Article Text
PDF
Abstract
Background Precise prognostic information, if available, is very helpful for guiding treatment decisions and resource allocation in patients with non-cancer non-communicable chronic diseases (NCDs). This study aimed to systematically review the existing evidence, examining prognostic models and factors for identifying end-of-life non-cancer NCD patients.
Methods Electronic databases, including Medline, Embase, CINAHL, Cochrane Library, PsychINFO and other sources, were searched from the inception of these databases up until June 2023. Studies published in English with findings mentioning prognostic models or factors related to identifying end-of-life in non-cancer NCD patients were included. The quality of studies was assessed using the Quality in Prognosis Studies tool.
Results The analysis included data from 41 studies, with 16 focusing on chronic obstructive pulmonary diseases (COPD), 10 on dementia, 6 on heart failure and 9 on mixed NCDs. Traditional statistical modelling was predominantly used for the identified prognostic models. Common predictors in COPD models included dyspnoea, forced expiratory volume in 1 s, functional status, exacerbation history and body mass index. Models for dementia and heart failure frequently included comorbidity, age, gender, blood tests and nutritional status. Similarly, mixed NCD models commonly included functional status, age, dyspnoea, the presence of skin pressure ulcers, oral intake and level of consciousness. The identified prognostic models exhibited varying predictive accuracy, with the majority demonstrating weak to moderate discriminatory performance (area under the curve: 0.5–0.8). Additionally, most of these models lacked independent external validation, and only a few underwent internal validation.
Conclusion Our review summarised the most relevant predictors for identifying end-of-life in non-cancer NCDs. However, the predictive accuracy of identified models was generally inconsistent and low, and lacked external validation. Although efforts to improve these prognostic models should continue, clinicians should recognise the possibility that disease heterogeneity may limit the utility of these models for individual prognostication; they may be more useful for population level health planning.
- Advance Care Planning
- Chronic conditions
- Chronic obstructive pulmonary disease
- Heart failure
- Prognosis
- Neurological conditions
Data availability statement
All data relevant to the study are included in the article or uploaded as supplementary information.
This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/.
Statistics from Altmetric.com
- Advance Care Planning
- Chronic conditions
- Chronic obstructive pulmonary disease
- Heart failure
- Prognosis
- Neurological conditions
WHAT IS ALREADY KNOWN ON THIS TOPIC
There is a recognised gap in understanding prognostic models and factors specific to identifying end-of-life in patients with non-cancer non-communicable chronic diseases (NCDs). The existing body of knowledge highlights the significance of precise prognostic information for guiding treatment decisions and resource allocation in these patients.
WHAT THIS STUDY ADDS
This study identified key predictors for non-cancer NCDs, including chronic obstructive pulmonary diseases, dementia, heart failure and mixed NCDs. While the study revealed significant predictors, it also highlighted the varying predictive accuracy of the identified prognostic models and their frequent lack of external validation; thus, emphasising both the need for robust, validated models in this context as well as the potential limits of such models.
HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY
This study serves as a foundation for further investigation, offering the potential to refine and develop more accurate prognostic models tailored to non-cancer NCD patients.
While models thus far have limited accuracy for individual prognostication, there is potential for healthcare providers and policymakers to use these models to improve the quality of care and optimise resource allocation for populations with non-cancer NCDs approaching the end of life.
Introduction
Non-communicable chronic diseases (NCDs) are currently a major global, national and individual health concern.1 2 Socioeconomic advancements, combined with technological innovations and improvements in healthcare systems, have resulted in increased life expectancy and, as a result, increased rates of chronic conditions.3 People are living longer and surviving a variety of chronic and progressive diseases.3 This leads to the accumulation of conditions that significantly affect well-being and cause disabilities, eventually leading to severe organ failure (eg, heart, lung, liver, kidney and neurological) in the final years of life.3 4 This emerging epidemiological paradigm, characterised by the co-occurrence of organ failure(s), advanced age and frailty in the same patient, has the potential to alter the previously established trajectories of functional and vital decline,3–5 posing a challenge to our approach to optimising medical and supportive care.
Several studies have found that patients with advanced non-cancer NCDs have the same or more symptoms than patients with various cancers.6–10 High symptom burden, combined with a poor prognosis and inadequate communication, creates a particularly complex care situation that adversely affects patient quality of life and autonomy.11–13 A more holistic approach to end-of-life care and improved communication are needed for these patient groups. Implementing these targeted services in non-cancer NCDs with severe organ failure has been shown to improve survival, patient and family satisfaction, and resource utilisation throughout the disease process, and particularly in the final days of life.14–16 In this endeavour, public health advisory organisations such as the WHO have advocated for the early integration of palliative care with other therapies aimed at extending life.17
Although some aspects of palliative care should be considered early in the course of non-cancer NCDs, the appropriate timing of end-of-life discussions and referral for palliative care is more difficult due to the unpredictable nature of disease progression and a lack of prognostic information.18 19 In advanced non-cancer NCDs, determining prognosis is more difficult than in cancer. Most of these diseases have death trajectories that involve episodic, acute exacerbations, frequent hospitalisation, stabilisation and steady decline, making the palliative status determination and referral to hospice care more difficult.18 19 This challenge is especially pronounced in high-risk populations, where care goals must be reassessed, medically necessary therapies must be redefined, symptom control must be prioritised, other physical, psychosocial, and spiritual issues must be assessed, and earlier interventions must be considered.20 Precise prognostic information, if available, is very helpful for individual decision-making but can also support shared decision-making, estimate healthcare utilisation and identify groups that would benefit from specific interventions.20 21 For instance, within the context of NCDs, understanding which patient groups are likely to experience either rapid or prolonged disease progression enables health planners to make informed decisions, ensuring that resources are allocated where they are most needed.22
Several variables have been identified that are useful in estimating prognosis in patients with NCDs, and scores that combine several variables have been developed.20 23 However, widely used prognostic models like the Palliative Prognostic Score, Palliative Performance Scale, Prognosis in Palliative Care Study (PiPS-A and B) and Palliative Prognostic Index were primarily or exclusively developed for advanced cancer populations.20 24 This raises concerns about the appropriateness of using these prognostic models for patients with advanced non-cancer NCDs such as organ failure, degenerative neurological conditions and frailty.
A systematic review focused on prognostication in chronic obstructive pulmonary diseases (COPD) patients identified several variables contributing to all-cause mortality.25 However, only a limited number of studies have been designed to assess or report the prediction of mortality with a specific focus on identifying end-of-life patients. The quality of evidence remains low, suggesting that no single variable or multivariable score can be reliably recommended.25 26 It is also reported that prognostic factors that are useful for predicting long-term mortality, such as 5-year or 10-year prognosis, may not be the most relevant predictors for identifying patients likely to die within the next year.26 As the last year of life is considered a reasonable timeframe for proactive identification of patients in need of palliative care,27 examining models or factors that predict mortality within this critical phase are highly needed. Understanding which tools and variables can effectively predict prognosis during this period becomes paramount in improving patient care and informed decision-making processes. Another study of palliative care and prognosis in COPD patients concluded that none of the suggested criteria for initiating palliative care based on an expected poor vital prognosis in the short or medium term provides sufficient reliability.28 Similarly, a study of prognostic indicators of 6-month mortality in elderly patients with advanced dementia found that the Functional Assessment Staging (FAST) criterion was not a reliable predictor of 6-month mortality, and there was a lack of prognostic concordance across the identified studies.29 These failures in prognostic modelling are partly attributed to methodological issues, as many were developed with small sample sizes and lack of validation in independent populations.30–33 For example, various modifications have been made to the available tools for advanced non-cancer NCDs (such as the National Hospice Organisation; now the National Hospice and Palliative Care Organisation criteria), but their calibration and/or discriminative power have not been optimal.23 34
Therefore, the primary aim of this systematic review was to comprehensively synthesise the available evidence on prognostic models and factors associated with mortality in patients with non-cancer NCDs. To achieve this aim, the review addressed the following specific questions: What key prognostic factors and prediction models of mortality are available for identifying end-of-life in non-cancer NCD patients? What are the methodological characteristics of these prognostic models and factors? By addressing these questions, the review aimed to provide a comprehensive understanding of the factors linked to end-of-life in non-cancer NCDs and assess the reliability and validity of existing prognostic models.
Methods
This systematic review was conducted specifically to synthesise available evidence on prognostic models and predictors of mortality among patients with non-cancer NCDs. The protocol for this review was registered in the PROSPERO international prospective register of systematic reviews (registration number: CRD42022342798). The results were reported following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines.35
Search strategy and study selection
A comprehensive systematic literature search was conducted using various approaches. The search strategy was developed by BMG, MLH and JRA, with assistance from the College of Health, Medicine and Wellbeing librarian to refine the search. The search strategy involved a combination of keywords and Medical Subject Headings (MeSH) terms related to prevalent non-cancer NCDs, specifically COPD, dementia, diabetes and heart failure, which are known for their significant morbidity and mortality burden and are associated with the need for palliative care.11 13 Additionally, the search incorporated prognostic models or factors, all-cause mortality as an outcome, and palliative or end-of-life care. The search was conducted across five electronic databases, namely Medline, Embase, CINAHL, Cochrane Library and PsycINFO, covering the inception of these databases up until June 2023. The full search strategy is available in online supplemental table 11. Additionally, the reference lists of included studies and other sources such as online theses, Google Scholar, Palliative Care Outcomes Collaboration and Palliative Care Social Work Australia websites were searched to identify potential studies for inclusion in the review.
Supplemental material
The studies identified through the systematic and manual hand search were exported into an EndNote library, and duplicate articles were identified and removed. The unique studies were subsequently imported into the web-based tool, Covidence, to facilitate the team screening process. Two reviewers (BMG and AGM) independently screened all retrieved studies based on title and abstract, using predefined eligibility criteria. Further screening was conducted on the studies included in the full-text review to make the final inclusion decision. Any conflicts or disagreements between the reviewers were resolved through discussion.
Eligibility criteria
All studies that focused on prognostic models to predict all-cause mortality risk or individual prognostic factor studies aimed at identifying predictors of mortality in patients with non-cancer NCDs who had stable visits or appointments with healthcare providers were included. We considered both observational and interventional study designs. There were no restrictions on age or ethnicity. Articles were eligible for inclusion if they provided details on at least one prognostic factor or a set of factors, indices or scores predicting mortality, included all-cause mortality as an outcome measure, and were reported in English. Articles were excluded if they included both cancer and non-cancer patients without separate reporting for the non-cancer subgroup, focused on identifying patients suitable for therapeutic rather than palliative interventions, or did not mention mortality estimates or establish a relationship between the factors, indices or scores and mortality. Additionally, letters to the editor, review articles and abstracts solely presented at seminars or conferences were excluded.
Data extraction
A data extraction form was prepared based on the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis checklist and piloted using five studies. Two reviewers (BMG and AGM) independently extracted all relevant information from the included full-text articles. This included details such as the first author and year of publication, country of study, data source, study design, diagnosis, sample size determination, outcome, the prevalence of an outcome, candidate predictors or prognostic variables, missing data, modelling methods, model performance (including overall performance, calibration and discrimination) and other relevant findings. The data extracted from the included studies were managed using an Excel spreadsheet throughout the review process.
Risk of bias assessment
Two reviewers (BMG and AGM) evaluated the quality of included studies using Quality In Prognosis Studies (QUIPS).36 The QUIPS tool was developed for assessing the methodological quality of prognostic studies; it consists of six domains, including study participation, study attrition, prognostic factor measurement, outcome measurement, study confounding, and statistical analysis and reporting.36 Reviewers rated the prognostic model and prognostic studies as having a low risk of bias (ROB) if no relevant shortcomings were identified (all domains had low ROB), high ROB if at least one domain had high ROB, and unclear ROB if unclear ROB was noted in at least one domain and all other domains had low ROB. The rating (low, high and unclear) of quality was judged by the reviewers, and any conflict was resolved by discussion.
Synthesis methods
A narrative synthesis was conducted using thematic and contextual analysis to summarise the findings based on the identified themes of NCDs. The synthesis encompassed the methodologies applied to develop and validate the prognostic models, including study design, sample size, data source, outcome measurement, candidate predictors included in the models, missing data, modelling methods and model performance. For prognostic factor studies, an overview of the type of predictors that have been investigated thus far for predicting all-cause mortality within the context of palliative care was provided. Tables and graphs were used as appropriate to present the data effectively. Due to the heterogeneity in the methodological approaches used to develop and validate the models, a meta-analysis was not conducted.
Results
Study selection and characteristics
A total of 8369 studies were identified through electronic database searching. Additionally, 57 studies were found from other sources. After removing 1051 duplicates, 7375 studies were screened based on title and abstract. The assessment of the titles and abstracts for relevance resulted in the exclusion of 7253 studies, and 122 studies were further assessed against the inclusion and exclusion criteria. Following a detailed evaluation, 81 studies were excluded. The main reasons for the exclusion of these studies were: no outcome of interest, the study population was not relevant, and not reporting the proportion or risk estimate for the subgroup of patients with non-cancer NCDs. As a result, a total of 41 studies were included in this systematic review. The systematic review process is presented in the flow diagram depicted in figure 1.
Preferred Reporting Items for Systematic Reviews and Meta-Analyses flow diagram illustrating the study selection.
32 of the 41 studies reported prognostic models or predictors of mortality risk in specific disease conditions, such as COPD (n=16),33 37–51 dementia (n=10)30 52–60 and heart failure (n=6).61–66 The remaining nine studies reported prognostic models or predictors of mortality risk in a mixed group of NCDs, including diabetes, cardiovascular, respiratory, renal, liver, neurological and nutrition-related diseases.31 67–74 Nearly half of the included studies were conducted in three countries: the USA (n=14, 34.1%), Spain (n=5, 12.2%) and the UK (n=4, 9.8%). Interestingly, the majority of studies (n=22, 53.7%) were conducted since 2016.
Prognostic models and factors for patients with COPD
A total of 16 studies were identified, which focused on prognostic models and factors that can predict mortality in patients with COPD. Among these, eight studies reported on the development of new prognostic models, and external validation and application of existing prognostic models.33 37–43 The remaining eight studies examined prognostic factors, specifically exploring the association between individual factors and mortality risk.44–51
A total of 16 distinct prognostic models were identified from the 8 studies focusing on prognostic models for patients with COPD. Among these models, four were evaluated in multiple studies, namely ADO (n=4), BODE (n=3), BODEX (n=2) and DOSE (n=2). The sample sizes of the development and external validation of cohorts varied, ranging from 93 to 54 990 participants. The prevalence of the outcome (number of participants with the outcome) ranged from 28 to 11 775 (online supplemental table 1).
In four studies, the included models exclusively examined all-cause mortality within a 12-month timeframe,37 39–41 while the models in the remaining four studies went beyond 12 months.33 38 42 43 The number of risk predictor variables included in the models ranged from 2 to 18. The most frequently included predictor variables in the final models to predict all-cause mortality were dyspnoea (n=11), forced expiratory volume in 1 s (FEV1) (n=9), functional status (n=7), the number of exacerbations in the previous year (n=7) and body mass index (BMI) (n=6) (figure 2). Furthermore, four models incorporated comorbidities such as asthma, atrial fibrillation, stroke, chronic kidney disease, dementia, lung cancer and lung fibrosis. In one model, laboratory variables including serum creatinine, haemoglobin and platelets were included as predictor variables. A comprehensive list of prognostic determinants included in the final models is presented in online supplemental table 2.
Commonly used predictors for developing mortality prognostic models in patients with chronic obstructive pulmonary diseases. Laboratory variables (serum creatinine, haemoglobin and platelets); comorbidities: asthma, atrial fibrillation, stroke, chronic kidney disease, dementia, lung cancer and lung fibrosis. BMI, body mass index; FEV1, forced expiratory volume in 1 s; HRQoL, Health Related Quality of Life.
Traditional statistical models, such as logistic regression and Cox regression, were commonly employed for the development and validation of all prognostic models. In model development studies that reported predictive performance, the reported area under the curve (AUC) or C-statistic for 12-month all-cause mortality ranged from 0.68 to 0.83; only two models underwent internal validation. Among the identified prognostic models, half of them (n=8, 50%) had at least one external validation, with their discrimination being slightly lower than the original results. The reported AUC or C-statistic for 12-month all-cause mortality in the external validation studies ranged from 0.63 to 0.84. Model calibration was assessed in some prognostic models (n=7) using methods such as the Hosmer-Lemeshow test, calibration slope and calibration plot. Most importantly, the majority of the prognostic models were presented using various formats, including risk prediction equations, nomogram, risk score and web application. The predictive performance of each prognostic model is summarised in online supplemental table 2.
In addition to the prognostic model studies, we identified eight additional studies that examined the association between individual predictor factors and mortality in patients with COPD.44–51 These studies revealed several independent predictor variables linked to short-term (within 12 months) and long-term (between 1 and 5 years) mortality risks. Functional performance, as measured by 2 and 6 min walking distances or a decrease in gait speed by 0.14 m/s,44 45 51 as well as dyspnoea during activities,46 a shift towards perceiving a very sedentary state, and feelings of upset or being downhearted, were identified as independent predictors of death.44 Additionally, having comorbid conditions such as asthma and being underweight,47 50 elevated serum C reactive protein levels,48 prebronchodilator and postbronchodilator lung function,49 a history of falls in the 6 months preceding hospital admission,46 and sociodemographic variables, including smoking and white ethnicity,50 were also significantly associated with mortality risk.
Prognostic models and factors for patients with dementia
A total of 10 studies were identified, which focused on prognostic models and factors for predicting mortality in patients with advanced dementia. Among these, six studies reported the development and validation of prognostic models specifically developed to assess 6-month to 2-year mortality risks.52 53 55 56 58 59 On the other hand, the remaining four studies examined prognostic factors and their association with mortality in patients with advanced dementia.30 54 57 60
A total of six distinct prognostic models were identified in the studies focusing on prognostic models for patients with dementia. All models were developed and validated using a cohort design, with the sample sizes ranging from 269 to 37 289 participants. The prevalence of the outcome (number of participants with the outcome) varied between 70 and 20 542 (online supplemental table 3). The number of risk predictor variables included in the models ranged from 6 to 14. Among the diverse range of predictors included, the most common variables found in the final models to predict all-cause mortality were comorbidity (n=5), age (n=4), gender (n=3) and insufficient oral intake (n=3) (figure 3). Comorbid conditions such as stroke, chronic kidney disease, cancer, liver cirrhosis, heart failure, arthritis and pressure ulcers were included in five of the models that considered comorbidity as a predictor. In one model, infectious diseases such as aspiration pneumonia, septicaemia, pyelonephritis and other upper urinary tract infections were included as predictor variables. A comprehensive list of predictors included in the final models is presented in online supplemental table 4.
Commonly used predictors for developing mortality prognostic models in patients with dementia. BMI, body mass index; FAST, Functional Assessment Staging.
Traditional statistical models (primarily logistic regression and Cox regression) were applied to develop and validate nearly all of the prognostic models. In the model development studies that reported predictive performance, the reported AUC or C-statistic for 12-month all-cause mortality ranged from 0.65 to 0.956, with four models undergoing internal validation (online supplemental table 4). Among the identified prognostic models, only three of them underwent external validation; the reported AUC or C-statistic for 6-month all-cause mortality ranged from 0.55 to 0.67. Calibration was assessed in a few prognostic models (n=3) using methods such as the Hosmer-Lemeshow test, calibration plot and calibration table. The majority of prognostic models were presented through the use of a risk score. The predictive performance of each prognostic model is summarised in online supplemental table 4.
In addition to the prognostic model studies, we identified four additional studies that examined the association between individual predictor factors and mortality within a 6-month timeframe in patients with dementia.30 54 57 60 In two of the studies, a high level of suffering as measured by the Mini-Suffering State Examination scale was associated with 6-month mortality.30 60 In the remaining two studies, factors such as age, anorexia, and the presence of both anorexia and higher functional impairment were statistically significantly associated with an increased risk of mortality.54 57 One of the studies further explored the impact of reaching stage 7 on the FAST scale as a predictor of survival time. The results revealed that patients who reached stage 7 on the FAST scale had a mean survival time of only 3.2 months, which was significantly shorter compared with the 18-month survival time of those who did not reach that stage.54
Prognostic models and factors for patients with heart failure
A total of six studies were identified focusing on prognostic models and factors predicting mortality in patients with advanced heart failure. These studies reported the development and validation of prognostic models within various timeframes, ranging from 7 days to 12 months.61–66 From these studies, a total of eight distinct multivariable prognostic models were identified. All prognostic models were developed and validated using randomised trial and cohort designs, with sample sizes ranging from 282 to 7508 participants. The prevalence of the outcome (number of participants with the outcome) varied from 43 to 1825 (online supplemental table 5). The number of risk predictor variables included in the models ranged from 2 to 36. Despite the wide range of predictors included, the most common variables for predicting all-cause mortality were comorbidity (n=6), blood tests (n=6), age (n=4) and gender (n=4) (figure 4). The identified prognostic models considering comorbidity as a predictor incorporated conditions such as cancer, atrial fibrillation, peripheral arterial diseases, liver disease, chronic skin ulcers and neurological disorders. Similarly, the identified prognostic models incorporating blood test results as predictors included blood urea nitrogen, complete blood count parameters (white blood cell count, haematocrit, red cell distribution width, platelet count), electrolytes (serum sodium and potassium), total cholesterol, albumin, lactate, troponin and uric acid levels. A comprehensive list of predictors included in the final models is presented in online supplemental table 6.
Commonly used predictors for developing mortality prognostic models in patients with heart failure. CRT, cardiac resynchronisation therapy; GFR, glomerular filtration rate; ICD, implantable cardioverter defibrillator; LVEF, left ventricular ejection fraction; NYHA, New York Heart Association.
Traditional statistical models, specifically logistic regression and Cox regression were applied to develop and validate nearly all of the prognostic models. In the model development studies that reported predictive performance, the reported AUC or C-statistic for 6-month and 12-month all-cause mortality ranged from 0.71 to 0.80 and 0.631 to 0.740, respectively, with only two models undergoing external validation (online supplemental table 6). One model reported an AUC or C-statistic of 0.701 for predicting 6-month all-cause mortality. The second model, which predicted mortality risk at various time points (ranging from 7 days to 6 months), demonstrated an AUC or C-statistic ranging from 0.689 to 0.802. Calibration was reported for few prognostic models (n=2) using the Hosmer-Lemeshow test. Importantly, some of the prognostic models provided a risk score as a model presentation. The predictive performance of each prognostic model is summarised in online supplemental table 6.
Prognostic models and factors for patients with mixed chronic diseases
A total of nine studies were identified that focused on prognostic models and factors predicting mortality in patients with mixed chronic diseases, including chronic lung, heart, liver, renal, endocrine and neurological diseases (online supplemental table 7). Among these studies, six reported the development and validation of prognostic models for predicting mortality risks within a timeframe of 1 day to 24 months.31 68 69 71 74 75 The remaining three studies were prognostic factor studies aimed at examining the association between individual factors and mortality risk.67 70 73 A total of five distinct multivariable prognostic models were identified from the prognostic model studies. All prognostic models were developed and validated using a cohort design, with sample sizes ranging from 100 to 1778. The prevalence of the outcome (number of participants with the outcome) ranged from 201 to 422 (online supplemental table 7). The number of risk predictor variables included in the models ranged from four to seven. Despite the wide range of predictors included in the models, the most common predictor variables for predicting all-cause mortality were functional status (n=3), age (n=2), dyspnoea (n=2), presence of skin pressure ulcer (n=2), oral intake (n=2) and level of consciousness (n=2) (figure 5). A comprehensive list of predictors included in the final models is presented in online supplemental table 8.
Commonly used predictors for developing mortality prognostic models in patients with mixed chronic diseases.
Traditional statistical models, specifically logistic regression and Cox regression were applied to develop and validate all of the prognostic models. In the model development studies that reported predictive performance, the AUC or C-statistic for 6-month and 12-month all-cause mortality ranged from 0.710 to 0.786 and 0.693 to 0.751, respectively. Only three models underwent external validation (online supplemental table 8). Two models reported an AUC or C-statistic of 0.70 and 0.943 for predicting 6-month all-cause mortality. The third model, which predicted mortality risk at various time points (ranging from 1 day to 3 months), demonstrated an AUC ranging from 0.68 to 0.85. Calibration was reported for three prognostic models using the Hosmer-Lemeshow test, calibration plot, Brier score and expected-to-observed risk ratio. Importantly, the model presentation was reported for only two prognostic models using a risk score and nomogram. The predictive performance of each prognostic model is summarised in online supplemental table 8.
In addition to the prognostic model studies, we identified three additional studies that examined the association between individual predictor factors and mortality within a timeframe of 1 week to 12 months in patients with mixed chronic diseases.67 70 73 The presence of comorbid conditions, such as solid organ malignancy, severely impaired kidney function, severe dementia with complete or severe physical dependence, and hypoalbuminaemia, was statistically significantly associated with increased in-hospital and 12-month mortality.67 73 Furthermore, the findings indicated that being bedridden, having moderate to severe weakness as measured by the Karnofsky performance score, and having been hospitalised within the past 6 months were also significantly associated with increased mortality risk within 3 months.67 70
ROB in studies
Only 3 out of the 26 prognostic model studies demonstrated an overall low ROB. The domains of participant selection, assessment of attrition and outcome assessment were the primary factors contributing to a low ROB in most studies. Importantly, none of the studies exhibited a high ROB in the outcome assessment domain. The high ROB in participant selection primarily resulted from unclear participant identification or diagnosis and ambiguous eligibility criteria. Likewise, the high ROB in the assessment of attrition was attributed to insufficient or no information on follow-up and missing data. The majority of prognostic model studies had a moderate ROB in predictor assessment, with no studies exhibiting a high ROB. This moderate risk was due to retrospective assessment of predictors after knowledge of the outcome (mortality) or a lack of clarity regarding assessor blinding. In terms of analysis, most studies displayed a high ROB. Some studies mishandled continuous variables, despite recommendations favouring the use of continuous variables without categorisation in prognostic model development. Variables such as age, functional status score (ADL score), comorbidity index and the number of hospitalisations were incorrectly categorised in certain models. Additionally, excluded participants and those with missing data were inadequately addressed or not reported. Univariate analysis-based predictor selection, appropriate evaluation of relevant model performance measures, and accounting for model overfitting and optimism were also not appropriately addressed or reported in the majority of models (online supplemental tables 9 and 10).
Discussion
This systematic review comprehensively investigated prognostic factors and models for identifying end-of-life in patients with non-cancer NCDs, including COPD, dementia, heart failure and mixed NCDs. The review analysed 41 relevant studies, with 16 focusing on COPD,33 37–51 10 on dementia,30 52–60 6 on heart failure61–66 and 9 on mixed NCDs.31 67–74 These studies examined all-cause mortality across various timeframes, ranging from short-term (within a week) to long-term (up to 10 years). Among the identified models, a wide range of predictors was identified. Dyspnoea, FEV1, functional status, exacerbation history and BMI consistently emerged as frequent predictors in COPD models. Models for dementia and heart failure consistently incorporated predictors such as comorbidity, age, gender, blood tests and nutritional status. Similarly, mixed NCD models consistently included predictors such as functional status, age, dyspnoea, the presence of skin pressure ulcers, oral intake and level of consciousness. This has indicated many of the same predictors are combined in different ways in different multivariable models, which implies a degree of consensus among investigators as to the important prognostic factors.
The findings of this study revealed that the identified prognostic models exhibited variable predictive accuracy, with the majority demonstrating weak to moderate (AUC: 0.5–0.8) discriminatory performance. This variability and lower performance can be attributed to several factors. First, the complexity of non-cancer NCDs poses challenges in identifying and incorporating all relevant prognostic factors.8 11 76 These diseases involve a complex interplay between various clinical, physiological and psychosocial factors that contribute to disease progression and mortality risk.8 11 76 The identification and measurement of these factors can be subjective and prone to variability across studies.20 77 Moreover, the use of certain prognostic models in specific patient groups may lead to biased risk estimates. For example, the application of the Seattle Heart Failure Model in very old patients often results in a significant overestimation of prognosis.78 Another possible explanation for the variable performance of the models is the methodological limitations in the included studies, such as issues related to data quality and availability, patient heterogeneity, analytical approaches employed and sample size. These issues were evident in the overall quality of evidence, which were low for the majority of the studies. Another interesting finding of this study was that prognostic models primarily developed using disease-specific predictors, which directly relate to the pathophysiology and progression of the diseases, did not demonstrate superior predictive performance compared with models incorporating non-specific predictors. Disease-specific predictors, such as lung function, dyspnoea severity and exacerbation history for COPD; left ventricle ejection fraction and New York Heart Association Class for heart failure; and cognitive status and neuropsychiatric symptoms for dementia, have often been highlighted in the included studies as important markers for predicting end-of-life outcomes.11 79 80 However, our findings suggest that these disease-specific factors may not provide substantial added value in prognostic modelling when compared with more general or non-specific predictors like age, functional status, nutrition status or comorbidities.
One possible explanation for the lack of superior performance by models mainly developed using disease-specific predictors is the complex and multifactorial nature of non-cancer NCDs.8 11 76 81 82 These conditions are often characterised by a wide range of comorbidities and overlapping symptoms, making it challenging to identify disease-specific markers that accurately predict end-of-life outcomes.8 11 76 81 82 Moreover, non-cancer NCDs frequently progress at different rates and trajectories, further complicating the identification of disease-specific prognostic factors.8 11 76 81 82 Consequently, the inclusion of more generic predictors that capture the overall health status and functional decline of patients may offer a more comprehensive approach to prognostic modelling.81 82 In this regard, our findings align with evidence from other areas of medicine, which also suggests that models with non-specific predictors effectively predicted future outcomes.71 83 84 For instance, studies examining end-of-life predictions in patients with cancer have shown that generic measures, such as performance status and functional decline, are powerful predictors of survival and treatment outcomes.71 83 84 Similarly, in the field of geriatrics, non-specific predictors like frailty and cognitive impairment have been found to be valuable indicators of prognosis in a variety of conditions.85 86 Overall, these findings emphasise the significance of adopting a comprehensive approach through the integration of both generic and disease-specific measures within a single model, as this approach optimises the performance of prognostic models for non-cancer NCDs.
The overarching unreliability of prognostic tools to date, particularly their inability to provide consistent and accurate predictions in various non-cancer NCD contexts, raises the possibility that accurate prognostication at the individual level may never be possible due to disease heterogeneity. As such, the search for precise prognostic information may be an elusive journey, hindering clinicians from providing personalised and comprehensive patient care, specifically within the context of palliative care.87 88 This pursuit often leads to ‘prognostic paralysis’, a state in which the search for precise prognostic information causes delays and uncertainties in clinical decision-making.88 89 This fixation not only detracts from focusing on the immediate and unique needs of each patient but also potentially delays the implementation of palliative measures designed to enhance quality of life and alleviate suffering.88 89 This scenario underscores the limitations of relying solely on prognostic information for clinical decision-making and indicates a need to integrate other more dynamic, patient-centred approaches.88 89 These approaches may prioritise the patient’s current condition, preferences, and values over the uncertain and often elusive precision of prognostic information. By doing so, care can become more responsive and tailored, ensuring that it aligns with the fundamental goals of palliative care—to comfort and support patients through a compassionate understanding of their individual journey, rather than being constrained by the limitations of prognostic accuracy.88 89 It may eventuate that these models may be more useful at the population level for helping in resource planning and allocation by policy makers.
Strength and limitations
This is the first comprehensive systematic review that included relevant studies on prognostic models and factors for identifying end-of-life in patients with non-cancer NCDs. The objective of the review and transparent methodology, adhering to a predefined protocol and transparent criteria, ensures a rigorous and unbiased approach to synthesising the evidence. The main challenges in obtaining information from this review primarily stem from the poor quality of the included studies. However, to gain a comprehensive understanding of the strengths and weaknesses of existing models, it is crucial to thoroughly assess their overall quality and characteristics. This assessment is important as it serves as a foundation for the development of better models in the future.
Major limitations of the included studies are the lack of adherence to reporting guidelines for risk prediction modelling and the lack of external validation. Another methodological limitation was the high ROB, which was evident in the majority of studies. These methodological limitations included the use of uncertain inclusion or exclusion criteria, inadequate sample sizes, lack of reporting and/or handling of missing data, inappropriate handling of continuous and categorical variables, failure to evaluate or report relevant model performance measures, failure to consider model overfitting and optimism in model performance assessment, lack of internal and external validation, and lack of providing model presentation formats. It is essential to acknowledge and take into account these limitations when interpreting the findings.
Implications
To advance prognostication and patient care for individuals with non-cancer NCDs, future directions should focus on utilising robust methods for developing, validating and updating prognostic models. The integration of high-quality data from large datasets, including pooled patient data and electronic health records, is essential for the development of robust and cost-effective prognostic models that can be applied to real-world populations. Given the inherent prognostic uncertainty and unpredictable disease trajectory in non-cancer NCDs, it is crucial to incorporate generic predictors that capture the complex interplay between various clinical, physiological and psychosocial factors. These may include age, comorbidities, functional status, nutrition status and measures of frailty, which can improve the precision of risk estimates. Furthermore, the inclusion of disease-specific predictors, often overlooked or under-represented, that directly relate to the underlying pathophysiology and disease progression, can improve the accuracy of prognostic models. Moreover, optimal prognostic models should be user-friendly, cost-effective and incorporate readily available prognostic factors. By developing better prognostic tools, systematic identification of patients who may benefit from interventions such as palliative care or advance care planning can be achieved, ultimately improving end-of-life patient outcomes in the care of non-cancer NCDs. We also need to face the possibility that disease heterogeneity will limit the accuracy of any prognostic model at the individual level and that these models will be most useful at the population level for policy makers and health managers.
Conclusions
Patients with non-cancer NCDs often face complex and unpredictable disease trajectories, making it challenging to determine the appropriate timing for palliative care intervention. Accurate prognostic information, if available, is very helpful for guiding healthcare professionals in identifying those who may benefit most from early palliative care interventions, thereby avoiding unnecessary aggressive treatments and enhancing the patient’s overall comfort and well-being. However, our systematic review identified a wide range of predictors for all-cause mortality across the identified non-cancer NCDs. The existing prognostic models have variable predictive accuracy, with the majority demonstrating weak to moderate discriminatory performance. Moreover, only a few were internally validated and most lacked independent external validation. Additionally, we identified a high ROB, with several studies failing to report model performance measures and model presentation formats. Given these findings, exploring alternative approaches for end-of-life identification and treatment guidance may substantially improve the practical management and care of these patients. Future studies should focus on developing robust, high-quality prognostic models that incorporate variables explaining the complex interplay between various clinical, physiological and psychosocial factors to improve prediction accuracy. The inclusion of easily accessible and cost-effective prognostic determinants enhances the practical application of the models. Independent external validation is also critical before implementing these prognostic models into clinical practice, or in policy and resource allocation at a population level.
Data availability statement
All data relevant to the study are included in the article or uploaded as supplementary information.
Ethics statements
Patient consent for publication
Ethics approval
Not applicable.
Acknowledgments
We would like to acknowledge the support provided to BMG through the University of Newcastle Vice-Chancellor’s Higher Degree by Research Training Scholarship. Additionally, we would like to acknowledge Jessica Birchall, a librarian at the College of Health, Medicine, and Wellbeing, University of Newcastle, for her assistance in refining the database search terms.
References
Footnotes
Contributors BMG, MLH and JRA conceptualised the study and developed the search strategy. BMG and AGM screened citations, extracted data and carried out the quality assessment. BMG conducted data synthesis and drafted the manuscript. MLH and JRA reviewed and edited the manuscript for its contributions to the field. All authors have read and approved the final manuscript. BMG is the guarantor.
Funding The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors.
Competing interests None declared.
Provenance and peer review Not commissioned; externally peer reviewed.
Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.