examination of Guttman errors to identify response patterns inconsistent with a Guttman-style ordering. Scalability was assessed using Loevinger’s H coefficients, including item-pair (Hij), individual item (Hi), and overall scale homogeneity (HS) coefficients. Generally, scales are considered weak if , moderate if
, and strong if
, with individual items ideally having
[24]. Since MSA served as a descriptive check of scalability/monotonicity rather than as a basis for IRT model specification, it was conducted on the full dataset to maximize precision of Loevinger’s H estimates.
EFA was performed to further identify the underlying latent constructs, and to empirically establish item–factor mappings for IRT model specification. Given the ordinal nature of the item responses, a polychoric correlation matrix was used. To reduce bias from discovering the latent structure and later constructing IRT models using the same data, a split-sample workflow was employed. The EFA was conducted using a randomly selected 20% subset, leaving the remaining 80% as a confirmation subset for IRT model fitting. The suitability of the EFA subset for factor analysis was assessed using the Kaiser-Meyer-Olkin (KMO) measure [25]. The number of factors to retain was guided by multiple criteria including Kaiser’s criterion (eigenvalues > 1), scree plot examination [26], and parallel analysis [27]. EFA was performed using Minimum Residual (minres) estimation with an oblique (oblimin) rotation to allow factors to correlate.
Item response modeling and score scale construction.
Once the dimensional structure was understood from MSA/EFA, we modeled item responses with graded response models (GRM) [28]. Guided by the EFA, we conducted confirmatory model selection on the remaining 80% confirmation subset, comparing: (i) a unidimensional GRM, (ii) a correlated-factor confirmatory GRM with two correlated factors (’Physical Functioning’ and ’Well-being and Mental Health’), and (iii) a bifactor GRM with one general frailty factor θ (common to all items) and two orthogonal specific factors corresponding to Physical Functioning and Well-being/Mental Health.
Model fit for all confirmatory models was assessed using the Root Mean Square Error of Approximation (RMSEA), the Standardized Root Mean Square Residual (SRMSR), the Comparative Fit Index (CFI), and the Tucker-Lewis Index (TLI) [29,30]. After exploring different IRT structures, the selected models were refitted using the full dataset to obtain factor scores for DIF and survival analyses. All IRT models were fitted using the mirt package [31].
For transparency and later parameter interpretation, the functional form of the bifactor model is given below. Let θ denote the general frailty factor (loaded by all items), and and
represent orthogonal specific factors for Physical Functioning and Well-being/Mental Health, respectively. For item j, the graded response model specifies the cumulative item score probabilities
where are item slopes (discriminations) on the general/specific factors, and djk are ordered intercept (
). Item score probabilities follow from adjacent differences
with the convention . All three factors were assumed orthogonal. Slopes on non-applicable specific factors were fixed to 0.
Interpreting slopes and intercepts.
In Eq (1), the vector of slopes for item j is . Larger slopes indicate greater discrimination (steeper response curves). The ordered intercepts djk determine where the cumulative curves switch along the latent dimensions; for a given score boundary k, the location along the item’s most sensitive direction is
, where
. Intuitively,
is the point on the latent continuum (projected onto
) where the probability of responding with a score
is 0.5. Throughout, we report factor scores on the general frailty factor, scaled to mean 0 and variance 1 (higher
worse).
Score fairness.
To ensure the frailty score is fair and comparable across different groups, we investigated whether items functioned differently for various subgroups of the population. A DIF analysis was conducted for three demographic variables: sex (male vs. female), education level (Primary, Secondary, and University), and age. For the latter, participants were dichotomized at the sample median of 76 years, such that individuals aged 76 years or below comprised the younger group and those older than 76 years comprised the older group. To identify DIF items and anchor items (items assumed free of DIF, required for model identification), we used the drop-sequential anchor-purification procedure based on sample-size–adjusted BIC [32] implemented in the mirt package [31]. To gauge the practical magnitude of DIF, we reported the Expected-Score Standardized Difference (ESSD) [33], a Cohen’s d–style index. Following conventional thresholds [34], we interpreted , 0.50, and 0.80 as small, medium, and large DIF, respectively. All DIF analyses used models refit on the full dataset to maximize precision for group comparisons.
External validity.
To evaluate the prognostic validity of the frailty scores, we examined their association with all-cause mortality. Individuals were tracked up to 1,100 days after completing the questionnaire. Latent trait estimates of frailty () were obtained using expected a posteriori estimation [35] from the final IRT models refit on the full dataset. Score estimates were computed using both the bifactor model (general dimension θ estimates) and the unidimensional model.
For descriptive analysis, was classified into quartiles (Q1–Q4) and Kaplan–Meier survival curves were estimated to visualize differences in survival between frailty levels. To formally quantify associations, we fitted Cox proportional hazards regression models with time-to-death as the outcome. We adjusted for age, sex, and education to reduce potential confounding by demographic characteristics. In addition, we specified interaction terms between
and each of the covariates to test for potential effect modification, i.e. whether the strength of the association between frailty and mortality differs across age, sex, or educational groups. Continuous variables (age and
) were mean-centered prior to modeling to facilitate interpretation of main effect regression coefficients.
Results
A total of 22,389 three–month questionnaires were available from patients with ischemic or hemorrhagic stroke in 2021–2022. We excluded 2,531 questionnaires that were completed without the patient present (11.3%). After coding “don’t know” as missing, a further 388 records (1.7%) were excluded due to having more than three missing responses across the nine items. The final analytic sample comprised 19,470 patients (87.0% of the starting cohort). Demographic characteristics of the analytic sample are presented in Table 2, and item–level response distributions are shown in Fig 1.
NA indicates missing responses. Refer to Table 1 for item response descriptions of the digits for each item.
Dimensionality
MSA conducted on the full dataset provided strong evidence that the nine items form a unidimensional and robust scale, with an overall Loevinger’s HS of 0.614. All items and item pairs met scalability criteria based on the conventional lower bound threshold of 0.3. This suggests that the items can be conceptualized along a single latent continuum, where higher scores indicate a greater level of the underlying trait (worse recovery and overall well-being). Refer to Table 3 for results using different thresholds. 5.5% of the respondents had notable Guttman errors, suggesting that the hierarchical ordering of items holds for most individuals.
H coefficient standard errors are shown in parentheses. Scalable items are denoted by 1 at a given threshold, while 0 indicates the item was unscalable.
EFA was conducted on 20% of the data set to investigate the underlying factor structure of the nine questionnaire items. The KMO measure of sampling adequacy indicated that the data were suitable for factor analysis (KMO item-level scores ranged between 0.79 and 0.95). Based on multiple criteria, including the examination of scree plots, Kaiser’s criterion (eigenvalues > 1), and parallel analysis, a correlated-factor solution was deemed most appropriate. This correlated-factor solution, derived using minimum residual estimation and an oblimin rotation, accounted for 76% of the total variance. Overall model fit indices were: RMSR = 0.013, TLI = 0.968, and RMSEA = 0.093 (90% CI [0.087, 0.099]).
The first factor explained 43% of the variance, and the second factor explained 33%. Factor loadings, communalities (h2), uniquenesses (u2), and item complexity (com) are presented in Table 4. Item communalities were generally high (ranging from 0.407 to 0.988). The mean item complexity of 1.13 indicated a relatively simple structure with items tending to load primarily on a single factor. The two factors exhibited a moderate positive correlation (r = 0.57). This suggests that there are two related but distinguishable content areas or facets within the items.
Loadings < 0.20 are omitted for clarity.
Factor 1 was primarily defined by items concerning mobility and activities of daily living (Item 2: Mobility; Item 3: Toilet help; Item 4: Dressing help; Item 5: Dependent on support) and was labeled Physical Functioning. Factor 2 was characterized by items related to mood, fatigue, pain, general health perception, and return to previous activities (Item 1: Return to life/activities; Item 6: Down/depressed/anxious; Item 7: General health status; Item 8: Increased fatigue; Item 9: New pain) and was labeled Well-being and Mental Health. Loadings larger than 0.2 are shown in Table 4.
Item response modeling and score scale construction
Based on the results of the MSA and EFA analysis, the fit of several IRT models was compared, and the results are presented in Table 5. The models were fit on the 80% confirmation subset as previously explained.
The initial unidimensional IRT model, which posited that all items measure a single underlying construct, demonstrated a poor fit to the data. The RMSEA was 0.117 (90% CI [0.115, 0.120]), and the SRMSR was 0.094, both exceeding commonly accepted thresholds for good fit [36]. The CFI (0.936) and TLI (0.914) were also poor. These results strongly suggested that a multidimensional structure would better represent the data.
A confirmatory correlated-factor GRM was constructed based on the Physical Functioning and Well-being/Mental Health structure from the EFA. This model showed acceptable fit, with RMSEA = 0.059 (90% CI [0.056, 0.061]), SRMSR = 0.054, CFI = 0.984, and TLI = 0.978. The two latent factors were highly correlated (r = 0.78), suggesting a strong overarching general dimension.
Finally, a bifactor GRM was estimated on the 80% confirmation subset, including one general factor (all items) plus the two specific factors identified above. All factors were assumed uncorrelated. This model provided the best overall fit, with RMSEA = 0.026 (90% CI [0.023, 0.030]), SRMSR = 0.033, CFI = 0.998, and TLI = 0.996 (Table 5). The bifactor structure effectively addressed the high correlation between the CFA factors by modeling their shared variance through a dominant general dimension.
Item-level interpretation.
Linking back to Eq (1), Table 6 reports each item’s discrimination (slopes on the general and specific factors) and the ordered intercepts djk for the fitted bifactor model. By model definition, specific factor slopes are 0 for items assumed not measuring on that factor. Larger general slopes ajg indicate that the item is more sensitive to differences in the frailty score. In particular, item 3 (Toilet help) shows very large slopes, which likely reflects both the relatively low frequency of option 2 responses (Fig 1) and the fact that patients actually receiving help when using the toilet tend to either be high on the mobility dimension (reflecting low mobility), have high frailty, or both with very few exceptions. For interpretability along the general frailty dimension, we also report the general–axis thresholds . Each
is the value of θ where the probability of responding at or above level k equals 0.5 when the specific factors are fixed at zero. Note that since θ,
and
are assumed to be standard normal (mean 0, SD 1 in the reference groups), a patient at
is roughly one standard deviation frailer than the sample average. Items with b(g) close to a patient’s θ are the most informative about that individual’s position on the frailty continuum. The physical functioning items (2–5) are most informative from moderate to high frailty. The other items provide information throughout the entire score scale but mainly in the mid to upper score range.
Values are estimates with SE in parentheses. ajg = general factor slope; ajs1 = Physical Functioning specific slope; ajs2 = Well-being/Mental Health specific slope. General–axis thresholds are computed with specific factors fixed at 0.
Score fairness
The expected item scores (ranging from 0 to the number of possible item responses) as functions of θ for each demographic group (age, education and sex) are shown in Fig 2. The items for which the curves differ between groups were identified as DIF items by the drop-sequential DIF analysis, while the others were used as anchor items for model identification. Visual inspection already suggests that most items behave similarly between groups. Using ESSD as an effect size metric (Table 7), items that exceed the ’small’ DIF threshold of are found mainly for the age variable. Exceptions are item 8 for primary vs. university education and item 6 for sex with
values barely exceeding 0.2. Item 6 and item 9 show medium DIF for age with
. Overall, the questionnaire demonstrates measurement invariance across sex and education, but small-to-moderate age-related bias on items 1, 6, 8 and 9.
The expected item score reflects a weighted average of possible responses based on their estimated probabilities.
Table 8 presents the subgroup-specific latent means and variances after anchoring on the items without DIF. Note that the score scale is assumed standard normal (in the reference groups) by assumptions of the model fitting algorithm. Thus, it theoretically ranges from to
, but our estimated values are all within the range (–2.6,2.6). Respondents older than 76 years exhibited a mean
that was 0.45 standard deviations higher than those 76 or younger, and females had a mean
0.28 standard deviations higher than males—both indicating worse recovery/well-being in these groups. In contrast, respondents with secondary or university education had mean
values 0.17 and 0.29 standard deviations lower than those with only primary education, respectively, suggesting comparatively better recovery, likely confounded by other variables. Estimated variances were similar within subgroups of education level, but slightly lower for females and people over 76 years of age compared to their counterparts.
External validity
Fig 3 displays Kaplan–Meier survival curves stratified by quartile of the unidimensional and bifactor model scores. There is a clear, monotonic decline in survival probability from Q1 (lowest
) through Q4 (highest
) over the 1,100-day follow-up after questionnaire completion. The survival probability is similar between the Q1 and Q2 groups, but decreases more rapidly for the subsequent quartiles, and the patterns were nearly identical for the unidimensional and bifactor-derived scores. The unidimensional
quartiles used for survival analysis corresponded to scores of: Q1 (<–0.44), Q2 (–0.44 to 0.23), Q3 ([0.23 to 0.85), and Q4 (>0.85). The bifactor
quartiles were: Q1 (<–0.71), Q2 (–0.71 to 0.04), Q3 ([0.04 to 0.57), and Q4 (>0.57). One should note that these scores are unitless and not directly comparable. A unidimensional model score represents a composite of the general trait mixed with specific factors, whereas the bifactor score mathematically separates these components to measure a more “purified” general trait.
Quartiles Q1–Q4 correspond to increasing (worse recovery); shaded bands are 95% CIs and vertical lines are events (death).
In Cox regression analyses, both scores were strongly associated with all-cause mortality. In models adjusted for age, sex, and education, each one-unit increase in corresponded to an approximately 80% increase in the hazard of death for a 74.12 year old (average age in the dataset) male with primary school education (unidimensional: HR = 1.76, 95% CI: 1.62–1.90; bifactor: HR = 1.83, 95% CI: 1.68–1.99; Table 9). Results were consistent across the two scoring approaches, reflecting their high correlation (r = 0.98). The models indicated modest evidence of effect modification (Table 9). The association between frailty and mortality was somewhat stronger in females than in males (e.g., bifactor: HR for
female = 1.13, 95% CI: 1.04–1.24) and among respondents with university education compared to those with primary education (HR for
university = 1.18, 95% CI: 1.05–1.33). Interaction with age was negligible. These differences did not materially alter the overall conclusion that higher frailty scores were associated with increased mortality risk. Note that relatively small effects are also significant due to the large sample size. We evaluated the linearity assumption for the continuous predictors, age and θ, by contrasting spline-based Cox models with models using linear terms. As no substantial deviations from linearity were observed, we retained the linear specification for ease of interpretation.
Age and θ scores were centered before fitting the Cox models for easier main effect interpretation.
Discussion
This study shows that a psychometrically robust, fair, and prognostically valid single-number frailty score can be derived from PROMs three months post stroke. Using MSA, EFA, and bifactor IRT modeling, we identified a dominant general frailty dimension alongside two clinically relevant subdomains, Physical Functioning and Well-being/Mental Health. DIF analyses demonstrated broad measurement invariance across sex and education, with modest age-related item bias. Importantly, the score exhibited strong external validity: Kaplan–Meier curves showed monotonic separation by frailty quartiles (Fig 3), and in Cox models adjusted for age, sex, and education, a higher frailty score was associated with substantially increased risk of death.
Our findings extend previous work on the measurement of frailty in stroke. Traditional approaches such as the Fried phenotype, Frailty Index, or administrative tools such as eFI and HFRS are largely rated by clinicians or based on coded data [14–16]. By contrast, the present score is derived directly from PROMs already collected at scale in Riksstroke, making measurement highly efficient in terms of cost and time. Previous studies of PROMs in stroke have applied IRT to new scale development [20], examined multidimensional structures [37], and shown that PROMs predict future use of healthcare care [21]. Our work adds to this literature by developing a frailty score, testing its fairness, and validating its prognostic value within a national registry.
Several limitations merit consideration. First, we did not compare our score with established frailty instruments, which limits direct comparability between measurement approaches and prevents the definition of specific clinical cut-offs (e.g., for mild, moderate, or severe frailty) without further calibration against these standards. Second, although DIF was generally small, some age-related differences were observed, particularly for mood and pain items; these may reflect true age-related vulnerability or attribution biases. Third, a notable exclusion from our score is a formal assessment of cognitive impairment. Although Riksstroke includes a screening item for memory and concentration, we omitted it to ensure a more homogeneous physical-physiological frailty construct. Cognition is undoubtedly a component of frailty in a broader sense, but including a single self-reported cognitive item could have compromised the unidimensionality of the scale. Future studies should evaluate how our physical frailty score interacts with objective cognitive screens. Lastly, “don’t know” responses were treated as missing, which may not fully capture patient uncertainty.
The strengths of this study include the large and nationally representative sample and the rigorous psychometric evaluation. Because Riksstroke captures nearly all patients hospitalized with acute stroke in Sweden, the findings are likely to be generalizable to similar populations in other high-income settings with comprehensive registry infrastructures. Nevertheless, cultural and health system differences may influence item interpretation and score performance; replication in other contexts is needed.
To maximize the clinical utility of these findings, the strong association between frailty quartiles and survival (Fig 3) suggests this score could form the basis for future risk stratification tools. Validated risk tables estimating, for example, 3-year survival probabilities—could eventually be integrated into the registry feedback to aid clinicians and families in long-term care planning. Moreover, because this assessment captures patient-reported deficits in mood and fatigue that are often overlooked in brief clinical encounters, a high frailty score could serve as a ’red flag’ to, for example, trigger a Comprehensive Geriatric Assessment or a review of home-care needs. Future studies should specifically link this score to healthcare utilization data, such as hospital readmissions and institutionalization, to further validate its value in guiding resource allocation.
In summary, we developed and validated a frailty score from PROMs in the Riksstroke three-month follow-up questionnaire. The score reflects a strong general frailty dimension, demonstrates fairness across demographic groups, and is associated with mortality. Beyond this specific setting, our findings illustrate how routinely collected PROM data can be leveraged to construct psychometrically sound and prognostically meaningful frailty measures. Such approaches may complement or substitute clinician-rated instruments, offering scalable ways to monitor recovery and guide quality improvement in stroke care. Future studies should test the generalizability of this strategy across different health systems, cultures, and measurement contexts, and benchmark PROM-derived frailty scores against established instruments.
References
- 1. GBD 2019 Stroke Collaborators. Global, regional, and national burden of stroke and its risk factors, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019. Lancet Neurol. 2021;20(10):795–820. pmid:34487721
- 2. Asplund K, Hulter Åsberg K, Appelros P, Bjarne D, Eriksson M, Johansson A, et al. The Riks-Stroke story: building a sustainable national register for quality assessment of stroke care. Int J Stroke. 2011;6(2):99–108. pmid:21371269
- 3. Naeem F, Quinn T. Frailty in stroke. Practical Neurology. 2024;24(6):448–55. pmid:38772626
- 4. Evans NR, Pinho J, Beishon L, Nguyen T, Ganesh A, Balasundaram B, et al. Frailty and stroke: global implications for assessment, research, and clinical care-A WSO scientific statement. Int J Stroke. 2025;20(8):905–17. pmid:40390672
- 5. Evans NR, Minhas JS, Beishon LC, Quinn TJ. Stroke medicine is frailty medicine: clinical and research priorities for frailty in stroke. Cerebrovasc Dis. 2025;54(6):910–7. pmid:40090312
- 6. Evans NR, Fearon P, Beishon L, Pinho J, Quinn TJ. The importance of frailty in stroke and how to measure it. Stroke. 2025;56(1):e8–11. pmid:39234696
- 7. Fried LP, Tangen CM, Walston J, Newman AB, Hirsch C, Gottdiener J, et al. Frailty in older adults: evidence for a phenotype. The Journals of Gerontology Series A: Biological Sciences and Medical Sciences. 2001;56(3):M146–57.
- 8. Morley JE, Malmstrom TK, Miller DK. A simple frailty questionnaire (FRAIL) predicts outcomes in middle aged African Americans. J Nutr Health Aging. 2012;16(7):601–8. pmid:22836700
- 9. Mitnitski AB, Mogilner AJ, Rockwood K. Accumulation of deficits as a proxy measure of aging. ScientificWorldJournal. 2001;1:323–36. pmid:12806071
- 10. Clegg A, Bates C, Young J, Ryan R, Nichols L, Ann Teale E, et al. Development and validation of an electronic frailty index using routine primary care electronic health record data. Age Ageing. 2016;45(3):353–60. pmid:26944937
- 11. Gilbert T, Neuburger J, Kraindler J, Keeble E, Smith P, Ariti C, et al. Development and validation of a Hospital Frailty Risk Score focusing on older people in acute care settings using electronic hospital records: an observational study. Lancet. 2018;391(10132):1775–82. pmid:29706364
- 12. Rockwood K, Song X, MacKnight C, Bergman H, Hogan DB, McDowell I, et al. A global clinical measure of fitness and frailty in elderly people. CMAJ. 2005;173(5):489–95. pmid:16129869
- 13. Rockwood K, Theou O. Using the clinical frailty scale in allocating scarce health care resources. Can Geriatr J. 2020;23(3):210–5. pmid:32904824
- 14. Clegg A, Young J, Iliffe S, Rikkert MO, Rockwood K. Frailty in elderly people. Lancet. 2013;381(9868):752–62. pmid:23395245
- 15. Tang X-X, Wang H, Yang J, Gu P, Zhang X-M, Tang Q-Y, et al. A comparative analysis of three frailty assessment tools for hospitalized patients with stroke. Clin Neurol Neurosurg. 2024;246:108600. pmid:39447223
- 16. Ng SSM, Chen P, Liu TW, Li CHL, Tang THH, Lau JLY, et al. Assessment of the psychometric properties of the frailty and injuries: cooperative studies of intervention techniques - 4 in people with stroke. Top Stroke Rehabil. 2025;32(7):712–22. pmid:39981986
- 17. Söderholm A, Stegmayr B, Glader E-L, Asplund K, Riksstroke Collaboration. Validation of hospital performance measures of acute stroke care quality. riksstroke, the Swedish Stroke Register. Neuroepidemiology. 2016;46(4):229–34. pmid:26975057
- 18. Palmcrantz S, Sommerfeld DK. Development and validation of the Swedish national stroke register Riksstroke’s questionnaires in patients at 3 and 12 months after stroke: a qualitative study. BMJ Open. 2018;8(2):e018702. pmid:29472261
- 19.
Hambleton RK, Swaminathan H. Item response theory. Netherlands: Springer; 1985. https://doi.org/10.1007/978-94-017-1988-9
- 20. Luo Y, Yang J, Zhang Y. Development and validation of a patient-reported outcome measure for stroke patients. Health Qual Life Outcomes. 2015;13:53. pmid:25953508
- 21. Katzan IL, Thompson N, Schuster A, Wisco D, Lapin B. Patient-reported outcomes predict future emergency department visits and hospital admissions in patients with stroke. J Am Heart Assoc. 2021;10(6):e018794. pmid:33666094
- 22.
Osterlind SJ, Everson HT. Differential item functioning. Thousand Oaks, CA: Sage Publications; 2009.
- 23.
Mokken RJ. A theory and procedure of scale analysis. Berlin, New York: De Gruyter Mouton; 1971. https://doi.org/10.1515/9783110813203
- 24.
Sijtsma K, Molenaar IW. Introduction to nonparametric item response theory. Thousand Oaks, CA: Sage Publications; 2002.
- 25. Kaiser HF. A second generation Little Jiffy. Psychometrika. 1970;35(4):401–15.
- 26. Cattell RB. The scree test for the number of factors. Multivariate Behav Res. 1966;1(2):245–76. pmid:26828106
- 27. Horn JL. A rationale and test for the number of factors in factor analysis. Psychometrika. 1965;30:179–85. pmid:14306381
- 28. Samejima F. Estimation of latent ability using a response pattern of graded scores. Psychometrika. 1969;34(S1):1–97.
- 29. Hu L, Bentler PM. Cutoff criteria for fit indexes in covariance structure analysis: conventional criteria versus new alternatives. Structural Equation Modeling: A Multidisciplinary Journal. 1999;6(1):1–55.
- 30. Browne MW, Cudeck R. Alternative ways of assessing model fit. Sociological Methods & Research. 1992;21(2):230–58.
- 31. Chalmers RP. mirt: a multidimensional item response theory package for the R environment. J Stat Soft. 2012;48(6).
- 32. Schwarz G. Estimating the dimension of a model. Ann Statist. 1978;6(2).
- 33. Meade AW. A taxonomy of effect size measures for the differential functioning of items and scales. J Appl Psychol. 2010;95(4):728–43. pmid:20604592
- 34.
Cohen J. Statistical power analysis for the behavioral sciences. 2nd ed. Hillsdale, NJ: Lawrence Erlbaum Associates; 1988. http://dx.doi.org/10.4324/9780203771587
- 35.
Baker FB, Kim S-H. Item response theory. CRC Press; 2004. https://doi.org/10.1201/9781482276725
- 36. Schreiber JB, Nora A, Stage FK, Barlow EA, King J. Reporting structural equation modeling and confirmatory factor analysis results: a review. The Journal of Educational Research. 2006;99(6):323–38.
- 37. Wu S-Y, Li Y-C, Chen Y-W, Chen C-L, Pan H-C, Lin K-C, et al. Construct validity, responsiveness, minimal detectable change, and minimal clinically important difference of the stroke self-efficacy questionnaire in individuals receiving stroke rehabilitation. Disabil Rehabil. 2024;46(25):6195–203. pmid:38433459