Predicting the functional outcome of intensive inpatient rehabilitation after stroke: results from the RIPS Study

Abstract

BACKGROUND

The complexity of stroke sequelae, the heterogeneity of outcome measures and rehabilitation pathways, and the lack of extensively validated prediction models represent a challenge in predicting stroke rehabilitation outcomes

AIM

To prospectively investigate a multidimensional set of variables collected at admission to inpatient post-stroke rehabilitation as potential predictors of the functional level at discharge.

DESIGN

Multicentric prospective observational study.

SETTING

Patients were enrolled in four Intensive Rehabilitation Units (IRUs).

POPULATION

Patients were consecutively recruited in the period December 2019-December 2020 with the following inclusion criteria: aged 18+, with ischemic/haemorrhagic stroke, and undergoing inpatient rehabilitation within 30 days from stroke.

METHODS

This is a multicentric prospective observational study. The rehabilitation pathway was reproducible and evidence-based. The functional outcome was disability in activities of daily living, measured by the modified Barthel Index (mBI) at discharge. Potential multidimensional predictors, assessed at admission, included demographics, event description, clinical assessment, functional and cognitive profile, and psycho-social domains. The variables statistically associated with the outcome in the univariate analysis were fed into a multivariable model using multiple linear regression.

RESULTS

A total of 220 patients were included (median [IQR] age: 80 [15], 112 women, 175 ischemic). Median mBI was 26 (43) at admission and 62.5 (52) at discharge. In the multivariable analysis younger age, along with better functioning, fewer comorbidities, higher cognitive abilities, reduced stroke severity, and higher motor functions at admission, remained independently associated with higher discharge mBI. The final model allowed a reliable prediction of discharge functional outcome (adjusted R²=77.2%).

CONCLUSIONS

The model presented in this study, based on easily collectable, reliable admission variables, could help clinicians and researchers to predict the discharge scores of the global functional outcome for persons enrolled in an evidence-based inpatient stroke rehabilitation program.

CLINICAL REHABILITATION IMPACT

A reliable outcome prediction derived from standardized assessment measures and validated treatment protocols could guide clinicians in the management of patients in the subacute phase of stroke and help improve the planning of the rehabilitation individualized project.

The increasing effectiveness and diffusion of thrombolytic and thrombectomy treatments are not sufficient yet to completely avoid residual neurological deficits in stroke patients, and recanalization strategies are not always applicable.^{1, 2} Therefore, despite recent improvements in acute phase management, stroke remains a major cause of disability.³ Given these premises, post-acute rehabilitation maintains a pivotal role in allowing patients to equally access adequate rehabilitation treatment and obtain a satisfactory functional outcome.^4-6

However, the economic resources available in healthcare organisations are limited,⁷ thus it is essential to intervene on treatment and assessment processes to maximize functional recovery and thus reduce the costs deriving from residual disabilities.^{8, 9} The first and necessary step is the introduction of intervention protocols within an evidence-based Integrated Care Pathway (ICP).¹⁰ Additionally, evidence-based treatments should be complemented by the standardisation of measures, including a reproducible and complete functional assessment¹¹ performed by an interdisciplinary team¹² before and after treatment. In fact, it is essential to apply standardized and validated protocols for an in-depth evaluation of rehabilitation progression and outcomes and a coherent modification of the treatment plans.

It is in this optic that predictive models of the functional outcome in stroke patients eligible for rehabilitation¹³ are developed. Such data-based predictive tools aim to support clinical decisions to improve several processes in the care of post-acute stroke patients. Their application would enable a more accurate selection of the most efficient intensive rehabilitation path, already in the subacute phase.¹⁴ Further, the possibility to estimate the days of hospitalisation necessary to achieve the pre-established functional goals¹⁵ and to predict the extent of the need for the assistance required by patients upon returning home¹⁶ may improve rehabilitation and service planning. Finally, identifying the profile of response to treatment may allow the customisation of the approach based on specific deficits, with targeted treatments (“tailored approach”).^17-19

The functional outcome in post-stroke rehabilitation has been extensively studied in the literature, often focusing on individual functional aspects as well as on complete recovery.^20-26 Many studies described the outcomes of various treatments and evaluated predictors of recovery, but with a limited ability to generalise the results obtained due to a non-comparability of the data.^27-30 Multiple reasons could be attributed to this limitation, such as the heterogeneity of the applied rehabilitation approaches, or the inconsistency of the baseline and outcome assessments of the patients.

For instance, García-Rudolp et al.,³¹ in their paper, systematically evaluated studies assessing the effect of a wide range of interventions for stroke rehabilitation, but the inconsistency in measuring outcomes hindered the possibility to compare the results obtained from different treatments. Some studies failed to accurately describe the adopted rehabilitation protocols and the choice of outcome variables only partially reflected post-stroke patients’ independence.^{32, 33} It has been stated in many papers^{34, 35} that an improvement in methodological quality in stroke rehabilitation-related studies is essential, especially regarding the reproducibility and reliability of patients’ evaluation at different time points. Moreover, Salter et al.³⁶ highlighted the considerable heterogeneity in functional assessment used in literature, which hinders the interpretation of data deriving from randomised controlled trials concerning stroke rehabilitation.

Within a knowledge-translation approach, Don Carlo Gnocchi Foundation, one of the largest Italian scientific research and rehabilitation, has recently developed and implemented an evidence-based interdisciplinary ICP for post-acute stroke inpatient rehabilitation,¹⁰ including a multidimensional assessment protocol based on validated tools. After a pilot study confirming the feasibility and suggesting improved outcomes of the ICP compared to previous practice, the ICP has been implemented in four of the Don Carlo Gnocchi Foundation intensive rehabilitation units (IRUs), in order to standardise the outcome definition and the process of care according to national and international stroke rehabilitation guidelines.^{6, 10}

The RIPS (inpatient rehabilitation post-stroke) study is a multicentre prospective study, involving the above-mentioned IRUs where the evidence-based stroke ICP had been applied; the study protocol has been extensively described elsewhere.³⁷ RIPS aimed to study multiple features and outcomes in the context of patients admitted to ICPs after stroke. This analysis is focused on answering the main question of RIPS, that is to investigate which features recorded from the multidimensional assessment performed at admission to intensive post-stroke rehabilitation, may result as independent predictors of the functional outcome at discharge, measured with the modified Barthel Index (mBI).

Materials and methods

All patients admitted to four IRUs that applied a shared stroke rehabilitation ICP (Florence, Massa, Fivizzano, and La Spezia) were systematically assessed for eligibility from December 2019 to December 2020. Patients aged 18+ years who had suffered an ischemic or haemorrhagic stroke within 30 days before admission were considered eligible for this study, and all those signing informed consent to participate were systematically enrolled. Patients with a transitory ischemic attack, and with severe hemorrhagic or ischemic stroke (inducing disorders of consciousness states and critical clinical care conditions) addressed to the severe brain injury high-complexity rehabilitation ward were excluded. Measures regarding clinical information, nutritional, functional, neurological, and neuropsychological assessments, were taken at admission (T0), at discharge (T1), and at an in-person or telephone follow-up 6 months after the stroke (T2).

The study was approved by the local ethical committee (CEAVC Em. 2021-007 ID 14513_bio) and registered on ClinicalTrials.gov (registration number: NCT03866057, https://clinicaltrials.gov/ct2/show/study/NCT03866057?term=fondazione+don+gnocchi%2C+stroke&draw=2&rank=1).

As to the sample size, calculation was performed using the following equation:

(1)

a = equal to 1.96 assuming a significance level of 0.05.

b = equal to 1.28 assuming a power of 90%

σ² = population variance (standard deviation)

MCID = Minimal Clinically Important Difference for the modified Barthel Index

Data assumptions for the population variance were obtained from data retrospectively collected between 2015 and 2017 at two participating hospitals (Massa and Fivizzano). The 527 patients enrolled in the retrospective study showed a standard deviation on the modified Barthel Index at the admission of 29 points. Those data were presented at the XIX SIRN National Congress (Perugia, 4th-6th April 2019). To determine the MCID, the value proposed by Hsieh et al.³⁸ for the Barthel Index in stroke patients was adopted. The value was adjusted according to the score range (100 points) of the modified scale version, leading to an MCID of 9 points.

(2)

A sample size of 218 patients was obtained.

Considering a 10% increase to account for possible drop-outs the total sample size was estimated in 240 patients.

For this analysis, we focused on data collected in person at T0 and T1. The reporting of this study followed the STROBE checklist for observational studies.³⁹

Measures and analyses

In this work, the functional status of patients at discharge from the IRUs (T1), as measured with the modified Barthel Index (mBI) total score, was assumed as the study outcome.

The factors analysed for association with the outcome were collected at admission, addressing five different domains: Demographics, Description of the event, Clinical features, Psycho-social features and Functional profile (body functions and activity).

Specifically, the detail of each considered variable for the identified domains is reported in Table I.

Table I. —List of variables, for each domain, considered for association with the outcome.

Domain	Variable evaluated
Demographics	Age
	Gender
	Educational level
	Cohabitation
Description of the event	Aetiology (ischemic or hemorrhagic stroke)
	Time from the event
	Recurrent event
	Side of the lesion
	Area of the lesion
	Acute phase treatment (i.e. thrombolysis or thrombectomy)
Clinical features	Markers of complexity Pain Acute infection Dysphagia Malnutrition Nasogastric tube or percutaneous enterostomy Pressure ulcers Bladder catheter Incontinence Central venous catheter Tracheostomy cannula Reduced vigilance Delirium Clinical instability Depression
	Anemia
	Dialysis
	Agitation
	Diplopia
	CIRS comorbidity index
	BMI
	NIHSS, Total score and item 9 sub-score (Aphasia)
	SDC
Psycho-social features	Adjusted score of the MoCA
	Equivalent score of the OCS heart test
	mFWC
	FAI
Functional profile	mBI
	TCT
	FAC
	mRS, baseline and anamnestic score
	SPPB
	MAS
	FMA

CIRS: Cumulative Index Rating Scale; BMI: Body Mass Index; NIHSS: National Institutes of Health Stroke Scale; SDC: Communication Disability Scale; MoCA: Montreal Cognitive Assessment; OCS: Oxford cognitive screen; mFWC: modified Functional Walking Classification; FAI: Frenchay Activity Index: mBI: Modified Barthel Index; TCT: Trunk Control Test; FAC: Functional Ambulation Categories; mRS: Modified Rankin Score; SPPB: Short Physical Performance Battery; MAS: Modified Ashworth Scale; FMA: Fugl-Meyer Assessment.

Details concerning each evaluation tool reference are reported as Supplementary Digital Material 1, Supplementary Table I.

Statistical analysis

The variables were first analysed by the multi-professional clinical team for clinical imputation of missing data, wherever possible. Specifically, the variable of Montreal Cognitive Assessment (MoCA) was imputed considering values of 0, based on the assumption of the impossibility to administer the scale, in case of the absence of communication (Communication Disability Scale, SDC=0) or presence of reduced vigilance/coma (marker of complexity=1) or presence of delirium (marker of complexity=1) or presence of global aphasia (National Institutes of Health Stroke Scale, NIHSS_9=3).

Subsequently, the variables were screened according to the percentage of missing data. More in detail, variables not reaching at least 80% completeness were discarded. Moreover, categorical variables with frequencies on at least one category not reaching the minimum of ten cases were discarded.

Finally, on the remaining variables, an automatic k-Nearest Neighbours-based single imputation method⁴⁰ was performed on Python for filling in missing data that could not be attributed with values during the clinical screening. Further, variables screening was clinically performed excluding those variables with overlapping information with others. Additional detail on the excluded variables is presented in Table I.

The statistical analyses were performed on IBM Corp. Released 2021. IBM SPSS Statistics for Windows, Version 28.0. Armonk, NY: IBM Corp. Firstly, descriptive analyses were performed using the mean and standard deviation (std), or median and interquartile range (IQR) when appropriate, for numerical variables and relative frequencies for the categorical variables. Comparisons over time of numerical variables were performed using the Wilcoxon Test.

The univariate analyses were performed using Pearson’s or Spearman’s correlation and the t-test of the Mann-Whitney Test, for numerical and categorical variables, respectively. Finally, the variables that resulted statistically associated with the outcome in the univariate step were fed into a multivariable model. The multivariate analysis was performed using multiple linear regression. The normality of the distributions was tested through the Shapiro-Wilk Test.

A simple calculator is available in the Supplementary Digital Material 2, offering the mBI estimation at discharge based on the coefficients obtained in this multivariable regression model. In order to have a better understanding of where the calculator might fail or succeed the most, error analyses were computed by comparing the baseline characteristics of patients with good estimated or wrongly estimated outcomes.

Specifically, the estimated values were further analysed by dividing the patients into three groups based on the values of the differences between estimated and measured mBI scores at discharge. In the absence of an established reference to define a cut-off for a Minimal Clinical Important Difference (MCID) in the mBI, we estimated a cut-off of 10 utilising the MCID=1.85 proposed for the Barthel Index Scale (scoring 0-20).³⁸ Then, we generated the three groups, considering the difference between the real and estimated mBI values (estimation errors). Patients with estimation errors below -10 were attributed to the group of overestimated mBI values, over 10 were identified as underestimated values, and the others were considered as correctly estimated. Lastly, the descriptive analyses of the variables in the multivariable model were repeated for each group and statistical comparisons of the variables among groups were presented. Specifically, the Kruskal-Wallis Test, with pairwise comparisons, was applied for numerical variables and the χ² Test was used for categorical ones. For each test applied a P value <0.05 was considered statistically significant.

Results

Out of the 278 stroke patients who were eligible for the RIPS study, 235 (85%) signed the informed consent and were enrolled. Fifteen patients (0.06%) were subsequently excluded since their discharge mBI (our primary outcome) had not been recorded, due to anticipated discharge preventing discharge evaluation (T1). Thus, data on 220 patients (80% of all eligible patients) were included in the analyses of this work. Of these, 125 were recruited from the centre in Florence, 34 from the centre in Massa, 13 from the centre in Fivizzano, and 48 from La Spezia.

The main characteristics of the included sample are presented in Table II.⁴¹ The patients reported a median [IQR] age of 80.0 [15.0] years, with 50.9% of women, and a prevalence of ischemic aetiology (79.5%). The mBI of patients significantly increased between admission and discharge (P value <0.001), passing from a median value of 26.0 [43.0] to 62.5 [52.0].

Table II. —Description of the sample included in the analyses.

Variables	Mean (std)/ Median [IQR] or Frequencies (%)a	N.
Age	80.0 [15.0]	220
Gender (male)	108 (49.1%)	220
Educational level (years)	8.0 [8.0]	205
Cohabitation (yes)	144 (74.6%)	193
Center	Florence: 125 (58.8%) Massa: 34 (15.5%) Fivizzano: 13 (5.9%) La Spezia: 48 (21.8%)	220
Aetiology (ischemic)	175 (79.5%)	220
Time from the event	11.0 [8.0]	220
Recurrent event (yes)	30 (13.8%)	217
Side of the lesion	Right: 96 (46.4%) Left: 96 (46.4%) Both: 15 (7.2%)	207
Area of the lesion	None: 16 (7.3%) Sub-tentorial: 24 (10.9%) Supra-tentorial: 172 (78.2%) Both: 8 (3.6%)	220
Thrombolysis (yes)	50 (28.9%)	173
Reduced vigilance and coma (yes)	14 (6.4%)	220
Clinical instability (yes)	21 (9.5%)	220
Delirium (yes)	12 (5.5%)	220
Acute infection (yes)	26 (11.8%)	220
Depression (yes)	58 (26.4%)	220
Dysphagia (yes)	117 (53.2%)	220
Malnutrition (yes)	13 (5.9%)	220
BMI	25.1 [5.08]	175
Bedsores (yes)	26 (11.8%)	220
Bladder catheter (yes)	90 (40.9%)	220
Incontinence (yes)	85 (38.6%)	220
Central venous catheter (yes)	9 (4.1%)	220
Tracheostomy (yes)	1 (0.5%)	220
Anaemia (yes)	76 (34.5%)	220
Dialysis (yes)	4 (1.8%)	220
Agitation (yes)	13 (6.5%)	201
Pain (yes)	47 (21.4%)	220
Diplopia (yes)	6 (2.7%)	219
CIRS_comorbidity	3.0 [2.0]	212
NIHSS	7.0 [8.0]	218
Aphasia (NIHSS_9)	No aphasia: 138 (63.0%) Mild to moderate aphasia: 34 (15.5%) Severe aphasia: 33 (15.1%) Mute or global aphasia: 14 (6.9%)	219
SDC	3.0 [2.0]	220
MoCA_adjusted_Santangelo*	18.73 [7.53]	191
OCS Space Asymmetry_raw	0.0 [4.0]	115
OCS Space Asymmetry_cut-off	Normal: 76 (66.1%) Right neglect: 14 (12.2%) Left neglect: 25 (21.7%)	115
mFWC	6.0 [0.0]	133
FAI	28.0 [8.0]	132
mBI	26.0 [43.0]	219
TCT	48.0 [75.0]	209
FAC	0.0 [2.0]	210
SPPB	0.0 [2.0]	206
mRS	4.0 [1.0]	209
mRS_anamnestic	0.0 [1.0]	208
MAS	0.0 [2.0]	200
FMA	158.0 [80.0]	183
LoS (days)	32.0 [28.0]	216
mBI	62.5 [52.0]	220

BMI: Body Mass Index; CIRS: Cumulative Index Rating Scale; FMA: Fugl-Meyer Assessment; IQR: Interquartile Range; LoS: length of stay; MAS: Modified Ashworth Scale; mBI: Modified Barthel Index; mFWC: Modified Functional Walking Classification; MoCA: Montreal Cognitive Assessment; mRS: Modified Rankin Score; N.: numerosity; NIHSS: National Institutes of Health Stroke Scale; OCS: Oxford cognitive screen; SDC: Communication Disability Scale; SPPB: Short Physical Performance Battery; std: standard deviation; TCT: Trunk Control Test. ^aDichotomous variables are presented with the frequency of one class only, indicated between brackets. *MoCA adjusted score on normative values by Santangelo et al.⁴¹

For what concerns the clinical imputation, the number of missing data on the MoCA reduced from 35% to 13.2%. After this process, among the considered variables, thrombolysis, Body Mass Index (BMI), Frenchay Activity Index (FAI), modified Functional Walking Classification (mFWC), and Oxford Cognitive Screening (OCS) heart test were excluded for insufficient completeness level (Table II). In addition, diplopia, and the complexity markers central venous catheter, tracheostomy, and dialysis were excluded from the analyses for insufficient cases in one of the categories (Table II).

From the univariate analyses (Table III), all variables in the Psycho-social features and Functional and cognitive profile domains and all variables in the Demographics and Clinical features domain, except for educational level, agitation, anaemia, and depression, resulted significantly associated with the outcome. Instead, for what concerns the domain of the Description of the event, only the time from the event was significantly associated with the outcome.

Table III. —Univariate analyses. The statistically significant P values are reported in bold.

Variables	mBI median [IQR] on the groups or Correlation coefficient (R)	P value
Demographics
Age	-0.361	<0.001
Gender	Male: 74.0 [48.0] Female: 52.5 [59.0]	0.009
Educational level (years)	0.065	0.337
Cohabitation	Yes: 72.0 [56.0] No: 52.0 [54.0]	0.046
Description of the event
Aetiology	Ischemic: 63.0 [54.0] Hemorrhagic: 60.0 [54.0]	0.502
Time from the event	-0.163	0.015
Recurrent event	No: 62.0 [52.0] Yes: 77.5 [62.0]	0.555
Side of the lesion	Right: 58.0 [50.0] Left: 63.0 [56.0] Bilateral: 86.0 [60.0]	0.192
Area of the lesion	None: 77.5 [47.0] Sub-tentorial: 76.0 [48.0] Supra-tentorial: 57.5 [59.0] Both: 66.5 [43.0]	0.100
Clinical features
Reduced vigilance and coma	Yes: 11.5 [40.0] No: 66.5 [51.0]	<0.001
Clinical instability	Yes: 16.0 [52.0] No: 67.0 [49.0]	<0.001
Delirium	Yes: 39.50 [54.0] No: 63.0 [53.0]	0.047
Acute infection	Yes: 36.5 [65.0] No: 67.0 [51.0]	0.003
Depression	Yes: 56.5 [49.0] No: 63.0 [53.0]	0.659
Dysphagia	Yes: 48.0 [60.0] No: 79.0 [42.0]	<0.001
Malnutrition	Yes: 13.0 [48.0] No: 66.0 [51.0]	<0.001
Bedsores	Yes: 17.5 [44.0] No: 69.0 [49.0]	<0.001
Bladder catheter	Yes: 37.5 [54.0] No: 79.0 [42.0]	<0.001
Incontinence	Yes: 43.0 [59.0] No: 77.0 [46.0]	<0.001
Anaemia	Yes: 58.0 [57.0] No: 63.0 [53.0 ]	0.596
Agitation	Yes: 55.0 [37.0] No: 63.0 [54.0]	0.864
Pain	Yes: 43.0 [58.0] No: 72.0 [52.0]	<0.001
CIRS_comorbidity	-0.181	0.007
NIHSS	-0.663	<0.001
Aphasia (NIHSS_9)	No aphasia: 74.0 [52.0] Mild to moderate aphasia: 53.0 [55.0] Severe aphasia: 52.0 [64.0]a Mute or global aphasia: 20.0 [39.0]b	<0.001
Psycho-social features
MoCA_adjusted_Santangelo*	0.502	<0.001
Functional profile
mBI	0.765	<0.001
TCT	0.721	<0.001
mRS_anamnestic	-0.137	0.043
MAS	-0.253	<0.001
FMA	0.788	<0.001

BMI: Body Mass Index; CIRS: Cumulative Index Rating Scale; FMA: Fugl-Meyer Assessment; IQR: Interquartile Range; LoS: length of stay; MAS: Modified Ashworth Scale; mBI: Modified Barthel Index; mFWC: Modified Functional Walking Classification; MoCA: Montreal Cognitive Assessment; mRS: Modified Rankin Score; N.: numerosity; NIHSS: National Institutes of Health Stroke Scale; OCS: Oxford Cognitive Screen; SDC: Communication Disability Scale; SPPB: Short Physical Performance Battery; std: standard deviation; TCT: Trunk Control Test. ^aPairwise comparisons highlighted significant P values in pairs of: No aphasia - Severe aphasia and Severe aphasia - Mute or global aphasia; ^bpairwise comparisons highlighted significant P values in all pairs involving Mute or global aphasia. *MoCA adjusted score on normative values by Santangelo et al.⁴¹

In the multivariable analysis, age, mBI, the comorbidity index of the Cumulative Illness Rating Scale (CIRS), the adjusted MoCA score, the NIHSS, and the Fugl-Meyer Assessment (FMA) resulted being significantly associated with the outcome (adjusted R²=77.2%). Specifically, younger age, higher independence in basic activities of daily living, reduced number of comorbidities, higher cognitive abilities, reduced stroke severity, and higher motor functions at admission, resulted to be independently associated with a more favourable functional outcome at discharge (Table IV).

Table IV. —Multivariate analyses. The statistically significant P values are reported in bold.

Independent variables	Unstandardized coefficients		P value	95% Confidence Interval
	B	Standard Error		Lower-limit	Upper-limit
Constant	45.268	23.117	0.052	-0.322	90.857
Age	-0.412	0.101	<0.001	-0.612	-0.212
Gender	-2.782	2.320	0.232	-7.357	1.792
Cohabitation	4.486	2.631	0.090	-0.704	9.675
Time from the event	-0.012	0.180	0.949	-0.367	0.344
Reduced vigilance and coma	-6.882	5.064	0.176	-16.870	3.106
Clinical instability	1.523	4.172	0.715	-6.704	9.750
Delirium	-3.571	5.186	0.492	-13.798	6.656
Acute infection	4.587	3.839	0.234	-2.984	12.159
Dysphagia	-4.145	2.491	0.098	-9.057	0.767
Malnutrition	-1.908	4.999	0.703	-11.767	7.950
Bedsores	6.765	3.705	0.069	-0.542	14.072
Bladder catheter	-1.646	2.790	0.556	-7.148	3.856
Incontinence	-1.865	2.650	0.483	-7.092	3.362
Pain	5.468	2.796	0.052	-0.046	10.981
CIRS_comorbidity	-1.621	0.701	0.022	-3.003	-0.239
NIHSS	-0.782	0.338	0.021	-1.448	-0.117
Aphasia (NIHSS_9)	1.494	1.570	0.342	-1.602	4.589
MoCA_adjusted_Santangelo*	0.414	0.180	0.022	0.060	0.768
mBI	0.305	0.069	<0.001	0.169	0.441
TCT	0.071	0.053	0.187	-0.035	0.176
mRS_anamnestic	-0.562	1.028	0.585	-2.590	1.465
MAS	0.116	0.353	0.742	-0.579	0.812
FMA	0.250	0.035	<0.001	0.182	0.319

*MoCA adjusted score on normative values by Santangelo et al.⁴¹

The mBI at discharge was correctly estimated with our model in 111 patients, while the expected outcome value was overestimated and underestimated respectively in 54 and 55 patients. The statistical comparisons of the variables among the 3 groups thus derived showed that only baseline and discharge mBI (P<0.001) and baseline FMA (P=0.049) were statistically different in incorrectly estimated outcome groups. Specifically, baseline mBI and FMA were lower in both the incorrectly estimated mBI groups (Table V).

Table V. —Descriptive analyses on the group of patients with overestimated, underestimated, and correctly estimated mBI values at discharge.

Variable	Mean (std)/Median [IQR] or Frequencies (%)			P valuea
	Correctly estimated (N.=111)	Overestimated (N.=54)	Underestimated (N.=55)
Age	81.0 [15.0]	80.5 [15.0]	80.0 [15.0]	0.782
Gender	Male: 53 (47.7%) Female: 58 (52.3%)	Male: 27 (50%) Female: 27 (50%)	Male: 28 (50.9%) Female: 27 (49.1%)	0.937
Cohabitation	No: 32 (28.8%) Yes: 79 (71.2%)	No: 12 (22.2%) Yes: 42 (77.8%)	No: 11 (20.0%) Yes: 44 (80.0%)	0.403
Time from the event	11.0 [9.0]	12.0 [9.00]	11.0 [8.00]	0.903
Reduced vigilance and coma	No: 106 (95.5%) Yes: 5 (4.5%)	No: 49 (90.7%) Yes: 5 (9.3%)	No: 51 (92.7%) Yes: 4 (7.3%)	0.412
Clinical instability	No: 98 (88.3%) Yes: 13 (11.7%)	No: 49 (90.7%) Yes: 5 (9.3%)	No: 52 (94.5%) Yes: 3 (5.5%)	0.475
Delirium	No: 106 (95.5%) Yes: 5 (4.5%)	No: 50 (92.5%) Yes: 4 (7.4%)	No: 52 (94.5%) Yes: 3 (5.5%)	0.687
Acute infection	No: 97 (87.4%) Yes: 14 (12.6%)	No: 48 (88.9%) Yes: 6 (11.1%)	No: 49 (89.1%) Yes: 6 (10.9%)	1.000
Dysphagia	No: 55 (49.5%) Yes: 56 (50.5%)	No: 23 (42.6%) Yes: 31 (57.4%)	No: 24 (43.6%) Yes: 31 (56.4%)	0.555
Malnutrition	No: 105 (94.6%) Yes: 6 (5.4%)	No: 51 (94.4%) Yes: 3 (5.6%)	No: 51 (92.7%) Yes: 4 (7.3%)	0.930
Bedsores	No: 100 (90.1%) Yes: 11 (9.9%)	No: 47 (87.0%) Yes: 7 (13.0%)	No: 47 (85.5%) Yes: 8 (14.5%)	0.627
Bladder catheter	No: 70 (63.1%) Yes: 41 (36.9%)	No: 30 (55.6%) Yes: 24 (44.4%)	No: 30 (54.5%) Yes: 25 (45.5%)	0.492
Incontinence	No: 76 (68.5%) Yes: 35 (31.5%)	No: 29 (53.7%) Yes: 25 (46.3%)	No: 30 (54.5%) Yes: 25 (45.5%)	0.097
Pain	No: 88 (79.3%) Yes: 23 (20.7%)	No: 41 (75.9%) Yes: 13 (24.1%)	No: 44 (80.0%) Yes: 11 (20.0%)	0.886
CIRS_comorbidity	3.00 [2.00]	3.00 [2.00]	3.00 [2.00]	0.857
NIHSS	6.00 [9.00]	8.00 [10.0]	7.00 [7.00]	0.184
Aphasia (NIHSS_9)	No aphasia: 70 (63.1%) Mild to moderate aphasia: 14 (12.6%) Severe aphasia: 18 (16.2%) Mute or global aphasia: 9 (8.1%)	No aphasia: 35 (64.8%) Mild to moderate aphasia: 9 (16.7%) Severe aphasia: 8 (14.8%) Mute or global aphasia: 2 (3.7%)	No aphasia: 34 (61.8%) Mild to moderate aphasia: 11 (20.0%) Severe aphasia: 7 (12.7%) Mute or global aphasia: 3 (5.5%)	0.848
MoCA_adjusted_Santangelo***	16.0 [21.0]	16.0 [13.0]	16.0 [12.0]	0.873
mBI	35.0 [53.0]	12.5 [39.0]**	20.0 [33.0]	<0.001
TCT	61.0 [63.0]	48.0 [62.0]	48.0 [62.0]	0.092
mRS_anamnestic	0.00 [1.00]	0.00 [1.00]	0.00 [1.00]	0.481
MAS	0.00 [2.00]	0.00 [3.00]	0.00 [3.00]	0.943
FMA	163 [87.0]	135 [83.0]	149 [74.0]	0.049
mBI dicharge	72.0 [47.0]	32.50 [44.0]**	81.0 [32.0]	<0.001

^aAll P values are connected to multiple comparisons among the three groups. The asterisks on the group of overestimated or underestimated patients are referred to the pairwise comparison of the respective groups with the correctly estimated group. Pairwise comparisons are not provided for P values not statistically significant. **P value <0.001; *P value <0.05. ***MoCA adjusted score on normative values by Santangelo et al.⁴¹

Discussion

The epochal improvements reported in acute phase treatments of stroke are not matched by reports of the same progress in the subacute phase. In addition to the innovation gap between acute and post-acute interventions, this may also depend on the poor quality of studies and inconsistency of treatments and evaluations, rather than a lack of effective rehabilitation strategies.^31-33 Our aim was to design and carry out a high-quality prospective study that would address multiple aspects. First, the evidence-based stroke ICP adopted had been consistently implemented and shared among all the participating centres,¹⁰ including the standardised assessment measures and the validated treatment protocols. This would reduce the variability regarding assessments and treatments often met in multicentric observational studies involving rehabilitation. Second, this study was designed and carried out to meet the requirements of the criteria of the Quality In Prognosis Studies (QUIPS) tool,⁴² for the risk of bias in studies of prognostic factors. Finally, we designed the study protocol and reported our results to comply with the STrengthening the Reporting of OBservational studies in Epidemiology (STROBE) Guidelines.⁴³

The results of our study identified a final model that explains the variance of the discharge mBI of the study participants with an adjusted R² of 77.2%. This model provides a reliable estimate of the functional outcome of stroke patients at discharge from the four IRUs, outreaching studies available in the current literature.^{44, 45}

To adequately describe the impact of rehabilitation on the outcome of stroke patients, a wide range of variables reflecting the complexity of this condition have been considered. This is essential to understand patients’ needs and promote the recovery of function and the reintegration of these patients in their pre-morbid setting, whenever possible.^{17, 46} In our analysis, among all the multidimensional features collected at admission, younger age, lower number of comorbidities, better cognitive function, reduced neurological deficits, and higher motor abilities were independently associated with a more favourable outcome at discharge.

Comparing our results to systematic reviews of the literature, Hakkennes et al.⁴⁴ confirmed the association with outcome of stroke severity, age, cognition, and functional level before rehabilitation as shown in our work, while we did not find an independent predictive role of continence in our cohort. The very old age of our population may explain why this feature was highly common and did not show a significant relationship with the outcome. However, a more recent review by Meyer et al.,⁴⁵ also did not identify urinary continence as a significant predictor of stroke discharge functional outcome. These authors reviewed multivariable predictive models of functional outcomes after post-stroke inpatient rehabilitation. Among a large number of models described, they identified 16 variables that were included in the final models in at least five studies. When considering papers investigating either by BI or by Functional Independence Measure (FIM) as the primary outcome, only 7 of the 16 variables resulted as significant predictors in more than 50% of the studies: age, admission functional level (BI or FIM), stroke severity (NIHSS), dysphasia, impulsivity, previous stroke, and neglect. Age and admission functional level (as measured with mBI), NIHSS and cognition, were significant also in our analyses, whereas previous history of stroke and dysphasia were not. This depend on their relative severity or on how these features were evaluated. For example, a previous ischemic event treated early and efficiently with recent therapies, as well as a mild production aphasia, may have a negligible impact on mBI at discharge.⁴⁷ In our study, any previous stroke was reported in anamnesis, regardless of the neurological sequelae, and the presence of language impairment was attested with the NIHSS sub-scale, rather than by a stroke-specific language assessment battery. This may have hidden the prognostic power of these variables and could be a starting point for future studies to implement a more in-depth assessment of previous strokes and a better characterization of the speech disorder. Among the other variables that Meyer et al.⁴⁵ identified, impulsivity was not assessed in our analysis, whereas neglect was not included because of exceeding missing values. This must be acknowledged as a limitation of the present study.

The need for a comprehensive and multidisciplinary patient evaluation is crucial to describe the result of the patient’s rehabilitation process. This concept was also highlighted by Weng et al.,⁴⁸ who retrospectively enrolled stroke patients included in a post-acute care program. The authors reinforced the usefulness of a systematic multidomain evaluation before rehabilitation to predict the outcome, showing that a higher baseline and greater improvement of cognitive and physical abilities were associated with shorter lengths of stay in the IRU, fewer hospital readmissions, and reduced 1-year mortality. Indeed, promoting a shared and standardized assessment protocol is the first step concerning the integration of evidence-based treatments in rehabilitation, and a strength of the present study. Standardisation can foster these processes by improving clinical research through increased data quality, supporting better data integration and reusability, enabling data exchange with partners, increasing the use of software tools, improving team communication, and facilitating regulatory reviews and audits.⁴⁶

Wei-Chieh Chen et al.⁴⁹ focused on the identification of clinical factors able to predict functional independence at hospital discharge, in stroke patients who received in-hospital rehabilitation. Similarly to our findings, they showed that baseline daily activity function (measured with mBI) and motor function impairment of the hemiplegic lower limb were the most important prognostic factors of functional independence. However, the generalisability of these results is limited due to the retrospective nature of the study and the different organization of the post-stroke care pathway in Taiwan.

Another retrospective study⁵⁰ exploited a multiparametric impairment assessment in stroke patients before and after inpatient rehabilitation. Plotting on a scattergram for each test the percentage of the highest score on admission and of patients whose score improved, Yagihashi et al.⁵⁰ defined three patterns of change in impairment during hospitalisation without however predicting the result of the rehabilitative treatment. Understanding which deficits can improve could allow for better decisions on when and how much to insist on the rehabilitation of impairment, rather than to provide a proper discharge functional outcome prediction, needed to plan the patient’s pathway including time of discharge and return to home.

Another aspect essential for a reliable and interpretable prognostic analysis of the stroke population is the choice of standardised, reliable and reproducible assessment tools, especially for what concerns the outcome measures. Many retrospective studies aimed to identify outcome predictors in stroke patients available before undergoing an intensive rehabilitation cycle. For instance, Bertolin et al.⁵¹ retrospectively examined acute predictors of cognitive and functional outcomes six months after stroke. The authors showed that demographic factors and cognitive, physical, and functional measures at stroke onset poorly explained cognitive sequels. However, the same variables only accounted for moderate degrees of variance in functional outcomes, measured with the basic activities of daily living (ADL)/Instrumental ADL (IADL) (ΔR²=4.2%) and Participation subscales of the Stroke Impact Scale (SIS) (ΔR²=2.3%). In their paper, brief screening instruments such as NIHSS and a short screener for deficits in orientation, registration, and attention⁵² (the Short Blessed Test) were the most consistent individual predictors. On the other hand, these authors’ choice of a self-report instrument like the SIS Score, to measure the independence in self-care and daily activity, in household chores and small purchases and the limitation in work and recreation, can be controversial. In fact, self-reports are subjected to a risk of report bias, mainly linked to invalid answers, due to lack of sincerity and introspective ability,⁵³ while the mBI assessed by a specialist, as we chose in our study, seems to be less prone to such bias. In fact, the choice of the primary outcome to adequately reflect the patient’s level of residual disability is also pivotal when studying rehabilitation functional outcomes. Many authors agree that it would be advisable to avoid measures focused on a narrow aspect of the function, such as binary outcomes or those describing single activities (e.g. ambulation) or measures that are possibly more related to contextual factors rather than to the rehabilitation intervention, such as whether the patient is discharged back home.⁵⁴ Many of the existing predictive models in stroke are aimed to obtain an automatic prediction of clinical binary outcomes, such as independence of walking,⁵⁵ or the likelihood of achieving at least one specific clinical score for independence⁵⁶ or upper limb functionality.⁵⁷ The results of modelling strategies using binary outcomes should be interpreted with caution, as pointed out by Dijkland et al.⁵⁴ For instance, Scrutinio et al.⁵⁶ developed an accurate prediction tool (area under the curve, AUC=0.866), externally validated by García-Rudolph et al.⁵⁸ (AUC=0.873), which incorporates age, sex, time from stroke onset to inpatient rehabilitation admission, baseline motor and cognitive functional independency scores, and neglect. However, their predictive model aimed to predict the probability of achieving an independence level requiring no more than supervision and a motor FIM score higher than 61 points. The use of a dichotomised measure may be a viable solution while targeting a very specific outcome, but its clinical application in support of the clinical decision for a comprehensive evaluation of the patient’s status may not be optimal. Indeed, binary outcomes often are of little relevance for the patient⁵⁹ and the family, whilst the continuous mBI^{60, 61} accurately reflects the level of independence and the burden that the caregiver will have to face at discharge. Thus, further development and validation of models with continuous outcomes should be promoted. This will allow a more granular prediction of the patient’s prognosis, with the possibility to evaluate alternative strategies to prevent the high risk of an unfavourable rehabilitative outcome, and also provide ground to improve the discharge planning with patients and families, already in the first days of rehabilitation hospitalisation. To avoid these limitations, we chose to preserve the continuous nature of our outcome variable (mBI at discharge). With this approach, our model obtains a good explanation of the outcome variance, with the possibility of distinguishing different outcomes at the end of hospitalisation such as complete recovery, sub-optimal or sufficient to continue the rehabilitation at home/assisted health service, which has a strong clinical impact.

Our descriptive analyses on the groups of patients with overestimated, underestimated, and correctly estimated mBI values at discharge showed that patients who were incorrectly predicted had significantly lower baseline mBI and FMA. This could mean that patients with a lower functional level at admission to IRU have less chance of being accurately estimated in their probability of reaching rehabilitation targets within their stay. It is possible that, while MBI and FMA provide a reliable admission assessment for some of these patients, others performed less well upon admission for reasons not strictly related to stroke severity, such as concurring acute clinical confounders, distress upon transfer, promoting agitation and lack of collaboration.⁶² These underlying clinical factors may either improve and resolve during the stay, explaining underestimation, or develop in full-blown clinical complications, explaining overestimation upon admission, but this hypothesis should be verified by further research. Possibly, including other assessment measures at admission, or repeating the same measures (MBI and FMA) at different times, particularly in those with lower scores upon admission, as well as integrating predicting models in time with updated clinical information of the patients may help improve the overall prediction.

Limitations of the study

This study has several limitations. Firstly, the model tested in this analysis needs external validation, which we plan to provide using datasets from ongoing prospective studies; further, we plan to compare and integrate it with machine learning-based approaches, for the analysis of the non-linear behaviour of data. In the meantime, a simple calculator, based on the coefficients obtained in the multivariable regression model, is available in the supplementary material, offering the mBI estimation at discharge based on these independent predictors, for stroke patients addressing post-acute intensive inpatient rehabilitation. The abovementioned lack of exploring impulsivity and neglect are also among our study limitations.⁶³ Finally, the number of patients, in reference to the multifactorial complexity of the baseline evaluation, may have underestimated the predictive role of some variables. On the other hand, the strengths of our work lie in the thorough description and reproducibility of the evidence-based rehabilitation protocol, and in the systematic assessment of all stroke patients addressing the IRUs in the considered time frame, with 85% of them enrolled and with 94% of them included in this prognostic analysis. Indeed, our study fulfils the criteria of the QUIPS tool,⁴¹ which assesses the risk of bias in studies of prognostic factors, for all the six considered areas of participation, attrition, prognostic factor measurement, confounding measurement and account, outcome measurement, and analysis and reporting. In fact, we prospectively assessed and proposed the study to all patients addressing the participating centres for post-acute intensive inpatient rehabilitation after a stroke, thus reducing selection bias. Thus, we provided a systematic enrolment of a cohort of patients actually representing the post-acute stroke patients addressing intensive inpatient rehabilitation. Our final model includes a set of reliable, easily collected variables at admission to intensive inpatient rehabilitation, predicting stroke patients’ discharge functional outcomes. These results can be considered the first step in understanding potential predictors of the mBI, while future research should involve the external validation of this model and the validation of algorithms for the development of a data-driven tool. While promoting a rationalisation of economic community resources (avoiding expenses for futile treatments), refining the prediction of stroke rehabilitation outcomes has great potential to facilitate an improvement in individual rehabilitation planning (personalised approach) and to optimie rehabilitation outcomes.

Conclusions

Among multidimensional potential predictors of stroke rehabilitation outcome, we found that younger age, as well as greater independence in basic activities of daily living, fewer comorbidities, higher cognitive abilities, reduced stroke severity, and higher motor function, recorded at admission, were independently associated with the ability to perform activities of daily living at discharge from intensive inpatient rehabilitation after stroke. The final model including these variables could help clinicians to provide a granular prediction of the functional outcome of post-acute stroke patients at their admission to inpatient rehabilitation. Additional investigations, for those patients who present very low mBI and FMA, and the further development of this model in larger settings are needed to further improve its performance.

Supplementary Digital Material 1

Supplementary Table I

Supplementary material, evaluation tools references.

Supplementary Digital Material 2

Supplementary Text File 1

Calculator