Intended for healthcare professionals


Collaborative systematic review of the randomised trials of organised inpatient (stroke unit) care after stroke

BMJ 1997; 314 doi: (Published 19 April 1997) Cite this as: BMJ 1997;314:1151
  1. Stroke Unit Trialists' Collaborationa
  1. a Stroke Unit Trialists' Collaboration
  1. Correspondence to: Dr Peter Langhorne, Academic Section of Geriatric Medicine, Royal Infirmary, Glasgow G4 0SF
  • Accepted 24 January 1997


Objectives: To define the characteristics and determine the effectiveness of organised inpatient (stroke unit) care compared with conventional care in reducing death, dependency, and the requirement for long term institutional care after stroke.

Design: Systematic review of all randomised trials which compared organised inpatient stroke care with the contemporary conventional care. Specialist stroke unit interventions were defined as either a ward or team exclusively managing stroke (dedicated stroke unit) or a ward or team specialising in the management of disabling illnesses, which include stroke (mixed assessment/rehabilitation unit). Conventional care was usually provided in a general medical ward.

Setting: 19 trials (of which three had two treatment arms). 12 trials randomised a total of 2060 patients to a dedicated stroke unit or a general medical ward, six trials (647 patients) compared a mixed assessment/rehabilitation unit with a general medical ward, and four trials (542 patients) compared a dedicated stroke unit with a mixed assessment/rehabilitation unit.

Main outcome measures: Death, institutionalisation, and dependency.

Results: Organised inpatient (stroke unit) care, when compared with conventional care, was best characterised by coordinated multidisciplinary rehabilitation, programmes of education and training in stroke, and specialisation of medical and nursing staff. The stroke unit care was usually housed in a geographically discrete ward. Stroke unit care was associated with a long term (median one year follow up) reduction of death (odds ratio 0.83, 95% confidence interval 0.69 to 0.98; P<0.05) and of the combined poor outcomes of death or dependency (0.69, 0.59 to 0.82; P <0.0001) and death or institutionalisation (0.75, 0.65 to 0.87; P<0.0001). Beneficial effects were independent of patients' age, sex, or stroke severity and of variations in stroke unit organisation. Length of stay in a hospital or institution was reduced by 8% (95% confidence interval 3% to 13%) compared with conventional care but there was considerable heterogeneity of results.

Conclusions: Organised stroke unit care resulted in long term reductions in death, dependency, and the need for institutional care. The observed benefits were not restricted to any particular subgroup of patients or model of stroke unit care. No systematic increase in the use of resources (in terms of length of stay) was apparent.

Key messages

  • Previous systematic reviews of organised inpatient (stroke unit) care have been limited by problems of interpretation and characterising stroke unit care

  • The important characteristics of stroke unit care within the randomised trials were the provision of coordinated multidisciplinary rehabilitation, staff specialisation in stroke or rehabilitation, and improved education and training

  • Patients managed in a stroke unit were more likely to survive, regain independence, and return home than those receiving conventional care

  • Apparent benefits were not restricted to any subgroup of stroke patients or model of stroke unit care

  • No systematic increase in length of stay was observed


The role of organised (stroke unit) care in managing inpatients with stroke has been controversial for over 30 years.1 The controversy arises because the moderate benefits that might be anticipated with stroke unit care can be reliably detected (or refuted) only with a very large randomised trial or a proper overview of the available small randomised trials. Evaluation of stroke unit care raises particular problems because of the complex and heterogeneous nature of the intervention and its potential interaction with other aspects of care. Even a prospective multicentre randomised trial could not guarantee a uniform intervention because the service characteristics would inevitably vary between centres.

Systematic review (including meta-analysis) methods combine the available evidence from randomised trials to draw more reliable and generalisable conclusions.2 Our review of randomised trials available up to October 1993 indicated that specialist stroke unit care may reduce death and institutionalisation after stroke.3 However, we did not have detailed descriptions of service organisation or detailed information on many outcomes or subgroups of interest and substantial new information has now become available from several recently completed randomised trials.

We conducted a further systematic review to determine whether the apparent benefits of organised stroke unit care were confirmed in a more extensive and updated analysis, examine outcomes in addition to death and institutionalisation, examine the effects in subgroups of stroke patients and with various models of specialist stroke unit care, provide a detailed description of stroke unit and control interventions, and identify the features associated with an improved outcome. To meet these objectives a collaborative review group was formed which included trialists from the available randomised controlled trials.


We aimed to compare any system of organised inpatient stroke care with the less organised conventional practice. We therefore included all prospective trials that used some form of randomisation to allocate patients to an organised stroke unit or conventional care, usually in general medical wards. Trials were included if treatment allocation was carried out on a strictly random basis or a quasi-randomised procedure (such as date of admission). We excluded studies which compared specific therapies within an organised stroke care setting.

Our objectives were to examine the effect of organised stroke unit care on the outcomes of death, dependency, and the requirement for institutional care (all recorded at the end of scheduled follow up in an intention to treat analysis).

Identification of trials

We identified relevant research reports up to December 1995 using several approaches.3 In summary, we carried out systematic hand searches of 22 core neurology and stroke journals and five Japanese journals and systematic searches of Index Medicus, Medline, and dissertation abstracts. We searched the reference lists of trials, review articles, and textbooks; Current Contents; and the proceedings of 43 recent conferences on neurology, geriatric medicine, and rehabilitation. Further information was obtained by talking to colleagues and publicising our preliminary findings at stroke conferences in the United Kingdom, Scandinavia, Germany, Switzerland, Spain, Canada, and South America.

Definition of interventions

Although the primary question was whether organised inpatient stroke care could improve outcomes compared with contemporary conventional care, we divided the organisation of service into one of the following three predefined groups to reflect the heterogeneity of services.

Dedicated stroke unit–A service provided by a discrete stroke ward or stroke team working exclusively in the care of stroke patients. This category included acute (intensive) stroke units, which accept patients acutely but discharge early (usually within seven days); rehabilitation stroke units, which accept patients after a minimum delay of seven days and focus on rehabilitation; and combined acute/rehabilitation units, which accept patients acutely but also provides rehabilitation for at least several weeks. Both the rehabilitation unit and combined acute/rehabilitation unit models offered prolonged periods of rehabilitation.

Mixed assessment/rehabilitation unit–A ward or team which has an interest and expertise in the assessment and rehabilitation of disabling illness but does not exclusively manage stroke patients.

General medical wards–A service provided in wards which focus on the management of acutely ill general medical patients but not on their subsequent rehabilitation. In most trials this formed the control group.

Definition of outcome measures

The primary analyses examined death, dependency, and the requirement for institutional care at the end of scheduled follow up. Dependency was categorised into two groups where independent was taken to mean that an individual did not require physical assistance for transfers, mobility, dressing, feeding, or toileting. Individuals who failed any of these criteria were considered dependent. The criteria for independence were roughly equivalent to a Rankin score of 0-2 or a Barthel score of >18/20.4 The requirement for long term institutional care was taken as meaning care in a residential home, nursing home, or hospital at the end of the rehabilitation period. Length of stay in a hospital or institution was also recorded.

Data from contributing trials

The principal investigators of all the trials that fulfilled the criteria of the overview were invited to join the Stroke Unit Trialists' Collaboration. All who could be contacted agreed to join. They were asked to provide details of their trial design, including the method of treatment allocation, selection criteria, characteristics of patients, details of service organisation, duration of interventions, duration of follow up, numbers in each outcome group, and additional services after discharge from hospital. The survey of trial characteristics included a structured interview with the trial coordinator, carried out by a single interviewer (PL), which focused on aspects of the structure, staffing, organisation, selection criteria, and procedures and practices within the stroke unit and control settings. For the three trials for which a coordinator could not be contacted we have used the best available published information.

Wherever possible we obtained basic outcome data at the end of scheduled follow up for all patients randomised (to permit an intention to treat analysis). Most trials could be analysed on this basis, at least for the outcomes of death and death or requiring institutional care. Those trials with incomplete follow up were analysed with the assumption that patients lost to follow up were alive and living at home. The implications of these assumptions were explored in a sensitivity analysis.

Outcome information was also sought for subgroups of patients based on age, sex, and stroke severity. Severity of stroke at the time of randomisation was defined by patients' initial dependency (within the first week after stroke). Where randomisation was carried out at different times after stroke, initial dependency was inferred from published information on the expected rate of functional recovery5:

Mild stroke–Patient can transfer and walk (with or without assistance) during the first week after the stroke. This is roughly equivalent to a Barthel score of >10/20 (Rankin score 0-3) within one week of the stroke or >13/20 (0-3) by two weeks after the stroke.

Moderate stroke–Patient is conscious and has sitting balance but is unable to stand or walk during the first week after stroke.

Severe stroke–Patient has reduced consciousness or no sitting balance, or both, during the first week after stroke; equivalent to a Barthel score of <3/20 (Rankin 5) within one week or <4/20 (5) by two weeks.

Although the inaccuracies inherent in this process are likely to have resulted in some misclassification of patients, the criteria were applied equally to stroke unit and control patients. Most trials used exclusion criteria such that patients with the mildest and severest strokes would be excluded.

We sought individual patient data for all trials, but unfortunately insufficient data were available to permit a comprehensive individual patient data analysis. The available data were, however, used to cross check the results obtained as summary data.

Statistical methods
Dichotomous outcomes

The formal statistical methods used to combine the results from different trials have been described elsewhere.6 Within each trial the standard quantity “observed minus expected” (together with its variance) was calculated for the numbers of events among patients allocated to treatment groups. The grand totals of the individual observed minus expected values and of their variance were used to calculate P values and odds ratios.6 The odds ratio gives the odds of an unfavourable outcome among patients in the treated groups compared with control patients stratified by trial.

This approach (fixed effects model) assumes that each trial result is sufficiently similar to the other results to differ only by chance. Where this is not the case, statistical heterogeneity exists. When heterogeneity was found the sources were explored7 and results confirmed by a random effects model analysis.8

Continuous variables

Continuous variable data (length of stay) were analysed as the weighted mean difference–that is, the difference between mean values in the treatment and control groups of individual trials and the mean difference weighted for trial size for groups of trials.9 The 95% confidence intervals of the weighted mean difference were calculated by using the mean and standard deviation data from the individual trials. Because the length of stay was calculated in different ways for different trials, results were calculated from both absolute values (days) and relative change in length of stay (expressed as a percentage of the length of stay in the control group).

Absolute outcomes

Absolute outcome rates, expressed as the proportion of patients in each outcome group are less statistically robust but more clinically meaningful than relative changes in outcomes. The number needed to treat to prevent one adverse outcome was calculated as the reciprocal of the difference in absolute outcome rates between the treatment and control groups.


A total of 19 trials were identified by December 1995.10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 Seventeen were formally randomised by using random numbers or sequentially numbered sealed envelopes and two used informal procedures based on bed availability27 or a strict admission rota.28 These two trials were evaluated separately to exclude significant bias in the conclusions.

Of the 19 trials identified, one has not yet been completed14 while the remaining 18 contained a total of 3249 patients. Eleven trials (2060 patients) compared a dedicated stroke unit with a general medical ward, six (647 patients) compared a mixed assessment/rehabilitation unit with a general medical ward, and four (542 patients) compared a dedicated stroke unit with a mixed assessment/rehabilitation unit (table 1). The total number of comparisons is greater than the number of trials because in three trials patients could be randomised to one of two conventional care groups; two of these trials21 22 used a stratified randomisation procedure and one11 did not.

Table 1

Characteristics of trials contributing data to the review

View this table:

Detailed descriptive information on service characteristics could not be obtained by structured interview for only three trials.10 16 19 In 18 trials stroke unit care included rehabilitation lasting several weeks if required; 10 of these units admitted patients acutely and eight after a delay of one to two weeks. Only one trial evaluated an acute stroke unit with no continuing rehabilitation.25

In 17 of the trials the organised care was provided in a geographically discrete ward; two trials examined peripatetic systems of care.18 28 Table 2) summarises the service comparisons within the trials. Stroke unit interventions were more likely to be reported to include coordinated multidisciplinary rehabilitation, staff with a specialist interest in stroke or rehabilitation, and regular programmes of education and training. Several factors indicating a more intensive or more comprehensive input of care were less significantly associated with stroke unit care.

Table 2

Frequency of various characteristics within organised (stroke unit) care and conventional care settings. Values are numbers (percentages) of arms of trials with available data

View this table:

Summary data on death, placement, and dependency at the end of scheduled follow up were available for 21, 20, and 20 comparisons respectively. In one trial the number of dependent patients had to be calculated from the mean and standard deviation Barthel score results.18 Six trials had minor omissions of data during follow up (total 10 stroke unit patients and 25 controls).10 11 12 18 23 27 As these patients were assumed to be alive and living at home, this may have introduced a minor basis in favour of the control group.

Within the stroke unit group 340/1626 (20.9%) patients were dead at the end of follow up (median one year after stroke), 304/1597 (19.0%) were in institutional care, and 519/1409 (36.8%) were dependent. The corresponding figures for controls were 413/1623 (25.4%) dead, 344/1600 (21.5%) in institutional care, and 543/1421 (38.2%)dependent. The minor variation in the denominator is due to placement and dependency data each being unavailable for one trial.

Death only

Figure 1) shows the odds of death by the end of scheduled follow up in different forms of stroke unit versus conventional care. The summary result (odds ratio 0.82, 95% confidence interval 0.69 to 0.98; P<0.05) was not complicated by significant heterogeneity between trials (Χ2=13.6, df=18; P>0.2). There was no detectable variation between the treatment effects in the three subgroup comparisons in figure 1). The odds of death was essentially unchanged if the analysis was restricted to trials where scheduled follow up was continued for a fixed period of six months or one year (0.84, 0.70 to 1.04; P<0.1).11 12 15 17 19 20 21 24 25 26 27 28 The exclusion of two trials with an informal randomisation procedure27 28 did not affect the conclusions (0.81, 0.67 to 0.98; P<0.05).

Fig 1
Fig 1

Odds of death occurring by end of scheduled follow up in stroke unit compared with conventional care. Odds ratios and 95% confidence intervals of individual trials are presented as a black box and horizontal line. The pooled odds ratio and 95% confidence interval for a group of trials is represented by an open diamond; the black diamond shows the pooled result for all trials. Data were not available for one trial14

Death or institutionalisation

The second outcome examined (fig 2) was the odds of death or requiring institutional care at the end of follow up (median one year after stroke). The summary result (0.75, 0.65 to 0.87; P<0.0001) was highly significant but some heterogeneity existed between trials (Χ2=25.9, df=19; P=0.1). Reanalysis of the results with a random effects model produced similar results (0.74, 0.62 to 0.89; P<0.0001). The observed heterogeneity was largely attributable to the five trials that had a short (less than six weeks) or variable period of follow up (Χ2=14, df=4; P<0.01).13 16 18 22 23 Trials with a fixed follow up period showed a significant reduction in death or institutionalisation (0.76, 0.64 to 0.90; P<0.01) with much less heterogeneity (Χ2=11.5, df=13; P>0.2). There was no significant variation between the treatment effects in the three subgroup comparisons. The estimate of apparent benefits was unaffected if informally randomised trials were excluded.

Fig 2
Fig 2

Odds ratio (95% confidence interval) of death or requiring institutional care at the end of scheduled follow up in patients receiving stroke unit compared with conventional care. Abbreviations and terms as for fig 1. Data were not available for two trials10 14

Death or dependency

The third outcome examined was the combined adverse outcome of being dead or dependent in activities of daily living at the end of follow up (fig 3). The overall odds ratio of being dead or dependent if given stroke unit care rather than conventional care was 0.71 (0.61 to 0.84; P<0.0001) but the summary result showed some heterogeneity (Χ2=16.1, df=19; P>0.2). Reanalysis with a random effects model produced similar results (0.72, 0.61 to 0.83; P<0.0001). The main source of heterogeneity seems to reflect the nature of the control group. Results were less heterogeneous (Χ2=10, df=12; P>0.2) and odds ratios remained significant (0.66, 0.55 to 0.79; P<0.0001) where either a dedicated stroke unit or a mixed assessment/rehabilitation unit was compared with a general medical ward. The conclusions were not altered by the exclusion of trials with a variable follow up period10 18 22 23 or informal randomisation procedure27 28 or where numbers of dependent patients were calculated from continuous data.18

Fig 3
Fig 3

Odds of death or dependency at the end of scheduled follow up with stroke unit compared with conventional care. Abbreviations and terms as for fig 1. Data were not available for two trials13 14

The main methodological difficulties with using dependency as an outcome was the degree of blinding of the final assessment and the potential for bias if the assessor was aware of the treatment allocation. Five trials used an unequivocally blinded final assessment for all patients.15 17 18 21 27 The odds ratio for death or dependency in that group was 0.72 (0.55 to 0.94; P<0.01).

Absolute outcome rates

The proportion of patients dead at the end of scheduled follow up was 340/1626 (20.9%) in the stroke unit group and 413/1623 (25.4%) in the controls. On this basis the number needed to treat to prevent one death is 22. Interpreting absolute outcome rates can be problematical if the baseline event rate is variable.29 As the baseline fatality rate varied from 0-50% in individual trials the number needed to treat might be expected to range from about 10 to infinity in the different study populations.

The proportion of patients who were unable to live at home at the end of follow up was 640/1597 (40.1%) in the stroke unit group and 755/1600 (47.2%) in the controls (number needed to treat of 14). The baseline rate in individual trials ranged from 21-81% thus the number needed to treat might range from 8 to 30.

In total 843/1409 (60.0%) stroke unit patients and 944/1421 (66.4%) control patients failed to regain independence (number needed to treat of 16). With baseline rates of death or dependency of 39-100%, the range in the number needed to treat would be about 10 to 25.

Length of stay

Mean or median length of stay was available for 18 trial comparisons. Length of stay was calculated in different ways (for example, acute hospital stay, total stay in hospital or institution). Mean length of stay ranged from 13-162 days in the stroke unit groups and 14-137 days in controls. Ten trials reported a shorter length of stay in the stroke unit group12 15 22 23 24 25 26 27 and eight a more prolonged stay.11 13 17 18 20 21 28 The calculation of weighted mean differences in length of stay was subject to methodological limitations. Five trials reported median rather than mean length of stay 13 14 15 16 17 18 19 20 21 22 23 24 25 and in six trials the standard deviation was inferred from the P value or the standard deviation results from similar trials.12 13 22 23 25 26 Overall, there was a relative reduction in length of stay in the stroke unit group of 8% (3-13%). When length of stay was calculated from absolute values (days) there was a non-significant reduction (−0.3, 95% confidence interval −1.8 to 1.1 days). Both the summary estimates were complicated by considerable heterogeneity which limits the extent to which general conclusions can be inferred.

Subgroup analysis

Figure 4) shows the subgroup analyses in terms of relative reduction of the combined adverse outcome of death or requiring long term institutional care. Details of important subgroups were available for most trials (at least 2000 patients randomised). There was no clear association of the patients' age, sex, or stroke severity with the effectiveness of organised stroke unit care. However, a relatively small number of events were observed, limiting the statistical power.

Fig 4
Fig 4

Analysis of patient and service characteristics on effectiveness of stroke unit care versus conventional care. Results are presented as odds ratio (95% confidence interval) of combined adverse outcome of death or requiring long term institutional care. Departmental setting refers to the medical department in which organised stroke unit care was established.

Figure 4) also outlines the relative reduction in adverse outcomes in a variety of service subgroups. Combined acute/rehabilitation stroke wards, stroke rehabilitation wards, and mixed acute/rehabilitation wards all tended to have better results than conventional care in general medical wards. There were insufficient data to comment on the acute stroke unit and roving stroke team evaluations. Benefits were apparent across units with different forms of admission policy, and within different departmental settings, and across all units which provided rehabilitation.

Publication bias

Publication bias (the selective non-reporting of trial results considered to be neutral or negative) is a potential problem for any systematic review.30 The degree to which the conclusions of the overview would be overturned by missing neutral trials can be estimated by calculating how many randomised patients (with a similar baseline event rate as in the overview) would have to be recruited from neutral trials (odds ratio=1.0) to render the overall result non-significant (P=0.05). These estimates for the mortality, combined death and institutionalisation, and combined death and dependency outcomes are >500, >4000, and >6000 respectively. We also examined the distribution of individual trial results in relation to the trial size in a funnel plot.30 No obvious deficiency of small, negative trials was observed.

Finally we examined the prospective sample of ongoing trials which were identified and recruited before any results were known.15 21 22 23 24 25 This included 1558 patients and the odds ratio for the combined outcome of death or institutionalisation was 0.73 (95% confidence interval 0.63 to 0.84; P<0.001).


Systematic reviews (including meta-analysis) provide a method for examining the results of randomised trials of interventions which may have modest but clinically important effects.2 There are several potential advantages in having a collaborative review approach, where representatives from each of the original trials are recruited into the study group. Firstly, the network of trialists recruited often have valuable information about unpublished or unfinished randomised trials, thus reducing the risk of publication bias. Secondly, the collaborative approach can allow standardised descriptions of intervention characteristics which would otherwise be reported in a manner which is not sufficiently detailed, standard, or consistent between trials. Thirdly, the collaborative review approach allows the collection of standardised subgroup and outcome information. Finally, interpretation of overviews of complex interventions can be problematical unless one can call on the collective experience and data of the trialists who are aware of the context and practical constraints within which the original randomised trials operated. Overviews based on a reanalysis of individual patient data provide the “gold standard” meta-analysis.31 We were not able to pursue this approach because these data were not available for a substantial number of trials. However, we have been able to provide standard data sets and provide much more information than could be obtained from published data alone.

Stroke unit characteristics

Our results indicate that the benefits of organised stroke unit care, as opposed to conventional care, are not clearly due to the structure, departmental setting, staff mix, or the amount of medical, nursing, and therapy input available. The most distinctive features seem to be those of organisation (coordinated multidisciplinary team care, nursing integration with multidisciplinary care, and involvement of carers in the rehabilitation process), specialisation (medical and nursing interest and expertise in stroke and rehabilitation), and education (education and training programmes for staff, patients, and carers). These characteristics were held in common within most stroke unit settings and were usually absent from the conventional care setting. The observation that stroke unit care was usually provided in a geographically discrete ward may reflect difficulties in developing coordinated care within a mobile stroke team.32

However, several methodological problems exist with this approach to analysing stroke unit services. Firstly, the information was obtained from the trialists who ran the stroke units and we were not able to obtain information from all staff who provided the conventional care. Therefore our findings could be biased by the expectations of the trialists as to which stroke unit features may or may not be effective. Secondly, this was largely a retrospective analysis and in some cases specific questions could not be answered by the trialist or were not explicitly stated in the original published reports. The information provided here may reflect a mixture of both the recollection of factual details and the recall of features which trialists believed were effective. At best, it represents a strictly factual account of service characteristics, while at the worst, it represents a consensus view from the stroke unit trialists as to which features of stroke unit care were important. Although the identification of characteristics which correlate with effective stroke care does not prove that these characteristics dictated that effectiveness, it does provide powerful circumstantial evidence.

Stroke unit outcomes

The primary question of this review was whether organising inpatient stroke care could improve patient outcomes compared with contemporary conventional care. Our results confirm and extend the findings of previous work3 33; compared with conventional care organised stroke unit care reduces the odds of death after stroke. This apparent effect, however, is not statistically robust and could be overturned by a relatively small number of unpublished randomised trials.

The observed reduction in the combined adverse outcomes is much more convincing. The reduction in death or the requirement for long term institutional care was statistically robust. While the requirement for long term care is a useful surrogate for disability that is not subject to assessor bias,34 the absolute rates of institutionalisation will be influenced by national and cultural factors. Our findings indicate that the reduction in the requirement for institutional care was not due to unreasonable hospital discharge policies because the benefits were sustained for up to one year. They also indicate that reduced institutionalisation was a result of fewer patients becoming dependent rather than more dependent patients being discharged home.

The observed reduction in the combined adverse outcome of death or dependency was also statistically robust. However, it is subject to potential observer bias where final assessments were not carried out in a blinded manner. The sensitivity analysis based on those trials which used an unequivocal blinded assessment suggest that such bias has not seriously influenced the results.

Subgroup analysis

The subgroup analysis indicates that the observed benefits of organised stroke unit care are not limited to any particular subgroup of patients or models of stroke unit organisation. The apparent benefits of stroke unit care were seen in both sexes, in patients aged under and over 75 years, and across a range of stroke severities. Combined acute and rehabilitation stroke units, rehabilitation stroke units, and mixed assessment/rehabilitation units all tended to be more effective than conventional care provided in a general medical ward setting. The limited amount of information from direct comparisons of dedicated stroke rehabilitation units with mixed assessment/rehabilitation unit was insufficient to provide conclusive results. Apparent benefits were seen in units with acute admission policies as well as those with delayed admission policies and in units operating within different departments.

Rational arguments can be made to support individual models of stroke unit care (for example, combined acute/rehabilitation units are likely to cater for a broader group of stroke patients, mixed assessment/rehabilitation units are more flexible in also offering a service to other patient groups). However, our analysis cannot indicate if one model of specialist stroke unit care is more effective than another. The aspects of care which were held in common by all stroke units concerned their provision of prolonged (up to several weeks) periods of rehabilitation and certain practices and procedures (such as the presence of a coordinated multidisciplinary team approach with specialist stroke interests of medical and nursing staff and programmes of ongoing training and education in stroke). All these aspects of stroke unit care are sufficiently fundamental to permit a flexible approach to improving services but are sufficiently specific to allow the audit of such stroke services.

Cost effectiveness

The results reported here indicate that relatively few stroke patients need to be managed in an organised stroke unit to prevent a death, dependency, or institutionalisation. Our calculations of the number of patients needed to treat to prevent one adverse outcome are very approximate. However, they do indicate the potential degree of benefit which might be achieved through improvements in the organisation of stroke patient care. This compares favourably with many routine medical interventions. However, at what cost would this be achieved?

There are insufficient reliable data available to permit a detailed cost effectiveness analysis of stroke unit care, although recent studies from Canada and Europe indicate that the main costs of inpatient stroke care are due to “hotel” and staffing costs.35 36 Therefore the length of stay in hospital may be a good surrogate measure of costs assuming that staffing levels are relatively constant. Our analysis of length of stay was complicated by varying definitions of inpatient stay with variable periods of follow up and different approaches to reporting results. However, the benefits of organised stroke unit care did not depend on an increased hospital stay and may even reduce it. It seems reasonable to conclude that organised (stroke unit) care is unlikely to be more expensive than conventional care in a general ward setting and may be less expensive.


Acute stroke patients should be offered early organised multidisciplinary care, ideally provided within a ward dedicated to stroke care, which can offer a substantial period of rehabilitation if required. Access should not be restricted by age, sex, or stroke severity. There are several approaches to providing this care but all stroke units should aim to replicate the main characteristics of those in the randomised trials.

Future trials should focus on examining the potentially important components of care and on direct comparisons of different models of organised stroke unit care. Preplanned collaboration between comparable trials could alleviate some of the problems of retrospective systematic review.37


List of collaborators (in alphabetical order): K Asplund (Umea, Sweden), P Berman (Nottingham), C Blomstrand (Gothenberg, Sweden), M Dennis (Edinburgh), J Douglas (Glasgow), T Erila (Tampere, Finland), M Garraway (Edinburgh), E Hamrin (Linkoping, Sweden), G Hankey (Perth, Australia), M Ilmavirta (Jyvaskyla, Finland), B Indredavik (Trondheim, Norway), L Kalra (Orpington), M Kaste (Helsinki, Finland), P Langhorne (Glasgow), H Rodgers (Newcastle), J Sivenius (Kuopio, Finland), J Slattery (Edinburgh), R Stevens (Dover), A Svensson (Gothenborg, Sweden), C Warlow (Glasgow), B Williams (Glasgow), S Wood-Dauphinee (Montreal, Canada). Important contributions were also made by D Deleo (Perth, Australia), A Drummond (Nottingham), R Fogelholm (Jyvaskyla, Finland), H Palomaki (Helsinki, Finland), T Strand (Umea, Sweden), and L Wilhelmsen (Gothenberg, Sweden).

We thank Carl Counsell and Hazel Fraser (Cochrane Collaboration Stroke Group) for invaluable assistance with literature searching and Patricia Kelly for secretarial support.

Funding: Chest, Heart and Stroke, Scotland.

Conflict of interest: None.


  1. 1.
  2. 2.
  3. 3.
  4. 4.
  5. 5.
  6. 6.
  7. 7.
  8. 8.
  9. 9.
  10. 10.
  11. 11.
  12. 12.
  13. 13.
  14. 14.
  15. 15.
  16. 16.
  17. 17.
  18. 18.
  19. 19.
  20. 20.
  21. 21.
  22. 22.
  23. 23.
  24. 24.
  25. 25.
  26. 26.
  27. 27.
  28. 28.
  29. 29.
  30. 30.
  31. 31.
  32. 32.
  33. 33.
  34. 34.
  35. 35.
  36. 36.
  37. 37.