Introduction
Managing unplanned admission in older people is a challenge for health care systems worldwide. In England alone there are more than 2 million unplanned admissions per year in older people, accounting for 68% of emergency bed days.Reference Imison, Poteliakhoff and Thompson1 Emergency admissions are costly, have adverse consequences for the individual and are detrimental to provision of planned care. As populations age, it is expected that unplanned admissions and bed use will rise further, placing increasing pressure on health care resources. In recent years, older people's emergency admissions have increased at approximately 6% per annum in England, a trend that has been described as unsustainable for the NHS.Reference White2–Reference Blunt, Bardsley and Dixon4 Internationally, policy and practice are increasingly focused on reducing unnecessary admissions in older people, but recent analyses suggest that unplanned admissions continue to be a major challenge.5–Reference Hawkes8 In 2006, Ham reviewed admission avoidance interventionsReference Ham9 and concluded that, although there was some evidence to support some approaches (such as risk stratification, self-management for chronic disease, integrated care, specialist teams), there was little high-quality evidence about many interventions (case management, telemedicine, care pathways) and that further research was required. This review will focus on new evidence in relation to older people's admissions in the period following Ham's review.
Search methods
Medline, PubMed, EMBASE and World of Science databases were searched for research relating to older people, avoidable admission, readmission, identification and targeting of admission risk and admission avoidance interventions. The review analyses current evidence about the unplanned admission and readmission of older people, focusing on the general medical or community-dwelling population rather than specific disease groups. There are numerous studies investigating readmissions in specific disease groups, but they tend to be small, of variable quality and context specific, making generalization to the wider population difficult. Search terms were: older people; admission; readmission; risk factors; prediction; avoidable admission; interventions. Truncation, Boolean operators and synonyms were used to combine these terms.
The search captured 481 papers, of which 425 were excluded because they focused on specific diagnoses rather than overall older population data, were not research studies, were not in English, were focused on surgical or elective admissions or on specific settings such as critical care and nursing homes, were protocol descriptions or were from prior to 2006. The remaining 56 papers, including systematic reviews and meta-analyses, were reviewed to explore recent evidence on risk factors for unplanned admission, the identification and targeting of individuals at risk and current evidence on effective interventions to reduce unplanned admissions.
Admission rates in older people
Admission rates reported in developed countries vary, although data collection is inconsistent and variation in age groupings and definitions of admission can make international comparisons difficult. In recent years, Inouye et al.Reference Inouye, Zhang, Jones, Shi, Cupples, Calderon and Marcantonio10 described a rate of 15% admission in their 12-month study of a large cohort of community-dwelling older people aged 70 and over in the USA. In contrast, both Lopez-Aguila et al.Reference Lopez-Aguila, Contel, Farre, Campuzano and Rajmil11 and Donnan et al.Reference Donnan, Dorward, Mutch and Morris12 report rates approximately 7% over 6 and 12 months, respectively, in Spain and Scotland, although the latter looked at the population aged 40 and over. Walsh et al.Reference Walsh, Roberts, Nicholls and Lattimer13 report rates of 5%, but only looked at specific, potentially avoidable admissions in a large population study in England. Reported rates of readmission in older people are reasonably consistent across patient groups and countries. Zerdevas & DobsonReference Zerdevas and Dobson14 and Laniece et al.Reference Laniece, Couturier, Drame, Gavazzi, Lehman, Jolly, Voisin, Lang, Jovenin, Gauvain, Novella, Sain-Jean and Blanchard15 reported a readmission rate of approximately 14% in people aged 75 and over in England and France. In a large population cohort study in Spain, Lopez-Aguila et al.Reference Lopez-Aguila, Contel, Farre, Campuzano and Rajmil11 found a readmission rate of 17% in 6 months. In the USA, Hastings et al.Reference Hastings, Oddone, Fillenbaum, Sloane and Schmader16 reported 19% readmission rates in older people discharged from the emergency department (ED), using analysis of routine data. Walsh et al.Reference Walsh, Addington-Hall, Roberts, Nicholls and Corner17 reported a rate of 10% in medical inpatients in England, and Moons et al.Reference Moons, De Ridder, Geyskens, Sabbe, Braes, Flamaing and Milisen18 a rate of 16% in older people discharged from the ED in Belgium, both over a period of 30 days. It should also be noted that within countries there is substantial variation in admission rates. In England, a 4-fold variation between different health organization areas has been reportedReference Imison, Poteliakhoff and Thompson1–Reference Robinson3. Imison et al.Reference Imison, Poteliakhoff and Thompson1 argue that, if all areas achieved rates of admission similar to those with the lowest bed use, around 7000 fewer beds would be needed in England. Further research is needed to understand the variation in admission rates within and between countries, but this will require large-scale longitudinal studies and a consistent approach to measurement.
Risk factors for unplanned admission
Data on factors influencing population admission rates are limited, but suggest that variation in admission may be influenced by factors that are not amenable to control; age is the strongest driver of admissions and different geographical areas may have substantially different age distributions in their populations. Urban locations also have higher admission rates.Reference Weeks, Lee, Wallace, West and Bagian19 Social deprivation is a risk factor and explains 45% of variation between GP practices.Reference Purdy7 Number and severity of chronic diseases are also predictors of admission within the general population.Reference Saver, Wang, Dobie, Green and Baldwin20 Relatively few national or population-level studies have been carried out to explore this question and data often emerges from other studies, such as those intended to validate screening or risk stratification tools. These studies have identified increased co-morbidity, prior hospitalization, higher levels of primary care visits, higher age, living alone, cognitive impairment, residence in nursing home, heart failure and social deprivation as key risk factors for unplanned admission.Reference Inouye, Zhang, Jones, Shi, Cupples, Calderon and Marcantonio10–Reference Donnan, Dorward, Mutch and Morris12, Reference Ben-Chetrit, Chen-Shuali, Zimran, Munter and Nesher21, Reference Crane, Tung, Hanson, Cha, Chaudhry and Takahashi22 However, it is unclear if these studies are sufficiently powered to answer broader questions about population risk factors. Studies of population risk stratification tools have the potential to illuminate these issues, and current evidence from these suggests that age, deprivation, cognitive impairment, social isolation and mental health problems are predictive of risk.Reference Gage, Ting, Williams, Drennan, Goodman, Iliffe, Manthorpe, Davies and Masey23, Reference Sargent, Boaden and Roland24
Recent retrospective and longitudinal studies have explored the factors related to increased risk of readmission in older people. Hastings et al.Reference Hastings, Oddone, Fillenbaum, Sloane and Schmader16 explored risk factors for readmission following discharge from the ED in older Medicare patients. They found that increased age, greater co-morbidity, higher levels of deprivation and prior recent ED or hospital admission all increased risk. Longman et al.Reference Longman, Rolfe, Passey, Heathcote, Ewald, Dunn, Barclay and Morgan25 focused specifically on those with multiple chronic conditions and found that those with repeated admissions were more likely to have heart failure, higher social isolation scores and higher levels of co-morbidity. There have been some attempts to explore the impact of frailty and functional problems and these suggest that symptom burden rather than diagnoses may be important.Reference Salanitro, Hovater, Hearld, Roth, Sawyer, Locher, Bodner, Brown, Allman and Ritchie26–Reference Walsh, Roberts and Nicholls28 Laniece et al. Reference Laniece, Couturier, Drame, Gavazzi, Lehman, Jolly, Voisin, Lang, Jovenin, Gauvain, Novella, Sain-Jean and Blanchard15 found that socio-demographic factors, co-morbidity and cognitive impairment were not associated with readmission, but functional problems and markers of frailty were reported as being predictive of readmission. However, this study did not use a recognized frailty indicator. The variables that were significantly associated with risk (poor body condition, pressure sores) could equally have indicated poor quality hospital care rather than frailty or geriatric syndromes. In contrast, the study of Drubbel et al.Reference Drubbel, de Wit, Bleijenberg, Eijkemans, Schuurmans and Numans29 of the predictive value of a frailty measure used a recognized approach (accumulation of deficits) to calculating a frailty index (FI). The FI tertiles discriminated low, medium and high risk for adverse outcomes, which included ED visits, out-of-hours GP visits and nursing home admission. However, this study did not explicitly address unplanned admission. Similarly, Walsh et al.Reference Walsh, Addington-Hall, Roberts, Nicholls and Corner17 reported that, when age, sex and co-morbidity are adjusted for, admissions for problems associated with frailty are predictive of potentially avoidable readmissions.
From these studies it is apparent that a range of risk factors might be important in admission and readmission in older people. High levels of co-morbidity (or the presence of specific high-risk conditions such as heart failure), predict admission risk in many, but not all, studies. However, the predictive value of co-morbidity is modest even where it is identified as an independent risk factor. Social deprivation and isolation are consistently important in studies where they are included as variables, but this is relatively unusual. There is also more questionable benefit in identifying risk factors that are not modifiable. Function, disability and frailty may be important, but there is a lack of consensus on measurement and relatively few studies have explored the impact of these factors directly on admissions. Overall, recent evidence is consistent in reporting the following factors that are associated with increased risk of admission: age, deprivation, co-morbidity and recent admission. Some studies additionally report living alone, social isolation, dependency, cognitive impairment and functional problems or frailty as risk factors. The variety of potential predictors, and to some extent the lack of consistency in the data, is probably a reflection of the range of variables explored across these studies. The emphasis of data collection in these studies ranges across service use, diagnoses, functional problems, and social or economic predictors in a range of combinations. There has been relatively less emphasis on organizational or socio-economic variables in the literature, but when these are included they often emerge as key factors associated with precursors of admissions, such as functional decline and frailty.Reference Walsh, Addington-Hall, Roberts, Nicholls and Corner17, Reference Walsh, Roberts and Nicholls28, Reference Hoorgerduijn, Buurman, Koravaar, Grobbee, de Rooji and Schuurmans30 The interaction between these factors is not addressed in the current evidence. Age, deprivation and co-morbidity may not be causal components, but they might mediate the effects of functional problems, social isolation and frailty. The issue of variation in admission rates has not been fully explored and the extent to which drivers of this variation are modifiable should be a priority for future research.
Tools for identification of older people at high risk of unplanned admission
Identification of older people at high risk of admission in the community-dwelling population is seen as key to organizing and targeting admission avoidance services. Risk adjustment models stratify populations according to risk of admission or high service use. They may be used for case-finding for admission avoidance services, resource allocation, performance management and evaluation, although they have primarily been used for case-finding in the UK.31 The first consideration is about which variables are included in the model. It is known that demographic variables such as age and gender yield low predictive power, but that this can be improved by the addition of diagnostic and service utilization variables. Although social data (employment, poverty, disability) can enhance the value of these models,Reference Donnan, Dorward, Mutch and Morris12 these are rarely used. Inclusion of social care data is challenging.31 Matching of cases is problematic, social care data is less standardized and different information governance procedures may be in place. Whilst prior utilization data can significantly improve the predictive value of a demographic-based model, it has previously been noted that targeting interventions based on these models is flawed because regression to the mean may result in excessive interventions being put in place for people whose risk has returned to baseline levels.Reference Roland, Dusheiko, Gravelle and Parker32 Despite this, most current models are heavily reliant on prior utilization data.
Newer models have attempted to build on existing ones with the addition of data on factors that are known to be risk factors for admission. Diagnosis is a key risk factor,Reference Saver, Wang, Dobie, Green and Baldwin20, Reference Longman, Rolfe, Passey, Heathcote, Ewald, Dunn, Barclay and Morgan25 but it is less clear if there is any additional advantage to using diagnostic data in risk stratification. Data on health status measures have not been widely used, possibly because these are less likely to appear in claims data that are often the basis for developing these types of model in insurance-based health care systems. Health status measures are not routinely collected and the evidence does not suggest that the benefit of their inclusion would outweigh the cost of administration. These measures may require direct patient contact and, in addition to cost, response rates may be poor.Reference Dorr, Jones, Burns, Donnelly, Brunker, Wilcox and Clayton33 Saver et al.Reference Saver, Wang, Dobie, Green and Baldwin20 reported that co-morbidity was the most important predictor of ambulatory sensitive hospitalization, but their model required the computation of a co-morbidity index and the setting of a co-morbidity flag within the routine data set. The predictive value of a variable may need to be balanced against the availability of data. For example, Graf et al.Reference Graf, Gianelli, Hermann, Sarasim, Michel, Zekry and Chevalley34 compared two existing validated assessment tools but found that adjustment for functional status and co-morbidity scores did not affect the predicted risk. The period of time over which data are available also needs to be considered; most models are based on one year's prior utilization data, but accuracy improves with more than one year. A further limitation of predictive models is that they require access to large routine data sets.Reference Purdy7 Whilst national data sets are relatively easy to access, they may not contain all the relevant variables. More local data sets can be difficult to access and require extensive data cleaning. Routine data is not always accurate. Therefore the accuracy of the model is constrained by both the scope and accuracy of the routine data.
PurdyReference Purdy7 notes that these models have varying degrees of accuracy, but that models with primary care data perform about 10% better than those using secondary care data alone. The performance of current risk prediction models is only moderate, with at best 50% positive predictive value (PPV) reported. Comparison between models is difficult due to the different outcomes used (costs, resource utilization, admission, readmission, etc.). The Patients At Risk of Re-hospitalization (PARR) tool, widely used in allocation to case management services, was designed to be used in real time, whilst the patient was in hospital or shortly after to identify risk of readmission to hospital. PARR2 used any emergency admission as a triggering event, but produced more false positives. PARR2 can be used retrospectively, exploring hospitalization data monthly or annually. Although PARR2 uses a broad range of variables about the patient, community and hospital, it does not include clinical data other than diagnoses.31, Reference Billings, Mijanovich, Dixon, Curry, Wennberg, Darin and Steinort35, Reference Billings, Dixon, Mijanovich and Wennberg36 The Combined Model uses both hospital and GP data and can therefore be used for risk stratification of an entire primary care population, an approach that has become widespread. The PPV is 58% for the Combined Model vs 50% for the PARR2 model. However, the developers note that their algorithms were originally designed for use in all of England but were also tested with a sample of data from individual primary care trusts (PCTs).37–Reference Curry, Billings, Darin, Dixon, Williams and Wennberg39 Comparable results were obtained at the PCT level, but there are substantial differences between PCTs in their demographic characteristics and performance/utilization patterns, as borne out by Imison's review of variation in admission rates.Reference Imison, Poteliakhoff and Thompson1 The authors therefore caution that false positives may vary considerably across PCTs when national algorithms are applied to local data. Although the King's Fund made the software available to allow local development and modelling based on the national algorithm it is not clear to what extent PCTs and other organizations have done this. It is unlikely that GP practices, the level at which these software are often applied, have the expertise and resources to do this local development. An interesting feature of the Combined Model is that it identifies different patients from PARR, even at the highest risk levels. Using the highest 1000 risk scores, the overlap between the two systems is only 484 patients. Both models have a high rate of false positives particularly when larger risk segments are identified. Overall, the literature reports R 2 from 0.02 to 0.46 and AUC of 0.6 to 0.7 (an AUC of 0.7 and over is necessary for acceptable discrimination) for risk prediction models. Only one recent studyReference Donnan, Dorward, Mutch and Morris12 reports a much higher PPV, but this is based on a historical split cohort for validation and not on prospective studies of predictive value. Data on performance of commercial models are difficult to ascertain.
Other risk prediction tools have been developed with the aim of identifying older people at risk of readmission following acute medical admission or attendance at the ED, but sensitivity of these tools tends to be variable and specificity tends to be too low. Ben-Chetrit et al.Reference Ben-Chetrit, Chen-Shuali, Zimran, Munter and Nesher21 found a PPV of 67%. In contrast, other authorsReference Graf, Gianelli, Hermann, Sarasim, Michel, Zekry and Chevalley34, Reference Deschodt, Wellens, Braes, Vuyst, Boonen, Flamaing, Moons and Milisen40–Reference Edmans, Bradshaw, Gladman, Franklin, Berdunov, Elliott and Conroy42 did not find that the Identification of Seniors at Risk (ISAR) tool was a useful predictor of adverse outcomes after discharge from the medical unit, with only modest AUCs. However, Edmans et al.Reference Edmans, Bradshaw, Gladman, Franklin, Berdunov, Elliott and Conroy42 used a composite outcome which included death, readmission and other outcomes and found that more than three-quarters of the sample experienced an adverse outcome within 90 days. The composite outcome and high rate of adverse outcomes may have reduced the predictive power of the model. Graf et al.Reference Graf, Gianelli, Hermann, Sarasim, Michel, Zekry and Chevalley34 noted that the ISAR, and another tool, the Triage Risk Stratification Tool (TRST), did, however, have high negative predictive values and could therefore have benefit in avoiding unnecessary assessment such as CGA. Moons et al.Reference Moons, De Ridder, Geyskens, Sabbe, Braes, Flamaing and Milisen18 also compared TRST and ISAR with the Runciman and Rowland tools and found that the latter had the best predictive values for readmission. It is worth noting that this tool measures functional IADL activities and independence. The authors found that all items positively associated with readmission were functional and that this tool is easily incorporated into the nursing assessment. Drubbel et al.Reference Drubbel, de Wit, Bleijenberg, Eijkemans, Schuurmans and Numans29 also explored the predictive value of functional data using a frailty index. The frailty index tertiles discriminated different levels of risk for adverse outcomes, which included emergency department attendance and nursing home admission and had moderate predictive ability, but the ability to predict admission or readmission was not measured. These studies do, however, suggest a potentially interesting area for further research.
Although all models highlight potentially important issues, no one study has captured all the relevant factors. In addition, models used for targeting service provision assume that appropriate alternatives to admission are in place for the identified ‘at-risk’ patients and that these are effective. It is not always clear that these patients’ problems are amenable to preventative care or that effective interventions are available to deploy.Reference Roland, Lewis, Steventon, Abel, Adams, Bardsley, Brereton, Chitnis, Conklin, Staetsky, Tunkel and Ling43 Commissioners need to be clear about what type of admissions they consider avoidable, which may be contingent on availability of alternative services.Reference Purdy7 For this reason the focus in risk stratification may move towards identifying those patients most likely to benefit from intervention rather than those at highest risk.31 Risk prediction models continue to develop, but the range of possible variables for inclusion in such models is large and the predictive value is variable. Organizational and socio-economic data are underused compared with diagnostic and service use data. However, more recent analyses that explore what might be considered precursors to admission, such as functional decline, suggest that these might also be the most important factors in risk prediction. Overall, no particular approach is markedly better in terms of performance.Reference Purdy, Paranjothy, Huntley, Thomas, Mann, Huws, Brindle and Elwyn44
Effective interventions to reduce admissions and readmissions
A complicating factor in interpreting and using data from current population risk stratification models is the lack of convincing evidence for benefit of many of the interventions targeted at the highest risk sector of the population.Reference Gage, Ting, Williams, Drennan, Goodman, Iliffe, Manthorpe, Davies and Masey23, Reference Purdy, Paranjothy, Huntley, Thomas, Mann, Huws, Brindle and Elwyn44–Reference Gravelle, Dusheiko, Sheaff, Sargent, Boaden and Pickard46 The most recent systematic review of case management interventionsReference Purdy, Paranjothy, Huntley, Thomas, Mann, Huws, Brindle and Elwyn44 concluded that there was no evidence of reductions in unplanned admissions in the majority of studies investigating case management, other than for specific groups such as those with heart failure. The most recent review of interventions for the general older population rather than specific disease populations, found that self-management interventions, hospital at home, care pathways, medication review and integration of secondary and primary care services are not effective in this group.Reference Purdy, Paranjothy, Huntley, Thomas, Mann, Huws, Brindle and Elwyn44 The same review concluded that there is insufficient evidence on home visit interventions. Another reviewReference Tappenden, Campbell, Rawdin, Wong and Kalita47 looked at nurse-led health promotion interventions for people aged 75+ years living in the community. They concluded that the experience and training of nurses was not consistently described, that the content and frequency of intervention visits was varied and that the focus was on monitoring, assessment and information-giving. From the 11 trials reviewed, there was a reduced risk of death, improved quality of life (QoL), but unclear data on functional outcomes. Organizational outcomes such as admission were not improved, but the authors concluded that these interventions have an overall benefit to health and well-being. Most recently, Lewis et al.Reference Lewis, Georghiou, Steventon, Vaithianathan, Chitnis, Billings, Blunt, Wright, Roberts and Bardsley48 concluded that virtual or community wards (a multi-disciplinary refinement of the case management approach) do not result in reductions in unplanned bed use or costs. Despite some potentially positive health benefits of these interventions, there is no evidence that they have any impact on admissions. This evidence must be considered in the context of largely qualitative evidence that case management may impact on other important outcomes, including total length of stay and quality of life.Reference Purdy7, Reference Williams, Smith, Chapman and Oliver49–Reference Challis, Hughes, Berzins, Reilly, Abell, Stewart and Bowns51 However, there is also some evidence that implementing the intervention is challenging and the content of these interventions continues to develop.Reference Sargent, Boaden and Roland24, Reference Challis, Hughes, Berzins, Reilly, Abell, Stewart and Bowns51 More recent trials have also failed to demonstrate any positive impact on admissions and variation in other outcomes.Reference Metzelthin, van Rossum, de Witte, Ambergen, Hobma, Sipers and Kempen GIJM52–Reference Van Hout, Jansen, van Marwijk, Pronk, Frijters and Nijpels54 The only recent study to report a reduction in admissions was a relatively small trial in which admission was a secondary outcome.Reference Imhof, Naef, Wallhagen, Schwarz and Mahrer-Imhof55
Ellis et al.Reference Ellis, Whitehead, Robinson, O’Neill and Langhorne56 carried out a meta-analysis of Comprehensive Geriatric Assessment (CGA) for older adults admitted to hospital. Although this analysis focused on patients who were admitted as an emergency and not the broader community-based population, it is important in its possible relevance to preventing readmission. The analysis found, across 22 trials of CGA, that recipients of this intervention were more likely to be alive, more likely to be living in their own home and less likely to be in residential care than patients receiving standard care. However, evidence for CGA in urgent care in order to divert patients from admission is limited. Whilst a small non-randomized study of geriatrician review of older people referred for admission to the acute unit suggests that CGA might have benefit in admission avoidance in urgent care,Reference Bouzon, Torregrossa, Larrain and Rodriguez Manas57 a small randomized trialReference Edmans, Bradshaw, Franklin, Gladman and Conroy58 and recent review of CGA at discharge for short-stay in frail older peopleReference Conroy, Stevens, Parker and Gladman JRF59 conclude that there is no benefit in terms of admission prevention.
Integrated care, which offers a range of services to older people across the risk strata, has been recommended as a means to improve patient experience and outcomes, including reducing avoidable admission.Reference Oliver, Foot and Humphries60 The evidence to support this approach is limitedReference Purdy7 and the suggestion that change must be rapid and large scale to be usefulReference Oliver, Foot and Humphries60, Reference Ham and Walsh61 does not appear to be evidence based. However, the current emphasis on integrated care does recognize that admission avoidance interventions are not usually deployed as single interventions but as part of a complex menu of service options.Reference Purdy, Paranjothy, Huntley, Thomas, Mann, Huws, Brindle and Elwyn44, Reference Oliver, Foot and Humphries60, Reference Walsh, Lattimer, Wintrup and Brailsford62 There is currently insufficient evidence on the effect of the range of interventions across the whole system.
Conclusions and recommendations for future research
Rising hospital admission and readmission rates in older people remain a challenge internationally and this review highlights both gaps and inconsistencies in the evidence. Large-scale studies of admission rates, particularly geographical or service-level comparisons, would provide better baseline information on which to plan services and evaluate screening tools and interventions. From a service perspective, a number of issues need to be considered in both identifying and intervening with those at risk of unplanned hospital admission. The evidence indicates that age, deprivation and co-morbidity are key drivers, but the benefit of using additional variables in predictive algorithms needs consideration. Regression to the mean calls into question the utility of models that are over-reliant on prior service use data.Reference Roland, Dusheiko, Gravelle and Parker32 The addition of socio-economic variables can increase the predictive power of risk stratification tools, but such factors may not be modifiable. They can be used to trigger additional input for some patients. The poor specificity and high false positive rates in risk stratification models may indicate that these systems are more useful for service-level planning, but for clinical management are of value only as a precursor to more detailed clinical assessment. Models that include measures of function, disability or frailty may be more effective in targeting relevant patients, but are likely to require additional data gathering beyond routine clinical measures, an approach associated with additional administrative and clinical costs. For the purposes of overall reduction in admissions, tools that are designed for post-admission screening may have more benefit. A number of generic and disease/context-specific tools exist, which have reasonable predictive power in relation to readmissions. In either case, identifying those at risk is only useful if suitable interventions exist and here the evidence base is lacking. Clearly, for those admitted to hospital, CGA-based interventions are likely to be of benefit. At the population level, having identified the risk strata of interest, it is less clear if CGA or case management approaches are beneficial. The lack of knowledge about identifying symptoms, warning signs for admission, appropriate interventions and social or personal factors that might limit self-management is problematic, as is variation in interventions being delivered. The King's Fund37 note that answers to these questions will be important in creating effective interventions.
It is clear that a key priority for future research will be exploring the diversity of the patient group identified as at high risk of unplanned admission. In recent years, there has been a move away from focused case management interventions on a relatively small extremely high-risk risk group towards provision of an integrated menu of services that are available to the top 5% risk stratum. Current models, when used to identify a larger stratum of risk, will undoubtedly generate a variable group of ‘at-risk’ patients given that risk might be attributable to diagnoses, prior service use, socio-economic status, functional problems and/or specific needs such as dementia care or end-of-life care. As the risk strata referred for admission avoidance services increases in size and complexity there will be a greater need to direct patients to appropriate services. These changes in service organization appear to have led to a considerable increase in caseload size and complexity, with a far greater range of patient problems ranging from end-of-life care needs to social care. There are indications that, where risk prediction software continues to be used, clinicians then need to make judgements about the nature and intensity of service required for individuals. Oliver et al.Reference Oliver, Foot and Humphries60 in their guidance on implementing integrated services, recommend that risk stratification be linked to additional clinical assessment and screening. Understanding the complexity of the target group and ensuring that patients can be referred to appropriate services will become a key challenge of this new approach.
Implementation of integrated care services requires further supporting evidence. These services are often implementing a range of admission avoidance interventions (such as case management or virtual wards), which themselves have little supporting evidence. The mixed picture regarding the effectiveness of admission avoidance interventions suggests another area that should be a priority for future research. The evidence to date does not support any particular admission avoidance intervention but there is evidence suggesting improvements in other outcomes including overall length of stay and quality of life. However, these outcomes are rarely the primary study end-points and it is not clear to what extent these studies are powered to detect differences in secondary outcome measures. Given the difficulty in demonstrating differences in admission rates, at least in part due to the problem of regression to the mean, there is a need to identify other important outcomes for these types of intervention. Finally, whole-system modelling approaches may allow current evidence on the effect of a range of interventions to be used to explore the summative effect on admissions when deployed in various configurations across the health economy.
