Introduction
Cancer patients and their families seek prognostic information to guide decision-making and emotionally prepare for end-of-life (Chu et al. Reference Chu, Anderson and White2020). Although physician survival prediction is widely utilized, it could be unreliable and unduly optimistic (Chu et al. Reference Chu, Anderson and White2020). To qualify for specialized care and guide treatment decisions, an accurate prognosis is necessary (Chu et al. Reference Chu, White and Stone2019; Kutzko et al. Reference Kutzko, Dadwal and Holt2022). Tools like the Palliative Prognostic Index (PPI) offer standardized estimates to address the limitations of clinician prediction. Other validated tools for advanced cancer patients include the Palliative Prognostic Score (Yoong et al. Reference Yoong, Bhowmik and Kapparath2024), the Suprise Question (van Lummel et al. Reference van Lummel, Ietswaard and Zuithoff2022), and the Prognosis in Palliative Care tool and the Objective Prognostic Score (Lee et al. Reference Lee, Lee and Choi2021).
The European Association of Palliative Care (Maltoni et al. Reference Maltoni, Caraceni and Brunelli2005) and the European Society for Medical Oncology identified the PPI as a key tool for predicting survival in advanced cancer patients (Stone et al. Reference Stone, Buckle and Dolan2023). Developing using data from a Japanese inpatient hospice (Morita et al. Reference Morita, Tsunoda and Inoue1999), the PPI score ranges from 0 to 15 and includes assessments of the Palliative Performance Scale, edema, dyspnea, and delirium (Morita et al. Reference Morita, Tsunoda and Inoue2001), with higher scores indicating shorter survival.
The PPI has been validated in various cancer settings, such as hospices (Kim et al. Reference Kim, Youn and Ko2014; Subramaniam et al. Reference Subramaniam, Thorns and Ridout2013), palliative care units (Gerber et al. Reference Gerber, Tuer and Yates2021; Miyagi et al. Reference Miyagi, Miyata and Tagami2021), community (Hamano et al. Reference Hamano, Kizawa and Maeno2014), and hematology wards (Lee et al. Reference Lee, Chou and Yang2022; Ohno et al. Reference Ohno, Abe and Sasaki2017). Palliative care nurses in the community hospitals easily integrated it into admission routines (Belanger et al. Reference Belanger, Tetrault and Tradounsky2015). A web-based prognostic calculator that included PPI increased doctors’ confidence and willingness to discuss prognosis with patients and ability to tailor treatments according to prognosis (Hui et al. Reference Hui, Maxwell and de la Rosa2024). Additionally, healthcare professionals in aged care teams found it easy to use and not burdensome, with most recommending it to colleagues (Gerber et al. Reference Gerber, Bloomer and Hayes2023). The PPI was particularly useful for uncertain prognoses, promoting end-of-life discussions and early recognition of dying. However, its challenges included distinguishing between acute and terminal delirium and when edema should be rated as present (Gerber et al. Reference Gerber, Bloomer and Hayes2023).
The original study’s survival analysis divided patients into 3 groups: PPI ≤ 2, 2 < PPI ≤ 4, and PPI > 4. Log-rank analyses showed that PPI could differentiate survival across these groups (Morita et al. Reference Morita, Tsunoda and Inoue1999). Validation studies typically presented log-rank tests and Kaplan–Meier curves but not hazard ratios (HR), odds ratios (OR), or risk ratios (RR). While Kaplan–Meier curves reveal crude survival differences among risk groups, they lack effect measures with 95% confidence intervals (CI) that adjust for other variables (Stel et al. Reference Stel, Dekker and Tripepi2011). Furthermore, validation studies did not always adhere to the original model’s risk group definitions, which may account for instances where survival differences were not significant (Palomar-Muñoz et al. Reference Palomar-Muñoz, Martín-Zamorano and Mogollo2018; Trejo-Ayala et al. Reference Trejo-Ayala, Ramos-Peñafiel and Santoyo-Sánchez2018; Yoon et al. Reference Yoon, Jung and Kim2014).
The only review on the prognostic utility of PPI pooled HR and did not differentiate between adjusted and unadjusted effect sizes (Liu et al. Reference Liu, Su and Wang2018), making it difficult to confirm an independent association between PPI scores and survival. A previous review on prognostic tools, including PPI, also highlighted inconsistent reporting of HR and 95% CI among the studies, preventing a meta-analysis (Simmons et al. Reference Simmons, McMillan and McWilliams2017). We previously conducted a meta-analysis evaluating the PPI’s performance in terms of discrimination and calibration for predicting cancer patients’ survival (Yoong et al. Reference Yoong, Porock and Whitty2023). Building on the previous review’s findings, this systematic review and meta-analysis aimed to evaluate the utility of PPI as a prognostic tool for advanced cancer patients (i.e. locally advanced, metastatic, or incurable cancers). This review focuses on advanced cancer patients who face an increased need to plan for end-of-life decisions, including treatment, palliation and personal matters. Compared to other predictive tools, the PPI offers a simple, standardized assessment that is easy for clinicians to use without extensive training or complex technology. Its evidence-based scoring system ensures quick and effective assessments. The findings from this review aim to provide clinicians with the best information to support patients and their families.
Methods
This review was reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (Table S1) (Page et al. Reference Page, McKenzie and Bossuyt2021). Its protocol was registered in PROSPERO (CRD42023475009).
Eligibility criteria
The inclusion criteria were as follows: (1) adults (≥18 years old) with advanced cancer of any type or those receiving palliative care; (2) studies reporting the association of PPI with death (HR, OR, RR, and 95% CI, including both adjusted or unadjusted effect sizes); (3) studies conducted in any setting, at any time and using any PPI cutoffs; (4) both prospective or retrospective studies (including peer-reviewed articles, dissertations/theses, and preprints); and (5) studies published in English, as the authors are fluent in English only.
Studies were excluded if they involved (1) adults without cancer (unless the noncancer participants were few, and ≥80% of the participants had cancer); (2) other versions of PPI, such as Functional PPI; (3) study designs other than those specified (e.g. experimental studies, reviews, letters to the editor); or (4) studies that only presented Kaplan–Meier curves, log-rank ratios or other descriptive analyses without reporting effect sizes.
Search strategy
We searched PubMed, ScienceDirect, Embase, Web of Science, CINAHL, ProQuest, and Google Scholar for relevant articles published from inception to 16 February 2024 (Tables S2–S7) and reviewed the reference lists of relevant studies and reviews. First, we searched PubMed using keywords and Medical Subject Headings such as “palliative prognostic index,” “palliative care,” and “cancer.” Second, other databases were searched with similar terms. Finally, Google Scholar and ProQuest were used to locate grey literature. The initial search results were uploaded to Rayyan, and after removing duplicates, SQY and DW identified potential studies by reviewing titles and abstracts. They independently assessed full-text articles for eligibility, with any discrepancies resolved by HZ.
Data extraction
Five studies were used to design and pilot test a standardized data extraction form. SQY extracted the data, which was then verified by DW and HZ. The extracted data included authors, country, study design, participant characteristics, and prognostic effect measures (e.g. HR, RR, OR). Any disagreements were resolved through discussion until a consensus was reached.
Quality appraisal
The risk of bias was assessed independently by SQY and DP using the Quality in Prognosis Studies tool, which evaluates 6 domains: (1) study participation, (2) study attrition, (3) prognostic factor measurement, (4) outcome measurement, (5) study confounding, and (6) statistical analysis and reporting. Each domain was rated as high, moderate, or low risk of bias (Grooten et al. Reference Grooten, Tseli and Äng2019; Hayden et al. Reference Hayden, van der Windt and Cartwright2013). A study was considered “low risk of bias” if all 6 domains, or 1 moderate domain, showed low bias. It was considered “high risk of bias” if at least 1 domain was rated high or 3 domains were rated moderate. Studies with intermediate ratings were classified as “moderate risk of bias” (Grooten et al. Reference Grooten, Tseli and Äng2019). Any discrepancies were resolved through discussion. Figures were generated using robvis (McGuinness and Higgins Reference McGuinness and Higgins2021).
The modified Grading of Recommendations, Assessment, Development, and Evaluations (GRADE) framework for prognostic factor reviews was used to assess the overall certainty of evidence (Huguet et al. Reference Huguet, Hayden and Stinson2013). It evaluated 6 factors: investigation phase, study limitations, inconsistency, indirectness, imprecision, and publication bias. Evidence with a moderate or large effect size, or an exposure-response gradient, could lead to an upgrade in the quality of evidence (Huguet et al. Reference Huguet, Hayden and Stinson2013). Studies with Phase 3 explanatory outcomes were initially rated as high-quality evidence (Huguet et al. Reference Huguet, Hayden and Stinson2013; Kent et al. Reference Kent, Cancelliere and Boyle2020). Outcomes based on at least 2 studies included in the meta-analyses were rated as high, moderate, low, or very low quality. Justifications were provided in the “Evidence Profile” tables using the GRADEproGDT software (GRADE handbook 2013; McMaster University and Evidence Prime Inc 2022).
Data analysis
Meta-analysis was conducted using restricted maximum likelihood in JASP (version 0.19.1) (JASP Team 2024). Cochran’s Q test and I 2 statistic were used to assess heterogeneity, with statistical significance set at p < 0.10. Heterogeneity was classified as unimportant (I 2 = 0–40%), moderate (I 2 = 30–60%), substantial (I 2 = 50–90%), or considerable (I 2 = 75–100%) (Higgins et al. Reference Higgins, Thomas and Chandler2019). Following Riley et al. (Reference Riley, Moons and Snell2019), we pooled adjusted and unadjusted effect measures, grouped similar categories of effect measures, and treated continuous effect measures separately. Extracted outcomes were standardized and reclassified into “high” versus “low” PPI risk groups, with effect sizes representing the risk of death as positive numbers. When meta-analysis was not feasible, results were summarized narratively.
Results
Search results
Figure 1 illustrates the study selection process. The initial search identified 946 records. After removing duplicates, 720 records were screened by title and abstract, and 74 articles were further assessed through full-texts review. Ultimately, 23 articles from 21 patient cohorts were included in this systematic review. Reasons for exclusion are detailed in Table S8.
Figure 1. PRISMA diagram showing the study selection process.
Characteristics of included studies
Characteristics of the included studies are presented in Table 1. Studies were published between 2008 and 2023, using prospective (n = 9) (Chen et al. Reference Chen, Mauro and Gabrielli2018; Fernandes et al. Reference Fernandes, Branco and Fernandez2021; Hung et al. Reference Hung, Wang and Kao2014; Kao et al. Reference Kao, Hung and Wang2014; Lee et al. Reference Lee, Lee and Yang2014; Miura et al. Reference Miura, Matsumoto and Hama2015; Palomar-Muñoz et al. Reference Palomar-Muñoz, Martín-Zamorano and Mogollo2018; Stone et al. Reference Stone, Tiernan and Dooley2008; Subramaniam et al. Reference Subramaniam, Thorns and Ridout2013) or retrospective designs (n = 14) (Ahn et al. Reference Ahn, Ahn and Park2021, Reference Ahn, Hwang and Lee2016; Al-Ansari et al. Reference Al-Ansari, Abd-El-Gawad and AboSerea2022; Arai et al. Reference Arai, Okajima and Kotani2014; Arkın and Aras Reference Arkın and Aras2021; Chang et al. Reference Chang, Lee and Huang2021; Cheng et al. Reference Cheng, Kao and Hung2012; Chou et al. Reference Chou, Kao and Wang2015; Gerber et al. Reference Gerber, Tuer and Yates2021; Iizuka-Honma et al. Reference Iizuka-Honma, Mitsumori and Yoshikawa2023; Inomata et al. Reference Inomata, Hayashi and Tokui2014; Kiuchi et al. Reference Kiuchi, Hasegawa and Watanabe2022; Shatri et al. Reference Shatri, Prasetyaningtyas and Putranto2021; Trejo-Ayala et al. Reference Trejo-Ayala, Ramos-Peñafiel and Santoyo-Sánchez2018).
Table 1. Characteristics of included studies
SD = standard deviation; IQR = interquartile range;, ✓ = Yes; ? = unclear; ✘ = No.
The majority of studies were conducted in Asia and Australia (n = 17) (Ahn et al. Reference Ahn, Ahn and Park2021, Reference Ahn, Hwang and Lee2016; Al-Ansari et al. Reference Al-Ansari, Abd-El-Gawad and AboSerea2022; Arai et al. Reference Arai, Okajima and Kotani2014; Arkın and Aras Reference Arkın and Aras2021; Chang et al. Reference Chang, Lee and Huang2021; Cheng et al. Reference Cheng, Kao and Hung2012; Chou et al. Reference Chou, Kao and Wang2015; Gerber et al. Reference Gerber, Tuer and Yates2021; Hung et al. Reference Hung, Wang and Kao2014; Iizuka-Honma et al. Reference Iizuka-Honma, Mitsumori and Yoshikawa2023; Inomata et al. Reference Inomata, Hayashi and Tokui2014; Kao et al. Reference Kao, Hung and Wang2014; Kiuchi et al. Reference Kiuchi, Hasegawa and Watanabe2022; Lee et al. Reference Lee, Lee and Yang2014; Miura et al. Reference Miura, Matsumoto and Hama2015; Shatri et al. Reference Shatri, Prasetyaningtyas and Putranto2021), followed by Europe (n = 3) (Palomar-Muñoz et al. Reference Palomar-Muñoz, Martín-Zamorano and Mogollo2018; Stone et al. Reference Stone, Tiernan and Dooley2008; Subramaniam et al. Reference Subramaniam, Thorns and Ridout2013) and the Americas (n = 3) (Chen et al. Reference Chen, Mauro and Gabrielli2018; Fernandes et al. Reference Fernandes, Branco and Fernandez2021; Trejo-Ayala et al. Reference Trejo-Ayala, Ramos-Peñafiel and Santoyo-Sánchez2018).
The review included 11,235 patients aged 18–100, with sample sizes ranging from 28 to 4,685. Most studies involved a mix of primary cancers, while 7 studies focused on a single cancer type (Arkın and Aras Reference Arkın and Aras2021; Chang et al. Reference Chang, Lee and Huang2021; Chou et al. Reference Chou, Kao and Wang2015; Iizuka-Honma et al. Reference Iizuka-Honma, Mitsumori and Yoshikawa2023; Inomata et al. Reference Inomata, Hayashi and Tokui2014; Kiuchi et al. Reference Kiuchi, Hasegawa and Watanabe2022; Trejo-Ayala et al. Reference Trejo-Ayala, Ramos-Peñafiel and Santoyo-Sánchez2018). The majority of studies were conducted in palliative care settings, with 1 conducted in acute wards (Iizuka-Honma et al. Reference Iizuka-Honma, Mitsumori and Yoshikawa2023).
Twenty studies reported HR, with 12 adjusting for covariates (Ahn et al. Reference Ahn, Ahn and Park2021, Reference Ahn, Hwang and Lee2016; Arai et al. Reference Arai, Okajima and Kotani2014; Chang et al. Reference Chang, Lee and Huang2021; Chou et al. Reference Chou, Kao and Wang2015; Hung et al. Reference Hung, Wang and Kao2014; Inomata et al. Reference Inomata, Hayashi and Tokui2014; Kao et al. Reference Kao, Hung and Wang2014; Kiuchi et al. Reference Kiuchi, Hasegawa and Watanabe2022; Lee et al. Reference Lee, Lee and Yang2014; Miura et al. Reference Miura, Matsumoto and Hama2015; Palomar-Muñoz et al. Reference Palomar-Muñoz, Martín-Zamorano and Mogollo2018). Most studies treated PPI as a categorical variable, while 5 analyzed it as a continuous variable (Arai et al. Reference Arai, Okajima and Kotani2014; Gerber et al. Reference Gerber, Tuer and Yates2021; Lee et al. Reference Lee, Lee and Yang2014; Stone et al. Reference Stone, Tiernan and Dooley2008; Subramaniam et al. Reference Subramaniam, Thorns and Ridout2013). Dichotomous outcomes were extracted or computed from 5 studies (Al-Ansari et al. Reference Al-Ansari, Abd-El-Gawad and AboSerea2022; Arkın and Aras Reference Arkın and Aras2021; Fernandes et al. Reference Fernandes, Branco and Fernandez2021; Gerber et al. Reference Gerber, Tuer and Yates2021; Trejo-Ayala et al. Reference Trejo-Ayala, Ramos-Peñafiel and Santoyo-Sánchez2018), with 2 reporting adjusted effect sizes (Al-Ansari et al. Reference Al-Ansari, Abd-El-Gawad and AboSerea2022; Gerber et al. Reference Gerber, Tuer and Yates2021). The findings from each study are presented in Table 2.
Table 2. Prognostic effect measures reported in the included studies
PPI = Palliative Prognostic Index; HR = hazard ratio; RR = risk ratio; OR = odds ratio;, CI = confidence intervals;, BMI = body mass index;, EGFR = estimated glomerular filtration rate; ECOG PS = Eastern Cooperative Oncology Group Performance Status; CRP = C-reactive protein; NLR = neutrophil-to-lymphocyte ratio; mPaP = modified Palliative Prognostic Score.
Risk of bias assessment
Figure 2 illustrates the risk of bias ratings. Nine studies were classified as having a low risk of bias (Ahn et al. Reference Ahn, Ahn and Park2021; Al-Ansari et al. Reference Al-Ansari, Abd-El-Gawad and AboSerea2022; Arai et al. Reference Arai, Okajima and Kotani2014; Chang et al. Reference Chang, Lee and Huang2021; Chou et al. Reference Chou, Kao and Wang2015; Hung et al. Reference Hung, Wang and Kao2014; Kao et al. Reference Kao, Hung and Wang2014; Lee et al. Reference Lee, Lee and Yang2014; Palomar-Muñoz et al. Reference Palomar-Muñoz, Martín-Zamorano and Mogollo2018), 3 as moderate risk (Ahn et al. Reference Ahn, Hwang and Lee2016; Gerber et al. Reference Gerber, Tuer and Yates2021; Miura et al. Reference Miura, Matsumoto and Hama2015), and 11 as high risk (Arkın and Aras Reference Arkın and Aras2021; Chen et al. Reference Chen, Mauro and Gabrielli2018; Cheng et al. Reference Cheng, Kao and Hung2012; Fernandes et al. Reference Fernandes, Branco and Fernandez2021; Iizuka-Honma et al. Reference Iizuka-Honma, Mitsumori and Yoshikawa2023; Inomata et al. Reference Inomata, Hayashi and Tokui2014; Kiuchi et al. Reference Kiuchi, Hasegawa and Watanabe2022; Shatri et al. Reference Shatri, Prasetyaningtyas and Putranto2021; Stone et al. Reference Stone, Tiernan and Dooley2008; Subramaniam et al. Reference Subramaniam, Thorns and Ridout2013; Trejo-Ayala et al. Reference Trejo-Ayala, Ramos-Peñafiel and Santoyo-Sánchez2018).
Figure 2. (A) Risk of bias ratings for each study and (B) risk of bias summary graph showing the overall distribution of ratings for each domain.
In the study participation domain, most studies reported population characteristics well, although some did not specify the recruitment period or exclusion criteria. Study attrition was low in most studies, but some only analyzed a subset of participants from larger cohorts, potentially limiting the generalizability of outcomes. In certain studies, those lost to follow-up were excluded, leading to unclear attrition rates. For prognostic factor measurement, the risk of bias was generally low for studies that recorded PPI assessments during the first consultation. However, retrospective studies that calculated scores from available data may have been affected by the quality of clinical documentation. Some studies did not specify who completed the assessments. Outcome measurements were generally well-reported, though a few studies did not specify the duration of follow-up or how the date of death was determined. The risk of bias for study confounding was high or moderate when studies did not adjust for or specify relevant covariates.
Synthesis results
Detailed GRADE ratings are provided in Table S9. Due to the limited number of studies (less than 10 per meta-analysis), subgroup analyses based on study design, setting, risk of bias, and assessment for publication bias could not be conducted.
PPI scores as categorical variables
PPI > 6 vs PPI ≤ 4 risk groups
The pooled adjusted HR was 5.42 (95% CI 2.01–14.59, p = 0.0009) (Chou et al. Reference Chou, Kao and Wang2015; Palomar-Muñoz et al. Reference Palomar-Muñoz, Martín-Zamorano and Mogollo2018), with considerable heterogeneity (I 2 = 84%, p = 0.012) (n = 539, high-quality evidence). The pooled unadjusted HR (Cheng et al. Reference Cheng, Kao and Hung2012; Shatri et al. Reference Shatri, Prasetyaningtyas and Putranto2021) was 5.05 (95% CI 4.10–6.17, p < 0.00001) with nonsignificant heterogeneity (I 2 = 0%, p = 0.40) (n = 783, high-quality evidence) (Fig. 3A).
Figure 3. (A) Forest plots for meta-analysis of adjusted and unadjusted hazard ratios for associations between PPI (as a categorical variable) and risk of death for cutoffs 4 and 6. (B) Forest plot for meta-analysis of hazard ratios for the association between PPI (as a continuous variable) and risk of death. (C) Forest plot for meta-analysis of unadjusted risk ratios for inpatient death. Note that the figures show log[hazard ratio] and 95% CI – the results in the main text are in hazard ratio and 95% CI.
4 < PPI ≤ 6 vs PPI ≤ 4 risk groups
Two studies were pooled (Chou et al. Reference Chou, Kao and Wang2015; Palomar-Muñoz et al. Reference Palomar-Muñoz, Martín-Zamorano and Mogollo2018), yielding as adjusted HR of 2.04 (95% CI 1.30–3.21, p = 0.002) with nonsignificant heterogeneity (I 2 = 17.4%, p = 0.271) (n = 539, high-quality evidence) (Fig. 3A).
PPI ≥ 6 vs PPI < 6 risk groups
Three studies (Ahn et al. Reference Ahn, Hwang and Lee2016; Chang et al. Reference Chang, Lee and Huang2021; Inomata et al. Reference Inomata, Hayashi and Tokui2014) were included in the meta-analysis, with a pooled adjusted HR of 2.52 (95% CI 1.39–4.58, p = 0.002), showing considerable heterogeneity (I 2 = 74.5%, p = 0.01) (n = 333, moderate quality evidence) (Fig. 3A).
PPI scores as a continuous variable
Four studies analyzed PPI as continuous variables (Arai et al. Reference Arai, Okajima and Kotani2014; Lee et al. Reference Lee, Lee and Yang2014; Stone et al. Reference Stone, Tiernan and Dooley2008; Subramaniam et al. Reference Subramaniam, Thorns and Ridout2013). The pooled unadjusted HR was 1.30 (95% CI 1.22–1.38, p < 0.00001) with substantial heterogeneity (I 2 = 60%, p = 0.06) (n = 815, low-quality evidence) (Fig. 3B). This indicates that for each 1-point increase in PPI score, there is a 30% higher risk of mortality.
Other comparisons
Only 3 studies used the PPI thresholds of 2 and 4 for survival analyses, as defined in the original study (Morita et al. Reference Morita, Tsunoda and Inoue1999) (Chen et al. Reference Chen, Mauro and Gabrielli2018; Iizuka-Honma et al. Reference Iizuka-Honma, Mitsumori and Yoshikawa2023; Kao et al. Reference Kao, Hung and Wang2014) (Table 2). Other comparisons that could not be meta-analyzed are presented in Table 2 (Ahn et al. Reference Ahn, Ahn and Park2021; Cheng et al. Reference Cheng, Kao and Hung2012; Kiuchi et al. Reference Kiuchi, Hasegawa and Watanabe2022; Miura et al. Reference Miura, Matsumoto and Hama2015).
PPI scores as dichotomous variables (RR or OR)
Six studies reported dichotomous outcomes (Al-Ansari et al. Reference Al-Ansari, Abd-El-Gawad and AboSerea2022; Arkın and Aras Reference Arkın and Aras2021; Fernandes et al. Reference Fernandes, Branco and Fernandez2021; Gerber et al. Reference Gerber, Tuer and Yates2021; Lee et al. Reference Lee, Chou and Yang2022; Trejo-Ayala et al. Reference Trejo-Ayala, Ramos-Peñafiel and Santoyo-Sánchez2018) (Table 2). Two studies (Arkın and Aras Reference Arkın and Aras2021; Gerber et al. Reference Gerber, Tuer and Yates2021) were included in the meta-analysis, yielding a pooled unadjusted RR of 3.48 (95% CI 2.46–4.91, p < 0.00001) for PPI ≤ 4 vs PPI > 6 in predicting inpatient death (n = 274, high-quality evidence). Heterogeneity was nonsignificant (I 2 = 0%, p = 0.64) (Fig. 3C). Findings from other studies that could not be meta-analyzed are presented in Table 2.
Change in PPI
Three studies investigated changes in PPI scores as a predictor of survival (Arai et al. Reference Arai, Okajima and Kotani2014; Hung et al. Reference Hung, Wang and Kao2014; Kao et al. Reference Kao, Hung and Wang2014) (Table 2). Hung et al. (Reference Hung, Wang and Kao2014) and Kao et al. (Reference Kao, Hung and Wang2014) examined the same patient cohort, but Hung et al. (Reference Hung, Wang and Kao2014) only involved those with PPI > 6 (poor prognosis). Arai et al. (Reference Arai, Okajima and Kotani2014) calculated the change in PPI per day. For Kao et al. (Reference Kao, Hung and Wang2014), the median survival was the shortest for the group with <0 change in score, followed by the 0 and >0 groups. In Hung et al. (Reference Hung, Wang and Kao2014), the group with >20% change in score had the shortest median overall survival, followed by the 0–20%, 0, −20 to 0%, and <−20% change groups. Although the studies used different methods to categorize and calculate changes in PPI, most comparisons indicated that changes in PPI score were a statistically significant prognostic factor for survival.
Discussion
Main findings
This review is the first to confirm an independent association between PPI scores and survival in advanced cancer patients. It expands on the findings of a previous systematic review, which concluded that higher PPI scores significantly predicted a shorter survival period based solely on unadjusted HR (Liu et al. Reference Liu, Su and Wang2018). Our review includes more recent studies, larger sample sizes, and an analysis of both adjusted and adjustable effect sizes. Additionally, we assessed the risk of bias and the certainty of evidence using the GRADE framework, an evaluation that was not conducted in the previous review (Liu et al. Reference Liu, Su and Wang2018). Building on the findings of the original PPI development study (Morita et al. Reference Morita, Tsunoda and Inoue1999), all meta-analyses in this review found that the association between PPI and survival remained significant even after adjusting for covariates, with significant differences in survival among the risk groups.
Most included studies conducted an initial patient assessment using the PPI upon admission. However, unexpected events at the end-of-life are common, and reasons for hospitalizations can vary, meaning that the first assessment might not entirely reflect the patient’s overall prognosis. One study found that cancer patients admitted to the palliative care unit with treatable acute conditions (e.g. infections, hemorrhage) were more likely to survive than those admitted due to cancer-related issues (e.g. refractory symptoms, disease progression). Additionally, no significant differences in survival were observed among risk groups when using PPI at discharge (Palomar-Muñoz et al. Reference Palomar-Muñoz, Martín-Zamorano and Mogollo2018). This highlights the need for caution when interpreting PPI scores across different patient populations and time points.
Changes in PPI scores could also serve as a significant prognostic factor in predicting survival, particularly in capturing sudden shifts in patients’ conditions during end-of-life care. A study found that worsened symptom scores 1 week after admission were associated with shorter survival compared to patients with improved symptoms. In contrast, those with stable and improved symptom scores showed no significant differences in survival (Suh et al. Reference Suh, Won and Hiratsuka2022). Similarly, 3 studies (Arai et al. Reference Arai, Okajima and Kotani2014; Hung et al. Reference Hung, Wang and Kao2014; Kao et al. Reference Kao, Hung and Wang2014) found significant associations between change in PPI scores and survival outcomes. In addition, Kao et al. (Reference Kao, Hung and Wang2014) found a model combining the initial PPI score with the change in score had the highest c-statistic, further supporting the importance of monitoring PPI score changes over time. The Model combining the initial PPI score with changes in the score proved to be a better predictor of 30-day survival than using the initial score, Week 1 PPI score or score change individually. Another study found that a second PPI assessment conducted on Days 3–5 in hospice residents had better discriminative performance than the first assessment at admission (Subramaniam et al. Reference Subramaniam, Dand and Ridout2019). The studies (Arai et al. Reference Arai, Okajima and Kotani2014; Hung et al. Reference Hung, Wang and Kao2014; Kao et al. Reference Kao, Hung and Wang2014) included employed different methods for calculating PPI. Future research should standardize these calculation methods to provide more reliable conclusions regarding the utility of PPI in prognostication.
PPI as a continuous variable may also have prognostic significance. This was observed in another study involving older adults receiving home palliative care (7.5% had cancer), which reported a 1.51-fold increased probability of death for each unit increase in PPI (Moretti et al. Reference Moretti, Cechinel and Espindola2019). An included study (Gerber et al. Reference Gerber, Tuer and Yates2021) reported an OR of 0.74, suggesting that a lower overall PPI score significantly predicted survival to discharge. However, this result became nonsignificant in multivariate analysis. Hence, the utility of PPI scores as a continuous variable warrants further research.
Finally, we found that the risk of inpatient death was significantly higher for patients with a PPI > 6 compared to those with a PPI ≤ 4. This finding aligns with a previous study, which reported that the mean PPI scores of patients who died in the hospital were significantly higher than those of patients who survived to discharge (8.2 ± 3.8 vs 3.2 ± 2.9, p < 0.001) (Alshemmari et al. Reference Alshemmari, Ezzat and Samir2012).
What this study adds
PPI is not only a reasonably accurate prognostic tool for predicting <3- and <6-week survival in cancer patients (Yoong et al. Reference Yoong, Porock and Whitty2023), but the findings of this review also suggest that a higher PPI score is a strong and independent prognostic factor for poorer survival outcomes in advanced cancer patients. Furthermore, the PPI could support current clinical practice guidelines, which recommend the early integration of palliative care into standard oncology treatment for patients with advanced cancer receiving concurrent active treatment (Corsi et al. Reference Corsi, Turriziani and Cavanna2019; Ferrell et al. Reference Ferrell, Temel and Temin2017; Lee et al. Reference Lee, Chou and Yang2022). By assisting clinicians in identifying cancer patients suitable for early palliative care, the PPI could enhance clinical decision-making, helping clinicians determine whether additional curative treatment may benefit the patient or if palliative care should be initiated (Cohen and Miner Reference Cohen and Miner2019; Hasegawa et al. Reference Hasegawa, Yamakado and Takaki2015; Pobar et al. Reference Pobar, Job and Holt2021).
PPI could also be valuable when an objective estimate of survival is needed, e.g. determining participants’ eligibility for clinical trials (Chu et al. Reference Chu, Anderson and White2020; Simms et al. Reference Simms, Barraclough and Govindan2013), conducting risk stratification in stratified randomized trials, or avoiding bias in treatment effect estimation by adjusting for PPI (Halabi and Owzar Reference Halabi and Owzar2010). It may also help identify patients with poorer outcomes, thereby encouraging clinical trial participation for novel or experimental treatments (Gospodarowicz et al. Reference Gospodarowicz, Mackillop and O’Sullivan2001). A study examining the impact of palliative radiotherapy on gastric cancer patients’ symptoms found that, after adjusting for baseline PPI (since patients with limited life expectancy often experience worsening symptoms), shortness of breath, pain, and distress significantly improved over 8 weeks. Additionally, higher PPI scores were associated with higher symptom scores at all time points (Kawamoto et al. Reference Kawamoto, Saito and Kosugi2022). Another study identified a baseline PPI of >2 as a reliable predictor of death within 2 months in patients with advanced gastric cancer patients, suggesting it may be suitable for guiding single-fraction radiotherapy (Sekii et al. Reference Sekii, Saito and Kosugi2023).
Although various prognostic factors and prediction models have been identified for cancer patients, many were specific to certain cancer types or complications, limiting their clinical applicability to the broader cancer population (Owusuaa et al. Reference Owusuaa, Dijkland and Nieboer2022). A prediction model that is simple to use, applicable to heterogeneous cancer populations, and accessible to medical specialists, general practitioners, and nurses is highly desirable, as it could aid in treatment planning and advance care decisions (Owusuaa et al. Reference Owusuaa, Dijkland and Nieboer2022). Some studies have pointed out the challenges of using certain prognostic tools due to the unavailability of blood test results (Baba et al. Reference Baba, Maeda and Morita2015; Kishino et al. Reference Kishino, Monden and Akisada2022). In addition, many existing prediction models lack external validation, and model calibration is rarely assessed, underscoring the need for well-performing, validated models that are applicable to most cancer patients (Kreuzberger et al. Reference Kreuzberger, Damen and Trivella2020; Owusuaa et al. Reference Owusuaa, Dijkland and Nieboer2022). The PPI tool could help address this gap, as it has been widely validated and accepted across diverse settings and cancer populations.
Strengths and limitations of the study
This is the first meta-analysis to report an independent association between PPI scores and survival, and it represents the most comprehensive systematic review on the prognostic utility of PPI to date. The finding may offer valuable insights that can benefit both clinicians and researchers.
This review has several limitations. First, only articles in English were included, which may have resulted in the exclusion of relevant studies published in other languages. There were also limited studies in each meta-analysis, so the results should be interpreted with caution. As a result, subgroup analysis and tests for publication bias could not be conducted. We also did not estimate HR from the published Kaplan–Meier curves, as most studies did not report numbers at risk, which hindered this estimation. Moreover, studies that did not provide effect sizes (e.g. only reporting a significant log-rank test) were excluded, meaning this review does not represent all available literature on the association between PPI and survival. Despite these limitations, this review aimed to evaluate whether PPI is a prognostic factor for survival; thus, making the synthesis of time-to-event outcome measures the most appropriate approach.
Implications for research and practice
The PPI was initially developed using a heterogeneous sample of patients with different types of cancers (Morita et al. Reference Morita, Tsunoda and Inoue1999). Subsequently, its utility has been investigated and validated in specific cancer types, including lung cancer (Arkın and Aras Reference Arkın and Aras2021; Inomata et al. Reference Inomata, Hayashi and Tokui2014), hematological malignancies (Chang et al. Reference Chang, Lee and Huang2021; Chou et al. Reference Chou, Kao and Wang2015; Iizuka-Honma et al. Reference Iizuka-Honma, Mitsumori and Yoshikawa2023; Trejo-Ayala et al. Reference Trejo-Ayala, Ramos-Peñafiel and Santoyo-Sánchez2018), and ovarian cancer (Kiuchi et al. Reference Kiuchi, Hasegawa and Watanabe2022). One study also found that PPI was associated with survival in patients with non-Hodgkin’s lymphoma but not in those with acute myeloid leukemia in the palliative care setting (Yamane et al. Reference Yamane, Ochi and Mimura2023). Future research should continue to explore whether the prognostic utility of PPI differs across cancer types, patient care settings (such as acute wards, home palliative care, hospices, etc.) and stages of the cancer treatment journey (e.g. during active treatment or palliative care), similar to how the Glasgow Prognostic Score has been comprehensively evaluated for various cancers (He et al. Reference He, Li and Cai2018; Tong et al. Reference Tong, Guan and Xiong2020; Wu et al. Reference Wu, Wang and Shi2021).
Most of the included studies had a moderate to high risk of bias, highlighting the need for improving reporting in future research. To strengthen credibility and ensure the findings are more reliable for practical application, future studies should adhere to established reporting guidelines (Altman et al. Reference Altman, McShane and Sauerbrei2012; Hayden et al. Reference Hayden, van der Windt and Cartwright2013). It is also crucial to report adjusted prognostic effect measures, as these are important for quantifying the extent of the increased mortality risk across PPI risk groups. We observed that the categorization of PPI risk groups was inconsistent, with only 3 out of 23 studies using the risk groups defined in the original development study, and a maximum of 3 studies testing the same comparison. As a result, our meta-analyses were limited by the small number of studies. Further research should validate our findings by further examining the predictive value of PPI score categories (PPI ≤ 2, 2 < PPI ≤ 4, and PPI > 4) as defined in the original development study.
Conclusion
Higher PPI scores were strongly associated with poorer survival outcomes in advanced cancer patients. While the limited number of studies in each risk group comparison constrained our meta-analyses, the findings were consistent in both direction and significance. Future studies should adhere to the risk categories defined in the original development study and report adjusted effect estimates with 95% CI to strengthen the evidence base.
Supplementary material
The supplementary material for this article can be found at https://doi.org/.10.1017/S1478951525000021.
Funding
This research did not receive any funding from agencies in the public, commercial, or not-for-profit sectors.
Competing interests
None.
Open access
