Objective To develop a screening tool to identify elderly patients at the end of life and quantify the risk of death in hospital or soon after discharge for to minimise prognostic uncertainty and avoid potentially harmful and futile treatments.
Design Narrative literature review of definitions, tools and measurements that could be combined into a screening tool based on routinely available or obtainable data at the point of care to identify elderly patients who are unavoidably dying at the time of admission or at risk of dying during hospitalisation.
Main measurements Variables and thresholds proposed for the Criteria for Screening and Triaging to Appropriate aLternative care (CriSTAL screening tool) were adopted from existing scales and published research findings showing association with either in-hospital, 30-day or 3-month mortality.
Results Eighteen predictor instruments and their variants were examined. The final items for the new CriSTAL screening tool included: age ≥65; meeting ≥2 deterioration criteria; an index of frailty with ≥2 criteria; early warning score >4; presence of ≥1 selected comorbidities; nursing home placement; evidence of cognitive impairment; prior emergency hospitalisation or intensive care unit readmission in the past year; abnormal ECG; and proteinuria.
Conclusions An unambiguous checklist may assist clinicians in reducing uncertainty patients who are likely to die within the next 3 months and help initiate transparent conversations with families and patients about end-of-life care. Retrospective chart review and prospective validation will be undertaken to optimise the number of prognostic items for easy administration and enhanced generalisability. Development of an evidence-based tool for defining and identifying the dying patient in hospital: CriSTAL.
Statistics from Altmetric.com
The natural progression of chronic disease involves periods of apparent remission interspersed by exacerbations and, in the year leading to death, multiple hospitalisations.1 Some indicators of poor prognosis can suggest a patient is nearing the end of life,2 and have been found useful for initiating discussions with families regarding pre-emptive care planning.1 Yet there is uncertainty of the time, frequency and duration of the next episode of decompensation as well as the ultimate prognosis causing doubts about whether to continue active management. Further, while the majority of people want to die at home, most will die in hospital.3–8 Patients nearing the end of life are high-level users of ambulance services,9 emergency services,2 ,10 hospital wards11 or intensive care units and many die in hospital.12 Significant numbers of patients with cancer or other terminal illnesses are suitable for palliative care but often are readmitted to acute hospitals multiple times with lengths of stay of just under a week.10 ,13 ,14 While there are accepted policies for de-escalating treatment in terminally ill patients,2 ,15–17 there are also inherent and societal pressures on medicine to continue utilising technological advances to prolong life even in plainly futile situations.15
The implications of a decision to administer or withhold aggressive treatment for terminal patients are enormous for clinicians, patients and their families, for the health system and for society as a whole. It can be difficult to reach a decision that balances the rights of patients to die with dignity18 ,19 and the expectations of families about satisfactory end-of-life care,20 ,21 while considering the limitations of health resources where opportunity costs cannot be disregarded.22–24 Delaying unavoidable death contributes to unsustainable and escalating healthcare costs, despite aggressive and expensive interventions. These interventions may not influence patient outcome; often do not improve the patient's quality of life; may compromise bereavement outcomes for families; and cause frustration for health professionals.20 ,25–29 This highlights the importance of developing more accurate ways of identifying patients near the end of life, involving both the patients and their carers in those discussions and then making more appropriate management plans.
For about two decades many acute hospitals have adopted rapid response systems to identify and manage seriously ill patients.25 ,30–32 They were initially developed to recognise at-risk patients early as a basis for triggering a rapid response to improve patient outcomes. In doing so, the system also identifies patients at the end of life who are predictably deteriorating. Up to one-third of all of rapid response team (RRT) calls have been related to end-of-life issues.33 ,34 This emphasises the failure of current hospital systems to recognise patients at the end of life. Often it is the patient and carers who initiate this conversation.25 ,35
Clinical decision aids are widely used to involve patients in informed treatment decisions that incorporate their personal preferences and values.36 Sensitive clinical decision rules have been used to discontinue futile resuscitation on patients who experience a cardiac arrest.37 However, we have not found a fit-for-purpose screening checklist or clinical decision tool for objective identification of end of life within days, weeks or months to minimise inappropriate treatment at hospital admission.29 There is a need to recognise patients at the end of life while at the same time acknowledging uncertainty around the exact time and circumstances when death will occur.38
The aims of the CrisTAL checklist are to assist clinicians to recognise these patients and to change the culture of the hospital to one where end of life is openly discussed and dealt with more appropriately.39
Accordingly, there is a need to collate evidence to assist clinicians, carers and families in decision-making about the most sustainable model for appropriate and best quality care in the last few months of life. The specific objectives of this research are to:
review literature to obtain definitions for dying patient and end of life;
review existing literature regarding screening tools for the prediction of death in hospitalised patients;
propose a checklist for screening of hospitalised patients at-risk of dying in the short to medium term.
Two common and important situations where patients at the end of life can potentially be identified are on admission to the emergency department (ED); and when a patient deteriorates and becomes the subject of a RRT call. This paper reports on the development of a clinical decision aid for use in both circumstances: CrisTAL (Criteria for Screening and Triaging to Appropriate aLternative care). It summarises the information available in the literature to construct the domains for such a screening instrument based on patient data items routinely available at the point of care.
The tool is intended to offer a starting point to begin discussions with the patient and relatives about priorities and preferences on type and place for end-of-life care.39 ,40 It also may identify elderly who will benefit from alternative care pathways instead of hospitalisation.5 ,41 The routine use of such a tool may also change the culture of the organisation to one which is more aware of patients who may be at the end of life and one where different management pathways are considered earlier. The tool is not meant to dictate whether or not a patient receives life sustaining therapy or is the subject or a do-not-resuscitate order. However, it may provide an objective assessment to inform and support that decision, made jointly by patients, their family and the treating team.
We undertook a narrative literature search in PubMed, Cochrane Library and Google Scholar for published and unpublished papers about explicit and practical definitions of ‘end of life’ and for tools or screening instruments to predict death. The search strategy included the following terms: (End of life, terminal, dying, inappropriate resuscitation, do-not-resuscitate, cardiopulmonary resuscitation order, limitations of treatment, discontinuation of care, futility, advanced directive) and (hospital, acute care facility, palliative care, ED) and (Screening tool, decision aid, algorithm, predictive, predictor of death). This was supplemented with manual searches through the reference lists of eligible papers.
The variables and thresholds explored for the screening tool were adopted from existing scales and published findings that demonstrated their association with either in-hospital or 30-day mortality or survival to 12 weeks.42–47 Based on the practicalities of applying the tool as decision-making support at the point of care, we used four criteria to decide whether the existing instrument was helpful for the purpose of objectively diagnosing dying and whether to discard items: ready availability in medical records,42 ,43 need for clinical judgement, use of value judgment and self-sufficiency of indicators. This review was followed by consultation with two doctors and three ICU nurses with intensive care qualifications and experience in end-of-life care, about the feasibility of acquiring or documenting these data items in routine care.
We found 112 relevant articles dealing with the definition of dying, determination of severity of deterioration, prediction of in-hospital death, preferred place of death and options for alternative end-of-life care. Among these, we identified 18 instruments and their variants validated in different settings. Below is a summary of the operational definitions and commonly used or cited tools to predict death in hospital.
Nine working definitions of end of life were found to assist in limiting the number of items for a screening tool to a manageable set (table 1). These were mostly impractical in their requirement of clinicians’ subjective assessment; or confined to patients imminently dying within hours; and of limited use for elderly patients with chronic disease, nearing end of life within days or weeks.
We defined inappropriateness of admission to hospital for patients at the end of life as those ‘admissions when the resources of the hospital will not have any significant impact on the clinical prognosis of the elderly patient with multiple life-threatening comorbidities’.
As pragmatic definitions of ‘dying patients’ were not prevalent in the literature, we searched for a suitable proxy measure that could be drawn from studies examining predictors of poor survival. These are abundant and cover both subjective and objective parameters anticipating death.
Subjective variables and their utility in predicting short-term to medium-term mortality
Of the instruments developed in the past 30 years for prognostication of death after admission, many still require value judgements and unstructured subjective assessments, which renders them less reassuring and hence less useful as a tool for deciding at the time whether to admit a patient.
Performance Status Scales designed as early as 1949 by Karnofsky53 ,54 and The World Health Organisation (ECOG PS) in 1982 are simple and popular instruments for determining appropriate intensity of care for patients. They have undergone adaptations over time55 where completion still involved major value judgements, which makes them impractical and unreliable for a standardised prognosis (table 2).
Various indices have been designed to identify illness severity and risk of death after admission (table 2). Some reliably capture the level of quality of life in terminal patients but do not focus on objective signs;54 some use nursing assessment of organic and psychosocial aspects;56 others suggest a checklist that combines objective (eg, semiconsciousness) and subjective items (eg, ‘irreversible deterioration’).51 Some emphasise application of survival prediction for in hospital-based palliative care services with high prognostic accuracy (85.6%) in estimating death within 3 days of admission to a palliative care facility, but only 54% and 57.6% accuracy in predicting death within 4–30 days and by 6 months.57
A global assessment of frailty using a subjective score between 1 (very fit) and 7 (severely frail) had good predictive validity for death within 18 months58 but required clinical and value judgements, and did not incorporate the impact of underlying conditions, hence reducing its ease of use for routine care by less experienced personnel. Clinician perception about risk of death has been found to be reasonably accurate in particular for patients with advanced chronic heart failure or chronic obstructive pulmonary disease as an adjuvant in the decision to more efficiently target palliative services and end-of-life care planning.59
Finally, the global self-rated question designed to assess patient perception of their own health in comparison with other people their age, not intended as a prognostic tool but since the early 1980s has been associated with predicting long-term mortality in the elderly,60 independently of ‘objective health status’ and across ethnic groups. The self-rated health question is now better understood,60–62 and has been validated as a screening tool for vulnerable people at higher risk of death in community.63 Its influence on imminent risk of death at hospital admission is not known and we will include it in our screening tool.
Objective variables and their utility in predicting short-term mortality
Several factors have been found to have an impact on the risk of death after hospital admission, including age 65 years and above,42 ,63 ,85 ,86 multiple comorbidities,18 multiorgan failure,44 physiological data from laboratory test results64 ,76 and type of service and urgency of admission.43
We propose a combined algorithm quantifying the aggregate risk estimation of some previously developed instruments to take us closer to a more accurate definition of dying. An historical exploration of 18 of these estimates has shed more light on the influence of these factors.
The diagnosis of advanced cancer has probably attracted the most attention for predicting prognosis and appropriate care. From a review of 24 studies and 18 prognostic indicators, there was general agreement that anorexia and weight loss showed the most significant association with poor survival, followed by cognitive impairment, dyspnoea and dysphagia.83 While several of these studies were conducted in small convenience samples, some with doubtful statistical methods,42 clinicians would agree that these are largely symptoms of imminent death. Uncertainty of what constitutes dying in the short term has led to the development of practical prediction tools to assist in treatment decision-making, guide family consultations, and minimise unnecessary expense to the health system (table 2).
Prognostic scales and indices
Performance Status Tools have been well received and modifications tested in various settings.
Table 2 summarises scales found a predicting outcome and time to death/discharge, some of which have been validated in similar or divergent populations and others have led to refinements and developments of further tools.72 many are cancer-specific scales, thus have limited value for wider use in ED.82 For instance the PaP score is good at reducing the prognostic uncertainty of death within 1 month of admission to palliative care services.76 However, it is only validated for patients with cancer and it can yield significant differences between the prediction of registered nurses and doctors.76 ,77 ,87
The Charlson Comorbidity Index (CCI), was designed to estimate 1–10-year mortality in longitudinal studies and is not validated as prognostic indicator for short-term outcomes in cancer or other conditions.88 The Elixhauser Comorbidity Index is a complex tool74 which uses administrative databases to estimate increased risk of in-hospital death or prolonged hospital stay.68 but clinicians may not find it user-friendly because it relies on administrative data and requires calculations. Further, the authors recommended a combination with other influential factors for a more accurate prediction of death in hospital.18
The Acute Physiology and Chronic Disease Evaluation (APACHE II) tool and its variants APACHE-L, APACHE III and APACHE IV and the Simplified Acute Physiology Score, SAPS II were designed to measure the severity of disease for adult patients and are all used to predict in-hospital death and risk-adjusted length of stay in intensive care units.65–67 ,70 ,89 The scores indicate the risk of death in patient groups rather than individual prognosis.90 Moreover, the APACHE instruments are heavily dependent on laboratory-based data not generally available in all EDs in Australia.
Multiple attempts have been made to enhance objective early warning scores (EWS) for identification of critical illness and deterioration on admission and in intensive care. Improvement in serial EWS within 4 h of presentation to hospital predicts improved clinical outcomes75 ,79 ,84 ,91 ,92 hence EWS has been deemed as a potential triage tool in the ED for acute medical patients.75 ,79 ,81 ,84 ,93 While the developers of some EWS have emphasised that they did not intend them as predictors of patient outcome,94 experience has shown that these scores are being used in practice to predict death. Accordingly, we chose to include these in the construction of algorithms defining the diagnosis of dying.
In 2012, the Rothman Index was found to be a strong predictor of both in-hospital mortality, hospital readmission and post-discharge mortality at 2 days, 30 days and 1 year.44 ,56 Unfortunately the Rothman Index relies on comprehensive collection of nursing or doctors’ assessments, not part of routine care in outpatients or ED in most hospitals.
Development of CriSTAL
To be considered useful on admission at ED or during an RRT attendance, the screening tool items should meet the following criteria: easily collected in routine practice,42 or readily available in electronic or paper medical records; does not require specialist clinical judgement; is sufficient to independently predict death in specific conditions; and with two exceptions, does not employ a value judgment. None of the 18 published predictive tools met the four criteria; five met three criteria but four of these instruments involved clinical judgements and one involved value judgment; nine tools met two criteria and four tools only met one criterion. Figure 1 shows the distribution of criteria to justify the need for a fresh tool that met the four criteria.
In the absence of a comprehensive instrument combining acute and chronic predictors to increase certainty of diagnosis of imminent death or death within weeks or months, this review gathered recognised predictors of death for elderly patients with complex health profiles from existing prognostic tools to create a new screening instrument. We anticipated that incorporating objective variables would enhance certainty of the screening tool and could assist in the decision to appropriately generate do-not-resuscitate orders25 and consider alternative end-of-life care orders.
The variables and values proposed for the CriSTAL screening tool were adopted from existing scales and from published research findings demonstrating their association with either in-hospital or 30-day mortality or survival to 12 weeks.42–45 ,56 ,76
Old age and RRT criteria are priorities on the checklist. Supplementation with a quantifiable level of severity based on EWS75 ,79 ,84 ,91; history of repeat hospitalisations with or without admissions to ICU95; emergency admission96; at least one of the predefined advanced comorbidities from the evidence-based list;35 an objective measurement of frailty;85 documentation of nursing home placement;33 ,46 evidence of cognitive impairment;25 ,42 ,63 ,73 ,83 and readily available test results: proteinuria and if ECG confirms abnormalities.97
Table 3 shows our resulting 29-item screening tool, named CriSTAL, to denote our intention to introduce transparency in the identification of the dying patient and enable objective clinician decisions about prognosis and justification for administering or de-escalating aggressive treatments.
A slight modification is proposed for the use of the tool following a RRT attendance (table 4). This might encourage reassessment of the need for continuing hospitalisation in an acute care facility and discussion about the need for limitations of treatment if death is imminent.
How would CriSTAL be used in practice? It may characterise ‘appropriateness of admission’ and appropriateness of subsequent treatment for patients at the end of life in a way that can be applicable to a wide variety of terminal health conditions. It could be used as a platform for beginning discussions with patients and their carers. It may also add more certainty identifying the irrevocably dying patient with chronic comorbidities and prevent further futile treatments to prolong life. We have omitted indicators of system failures or nursing staff workloads such as TISS or NEMS90 that may potentially influence risk of death, as these did not fulfil our inclusion criteria of being routinely available
The Scottish health system implemented a national action plan for care at the end of life deriving from the realisation that 30% of all hospital bed-days were accounted for by multiple admissions of people in the final year of life.41 The ‘Dying well’ premise in Scotland is that alternative care is integral to continuity of care outside the hospital. The strategy includes among others, early identification of care needs for any terminal illness, holistic assessment and involvement of patients and families in the coordination of alternative care.2
Inspired by this development, our definition of inappropriate hospital admission is linked to the more objective scoring factors of the CriSTAL tool, whose accuracy is to be validated to more precisely establish the main determinants of death in the short term. Our review indicated that old age42 ,63 ,85 ,86 and concurrent illnesses18 ,25 ,42 ,46 ,63 ,101 were the strongest predictors of death in and outside intensive care.45 ,104
Strengths of this developmental work are the evidence-base source of variables in the tool and the extensive range of predictors covering demographic, physiological and diagnostic prediction measures.
A limitation of this research is that the item selection was based on a narrative review with focused set of search terms. This may have led to overlook of some articles that would have been captured in a systematic and broader search strategy. However, the comprehensive search for tools and the breadth of instruments found using this approach provided a sufficiently large number of items to start the discussion on possible amalgamation of variables from existing instruments to meet our targeted need. Other researchers among the readership may choose to expand the search or enhance the tool. In fact, a limitation of CriSTAL's development at this stage is its length for routine administration, and the number of potential predictors which may lead to model ‘overfitting’. The testing of too many variables is known to reduce the generalisability of the predictive model.105 By retrospective testing and future prospective validation we hope to reduce the total number of items without sacrificing predictive accuracy or generalisability. Initially, CriSTAL's 29 subitems will be tested in a retrospective data review using a case–control study design where cases are all deaths reported from the RRT attendances system in a teaching hospital during 2012–2013. Controls will be age-sex-ward matched records of patients admitted in the same period with an RRT call but did not die before or within 3 months of discharge. Sensitivity, specificity and positive and negative predictive values will be calculated from logistic regression models of matched cases and controls. This retrospective validation has been endorsed by the South Western Sydney Local Health District Ethics Committee. The next step after retrospective testing will be the prospective administration of the validated tool as part of the admissions procedure in emergency and after the RRT calls on general wards.
The accuracy of models with different number of variables will be determined using the area under the receiver-operating characteristics (AUROC) curve.106 Minimum accuracy will be defined as area under the ROC curve >80%, and variables not contributing significantly to the model will be dropped from the instrument. Survival analysis and Cox proportional hazards regression will investigate the most significant predictors of imminent death. A 5% chance of survival to hospital discharge among those predicted to die will be chosen as the maximum error allowed for the tool to be considered useful.
While it is acknowledged that predictions based on population subgroups are not meant to be used for individuals,107 the calculated risk can be used as a reference to inform the decision by the individual under the clinician's guidance, on whether or not to continue aggressive treatment, given the odds of dying based on the well-established predictors. Careful use of the CriSTAL tool care for decision-making would involve alignment with quality of care principles and patient values and preferences, and should not be driven by hospital financial pressures or need to meet health system performance indicators.24
Finally, it is important to recognise that the use of a screening tool for identifying patients who have a high probability of dying within 3 months can only provide an indication of those who with a low probability of survival and will not be a signal of absolute certainty.50 Testing its appropriateness, reliability and predictive value in different patient subpopulations will help reduce this uncertainty but its predictive value may vary in different settings and for different timeframes and this needs to be ascertained. Further, its values after an RRT response will need to be assessed in relation to its value at the time of admission for patients when trialled. As emphasised before, testing in different settings could yield different predictive performance depending on the patient profile and possibly the influence of factors not accounted for in the tool. Readers and researchers are encouraged to train and validate the CriSTAL tool in their facility to generate the most valid and relevant set of variables for their subpopulations.
Conclusions and recommendations
This tool does not intend to preclude access to healthcare for the terminal elderly, but to provide an objective assessment and definition of the dying patient as a starting point for honest communication with patients and families,3 about recognising that dying is part of the life cycle. Dignified withdrawal of intensive and inappropriate treatment29 ,52 and triage into alternative care in non-acute facilities10 ,38 is an area where there is still ample room for improvement.1 ,29 ,39 Standard guidelines for alternative end-of-life care are not yet broadly adopted in Australia and discussions with policy-makers need to continue.2 However, increasing evidence of alternative out-of-hospital care acceptable to clinicians108 and others are known to include sedation to minimise distress, pain management,109 spiritual support,41 music therapy and home-based palliative care.110 If proven accurate in the prediction of short-term death, a reduced version of CriSTAL could be proposed for routine use at hospital admission. We acknowledge that the Australian health system may not yet be equipped to respond to the demand for alternative healthcare facilities for the dying.111 However, it is hoped that using such predictive tools may encourage more appropriate services for managing patients at the end of life.
Training for nurses and doctors in the use of the screening tool and in approaching patients and families with concrete information about inevitability of death and lack of benefit of further intensive treatment are paramount.27 ,112 They will be better equipped to communicate the responsible decision to suspend efforts and handle potential requests for futile treatment.41 ,49
Automation of CriSTAL and its scoring would facilitate use at time of admission and production of instant or retrospective locally relevant profiles of patients imminently dying. Potential uses include as a clinical support tool for decision-making on triage to appropriate end–of-life care facilities; to prevent death in some cases; and to examine variation in risk-of-death levels, differences in admission practices, and inform triage policies across hospitals,43 as a first step into cost-effectiveness and patient satisfaction studies.
- Received 31 July 2014.
- Revision received 23 October 2014.
- Accepted 23 November 2014.
Contributors MCM conceived the idea of an objective assessment. KH and MCM led the literature searches. MCM summarised findings, wrote the first and final draft manuscript and refined all versions. KH contributed to the literature review, provided ongoing intellectual input and revised subsequent manuscript drafts.
Funding This article presents independent research funded by a grant from the National Health and Medical Research Council of Australia (# 1054146).
Competing interests KH is an advocate for dying with dignity and has written books and articles on the subject.
Provenance and peer review Not commissioned; externally peer reviewed.
Provenance and peer review Not commissioned; externally peer reviewed.
Data sharing statement No additional unpublished data exists from this study. A retrospective validation of the screening tool is in preparation. Protocols and materials for the conduct of that validation could be shared with other research groups interested in testing the screening tool in their patient populations.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.