Prognostic value of inflammation-immunity-nutrition score in cancer patients: a meta-analysis

Abstract

Background

Cancer remains a leading cause of mortality worldwide, with inflammation and immune dysregulation playing crucial roles in tumor progression. The inflammation-immunity-nutrition score (IINS), integrating high-sensitivity C-reactive protein, lymphocyte count, and albumin, has emerged as a potential prognostic biomarker in cancer, but its generalizability across tumor types is unclear.

Methods

PubMed, Cochrane Library, and EMBASE were searched for studies up to September 2025 assessing the relationship between IINS and survival outcomes in cancer patients. Hazard ratios (HRs) for overall survival (OS) and progression-free survival (PFS) were pooled using appropriate statistical models.

Results

Eight eligible articles involving 3,950 unique participants contributed 11 non-overlapping analytic cohorts. All participants were from Chinese populations, and the evidence was predominantly retrospective, with one prospectively assembled validation cohort. Elevated IINS was significantly associated with poorer OS (HR = 2.13, 95% CI: 1.56–2.90, p < 0.001) and shorter PFS (HR = 1.95, 95% CI: 1.65–2.31, p < 0.001). The binary high-versus-low analyses and the graded IINS-category analyses were based on different sets of cohorts according to the form in which effect estimates were reported. Higher IINS categories were associated with progressively less favorable survival outcomes.

Conclusions

Elevated IINS is associated with poorer survival outcomes and may serve as an accessible prognostic marker in cancer. However, the current evidence is predominantly retrospective and entirely derived from Chinese populations. Prospective, multicentre, multiethnic validation and direct comparison with established prognostic indices are required before clinical implementation.

1. Introduction

Despite substantial advances in cancer prevention, early detection, and therapeutic strategies, malignancies remain the second leading cause of death worldwide [1]. Although survival outcomes for several tumor types have improved, the overall incidence and mortality of cancer continue to rise [1]. Projections estimate that cancer-related deaths—nearly 10 million in 2022—could reach 16.3 million by 2040, largely driven by population aging and changing environmental and lifestyle factors [2,3]. With treatment innovations prolonging survival, the number of cancer survivors is expanding rapidly. However, many of these individuals experience long-term complications attributable to both the underlying disease and its treatment [4,5]. This has prompted an urgent need for robust prognostic biomarkers that can better stratify risk, allowing intensified therapy for patients with aggressive disease while minimizing overtreatment and treatment-related toxicity in those with more favorable trajectories [6,7].

Chronic inflammation and dysregulated innate immunity are now recognized as important features of tumorigenesis, promoting angiogenesis, immune escape, and disease progression. An increasing body of evidence links systemic inflammatory responses and host immune competence to cancer outcomes [8,9]. Several routinely available blood-based indices, including the neutrophil-to-lymphocyte ratio (NLR) and platelet-to-lymphocyte ratio (PLR), have therefore been investigated as prognostic markers in different malignancies [10]. Among individual biomarkers, high-sensitivity C-reactive protein (hsCRP) reflects systemic inflammatory activity [11,12], whereas lymphocyte count (LYM) and serum albumin (ALB) provide information on immune competence and nutritional or metabolic reserve, respectively [13,14]. Composite indices combine these variables in different ways. The prognostic nutritional index (PNI) integrates ALB and LYM [15]; the systemic immune-inflammation index (SII) incorporates platelet, neutrophil, and lymphocyte counts [16]; the systemic inflammation response index (SIRI) combines neutrophil, monocyte, and lymphocyte counts [17]; and the modified Glasgow prognostic score (mGPS) is based on C-reactive protein and ALB [18]. Thus, these indices capture overlapping but non-identical dimensions of host inflammation, immunity, and nutritional status. The inflammation-immunity-nutrition score (IINS) similarly integrates hsCRP, LYM, and ALB within a single scoring system. Although this structure is conceptually comprehensive, it should not be interpreted as inherently superior to established indices without direct comparative validation.

Building on this rationale, Li et al. [19] introduced the IINS and reported that elevated IINS was associated with treatment failure and poorer prognosis in patients with resectable colorectal cancer. Zhang et al. [20] subsequently reported that IINS showed prognostic discrimination for overall survival (OS) in a cohort of patients with hepatocellular carcinoma receiving programmed cell death protein 1 (PD-1) inhibitor therapy. Nevertheless, the available evidence has remained fragmented across individual cancer types, treatment settings, study populations, and IINS scoring thresholds. At the time of our literature search, no meta-analysis had quantitatively synthesized hazard ratios for OS and progression-free survival (PFS) across multiple malignancies. Consequently, the overall magnitude and consistency of the prognostic association, as well as its generalizability across cancer settings, remained unclear. To address these gaps, we conducted a meta-analysis to quantify the associations between IINS and survival outcomes in patients with cancer and to evaluate the potential utility and current limitations of IINS as a prognostic biomarker.

2. Methods

2.1. Search strategy, inclusion, and exclusion criteria

This analysis adhered to the PRISMA recommendations [21]. A comprehensive literature search was performed in three primary databases—PubMed, Cochrane Library, and EMBASE—to identify all relevant studies published up to 2 September 2025. Predefined search terms, including ‘inflammation-immunity-nutrition score’ and ‘IINS,’ were applied to ensure exhaustive retrieval of pertinent articles. Additionally, reference lists of included studies were screened manually to capture any citations missed during database searches. To enhance methodological rigor and reduce selection bias, two investigators independently conducted study selection, with disagreements resolved by consultation with the senior author.

Studies were considered eligible if they met the following criteria [1]: retrospective or prospective research examining the association between IINS and survival endpoints, such as progression-free survival (PFS) or overall survival (OS) [2]; reported hazard ratios (HRs) with corresponding 95% confidence intervals (CIs) for relevant comparisons; and [3] availability of complete English-language text. Exclusion criteria were [1] duplicate publications [2]; case series, conference abstracts, case reports, narrative reviews, editorials, or guidelines. When multiple studies reported on overlapping patient populations, priority was given to the most methodologically robust analysis with the largest sample size [22].

2.2. Data extraction and quality assessment

Two investigators independently extracted the first author, publication year, study period, geographic region, cancer type, treatment modality, sample size, patient characteristics, IINS definition and cutoff value, survival endpoint, and follow-up information. For each survival outcome, we also recorded the source of the hazard ratio (HR), including whether it was obtained from a multivariable analysis, univariable analysis, or reconstructed from a Kaplan–Meier curve, as well as the covariates included in the reported model. When multiple estimates were available, the HR from the most fully adjusted multivariable model was preferentially extracted. If an adjusted estimate was unavailable, an HR from a univariable analysis or one reconstructed from a Kaplan–Meier curve was used. Disagreements were resolved through discussion or consultation with a senior investigator. Participant counts were calculated separately for each pooled comparison and refer to unique participants contributing to that comparison. Training and validation cohorts were treated as separate analytic cohorts only when they comprised independent, non-overlapping participants and provided separate survival-effect estimates. For He et al. 435 participants were randomly divided into training and validation cohorts at a 7:3 ratio [23]. Because cohort-specific baseline characteristics were not separately available, the characteristics are presented as a single article-level entry in Table 1, whereas the two cohort-specific effect estimates were entered separately in the relevant meta-analyses. Binary high-versus-low analyses and graded IINS-category analyses were based on different cohort sets according to the reporting format of the original articles.

Table 1.

Main characteristics of the included articles and analytic cohorts.

Study	Country	Cancer type	Sample size	Age, years	Gender (male/female)	Treatment	Study period	The definition of IINS	NOS score
Wang et al. 2025	China	Breast cancer	200	55.5 (34–75)a	0/200	Surgical treatment	01/2017–12/2018	hsCRP: score 0, ≤6 mg/L; score 1, >6 to ≤12 mg/L; score 2, >12 mg/L. LYM: score 0, >1.6 × 10⁹/L; score 1, >1.1 to ≤1.6 × 10⁹/L; score 2, ≤1.1 × 10⁹/L. ALB: score 0, >38.4 g/L; score 1, >34.8 to ≤38.4 g/L; score 2, ≤34.8 g/L. IINS >2 was defined as high IINS.	8
Cheng et al. 2025 (T)	China	Non-small cell lung cancer	302	152/ 150b	207/95	R0 resection	02/2008–03/2013	hsCRP: score 0, ≤2.20 mg/L; score 1, >2.20 to ≤3.76 mg/L; score 2, >3.76 mg/L. LYM: score 0, >1.98 × 10⁹/L; score 1, >1.00 to ≤1.98 × 10⁹/L; score 2, ≤1.00 × 10⁹/L. ALB: score 0, >44.20 g/L; score 1, >35.40 to ≤44.20 g/L; score 2, ≤35.40 g/L. IINS >3 was defined as high IINS. The training-cohort definition and cutoff were applied unchanged to the validation cohort.	7
Cheng et al. 2025 (V)e	China	Non-small cell lung cancer	82	45/37b	56/26	R0 resection	01/2018–01/2021		7
He et al. 2024f	China	Extranodal natural killer/T-cell lymphoma	435	46 (7–83)a	295/140	–	01/2017–10/2021	hsCRP: score 0, ≤10.1 mg/L; score 1, >10.1 mg/L. ALB: score 0, >41.1 g/L; score 1, ≤41.1 g/L. LYM: score 0, >1.12 × 10⁹/L; score 1, ≤1.12 × 10⁹/L. Participants were classified as low IINS (0), medium IINS [1], or high IINS [2–3].	8
Zhang et al. 2022	China	Hepatocellular carcinoma	101	55 (49–63)c	84/17	PD-1 inhibitors	01/2018–08/2020	hsCRP: score 0, ≤2.8 mg/L; score 1, >2.8 to ≤34.3 mg/L; score 2, >34.3 mg/L. LYM: score 0, >1.18 × 10⁹/L; score 1, >0.91 to ≤1.18 × 10⁹/L; score 2, ≤0.91 × 10⁹/L. ALB: score 0, >37.7 g/L; score 1, >33.5 to ≤37.7 g/L; score 2, ≤33.5 g/L. IINS ≤3 was defined as low and IINS >3 as high.	7
Liang et al. 2022	China	Hepatocellular carcinoma	144	58 ± 12	123/21	Hepatectomy	07/2017–07/2021	hsCRP: score 0, ≤2.47 mg/L; score 1, >2.47 to ≤11.33 mg/L; score 2, >11.33 mg/L. LYM: score 0, >1.60 × 10⁹/L; score 1, >0.84 to ≤1.60 × 10⁹/L; score 2, ≤0.84 × 10⁹/L. ALB: score 0, >39.8 g/L; score 1, >35.2 to ≤39.8 g/L; score 2, ≤35.2 g/L. IINS >2 was defined as high IINS.	6
Li et al. 2021	China	Colorectal cancer	719	58 (50–66)c	427/292	Surgical treatment	01/2015–12/2017	hsCRP: score 0, ≤3.22 mg/L; score 1, >3.22 to ≤73.10 mg/L; score 2, >73.10 mg/L. LYM: score 0, >1.34 × 10⁹/L; score 1, >0.89 to ≤1.34 × 10⁹/L; score 2, ≤0.89 × 10⁹/L. ALB: score 0, >42.60 g/L; score 1, >36.10 to ≤42.60 g/L; score 2, ≤36.10 g/L. IINS >2 was defined as high IINS.	7
Song et al. 2022	China	Hepatocellular carcinoma	701	302/ 399d	601/100	R0 resection	01/2014–12/2017	CRP: score 0, ≤4.93 mg/L; score 1, >4.93 mg/L. LYM: score 0, >1.22 × 10⁹/L; score 1, ≤1.22 × 10⁹/L. ALB: score 0, >37.5 g/L; score 1, ≤37.5 g/L.	8
Jiang et al. 2022 (T)g	China	Endometrial cancer	775	53.8 ± 9.3	0/775	Surgical treatment	10/2013–05/2018	CRP: score 0, <4.76 mg/L; score 1, ≥4.76 mg/L. LYM: score 0, ≥1.37 × 10⁹/L; score 1, <1.37 × 10⁹/L. ALB: score 0, ≥37 g/L; score 1, <37 g/L.	7
Jiang et al. 2022 (V)g	China	Endometrial cancer	491	53.3 ± 9.5	0/491	Surgical treatment	10/2013–05/2018	The training-cohort CRP, LYM, and ALB definitions were applied unchanged to the validation cohort.	7

Abbreviations: ALB, albumin; CRP, C-reactive protein; hsCRP, high-sensitivity C-reactive protein; IINS, inflammation-immunity-nutrition score; IQR, interquartile range; LYM, lymphocyte count; NOS, Newcastle–Ottawa Scale; PD-1, programmed cell death protein 1; R0, microscopically margin-negative resection; SD, standard deviation; T, training cohort; V, validation cohort.

^a Median (range).

^b Numbers of participants aged ≥60/<60 years, respectively.

^c Median (interquartile range).

^d Numbers of participants aged ≥50/<50 years, respectively. Values shown as mean ± SD are presented without a footnote.

^e Cheng et al. (2025) (V) was a prospectively assembled validation cohort; all remaining cohorts were retrospective.

^f He et al. (2024) randomly divided 435 participants into independent, non-overlapping training and validation cohorts at a 7:3 ratio. Because cohort-specific baseline characteristics were not separately available, article-level characteristics are shown, whereas separate cohort-specific effect estimates were used in the meta-analysis.

^g Jiang et al. (2022) reported distinct, non-overlapping training and validation cohorts covering the same study period; the cohorts were analyzed separately.

The methodological quality of the included non-randomized studies was independently assessed by two investigators using the Newcastle–Ottawa Scale (NOS) for cohort studies. The NOS evaluates three domains: Selection of the study population (0–4 points), Comparability of the cohorts (0–2 points), and Outcome assessment (0–3 points), yielding a maximum total score of 9 points [24]. Domain-specific and total NOS scores for each included study or cohort are presented in Supplementary Table S1. Studies with a total NOS score of six points or higher were considered to have relatively high methodological quality.

2.3. Statistical methods

All meta-analyses were performed with Stata version 18.0, and pooled results were illustrated using forest plots. Between-study heterogeneity was quantified using Cochran’s Q statistic together with the I² metric, considering variability to be significant when I² exceeded 50% or when the Q-test yielded a p-value below 0.1 [25]. When substantial heterogeneity was detected, the DerSimonian–Laird random-effects model was applied; in its absence, a fixed-effect model based on the inverse-variance approach was used. Publication bias was explored using Begg’s and Egger’s tests [6]. Because fewer than 10 analytic cohorts contributed to the OS and PFS analyses, these tests were expected to have limited statistical power; consequently, non-significant results were not interpreted as evidence that publication bias was absent. Sensitivity analyses were performed by sequentially removing each analytic cohort to examine its influence on the overall summary effect [26]. A two-sided p-value of less than 0.05 was regarded as statistically significant.

A post hoc exploratory subgroup analysis was performed for the high-versus-low OS comparison according to cancer type, using the same eight analytic cohorts included in the primary binary OS analysis. Cohorts were grouped as other cancers (breast and colorectal cancer), non-small cell lung cancer, extranodal natural killer/T-cell lymphoma, and hepatocellular carcinoma. The two endometrial cancer cohorts were not included because they contributed graded IINS-category estimates but did not contribute a binary high-versus-low OS estimate. Fixed-effect inverse-variance models were used within each subgroup, and subgroup differences were assessed using Cochran’s between-subgroup Q statistic with degrees of freedom equal to the number of subgroups minus one. Training and validation cohorts from the same article were non-overlapping and were treated as separate analytic units. Nevertheless, because cohorts derived from the same publication shared aspects of study design and reporting, residual within-publication methodological dependence could not be excluded; no multilevel adjustment was performed because only a small number of articles contributed multiple cohorts.

3. Results

3.1. Search results and study characteristics

The initial database search, supplemented by manual screening of reference lists, yielded 319 potentially eligible records. After the removal of 55 duplicates, 225 records were excluded during title and abstract screening because they did not meet the eligibility criteria. The full texts of the remaining 39 articles were assessed, of which 31 were excluded for not satisfying the predefined inclusion criteria. Consequently, eight eligible articles involving 3,950 unique participants were included in the quantitative synthesis [19,20,23,27–31] (Figure 1). These articles contributed 11 non-overlapping analytic cohorts: one cohort each from Wang et al. (2025), Zhang et al. (2022), Liang et al. (2022), Li et al. (2021), and Song et al. (2022), and separate training and validation cohorts from Cheng et al. (2025), He et al. (2024), and Jiang et al. (2022). For He et al. (2024), cohort-specific baseline characteristics were not separately available; therefore, the training and validation cohorts are presented together as a single article-level entry in Table 1, with a combined sample size of 435 participants, whereas their separate effect estimates were included as two analytic cohorts in the relevant pooled analyses.

Figure 1.

Flow diagram of the identification and selection of eligible articles.

Table 1 summarizes the principal characteristics of the included articles and cohorts. Article-level or cohort-level sample sizes ranged from 82 to 775 participants. All participants were from Chinese populations. The included evidence covered hepatocellular carcinoma, non-small cell lung cancer, endometrial cancer, breast cancer, colorectal cancer, and extranodal natural killer/T-cell lymphoma. Three articles focused on hepatocellular carcinoma, whereas the articles on non-small cell lung cancer, extranodal natural killer/T-cell lymphoma, and endometrial cancer each contributed separate training and validation cohorts where applicable. Eight of the 11 analytic cohorts included patients who underwent surgical treatment. The evidence was predominantly derived from retrospective cohorts; the Cheng et al.’s (2025) validation cohort was prospectively assembled. Newcastle–Ottawa Scale scores ranged from 6 to 8, indicating moderate to relatively high methodological quality according to the prespecified NOS threshold, rather than establishing a low risk of bias (Table 1).

Domain-specific NOS assessments for the included articles or cohorts are presented in Supplementary Table S1. The sources of the survival effect estimates and the covariates included in the original models are summarized in Supplementary Table S2. Most eligible effect estimates were obtained from multivariable Cox regression models, whereas univariable or Kaplan–Meier-derived estimates were used when adjusted HRs were unavailable. The covariate adjustment sets varied across studies and included different combinations of demographic characteristics, performance status, tumor stage and burden, pathological features, treatment-related factors, and laboratory variables. Therefore, although the most fully adjusted estimates were preferentially selected, the extent of confounder adjustment was not uniform across the included studies. Supplementary Tables S3 and S4 identify the articles, analytic cohorts, and participant counts contributing to each pooled analysis and cancer-type subgroup.

3.2. Baseline IINS and overall survival

The high-versus-low OS analysis included eight non-overlapping analytic cohorts from six articles, comprising 1,983 unique participants: Wang et al. (2025) (n = 200), the training and validation cohorts of Cheng et al. (2025) (n = 302 and n = 82, respectively), the training and validation cohorts of He et al. (2024) (combined n = 435), Zhang et al. (2022) (n = 101), Liang et al. (2022) (n = 144), and Li et al. (2021) (n = 719). Pooled hazard ratios indicated that elevated IINS was associated with poorer OS (HR = 2.13, 95% CI: 1.56–2.90, p < 0.001; Figure 2A). Substantial heterogeneity was observed across analytic cohorts (I² = 72.6%, Cochran’s Q test p < 0.001), and a random-effects model was therefore used.

Figure 2.

Association between baseline inflammation-immunity-nutrition score and overall survival. The analysis included eight non-overlapping analytic cohorts from six articles and 1,983 unique participants. (A) Forest plot of the association between elevated inflammation-immunity-nutrition score (IINS) and overall survival (OS). The pooled estimate was calculated using a DerSimonian–Laird random-effects model (HR = 2.13, 95% CI: 1.56–2.90; I² = 72.6%, Cochran’s Q test p < 0.001). (B) Leave-one-out sensitivity analysis. CI, confidence interval; HR, hazard ratio; IINS, inflammation-immunity-nutrition score; OS, overall survival; T, training cohort; V, validation cohort.

A leave-one-out sensitivity analysis was performed to evaluate the robustness of the pooled estimate. Exclusion of any single analytic cohort did not materially alter the combined effect estimate (Figure 2B). Begg’s and Egger’s tests did not reach statistical significance (Begg p = 0.536; Egger p = 0.124). However, because only eight analytic cohorts contributed to this analysis, these tests had limited power, and the findings do not exclude the possibility of publication bias.

The graded OS analyses included different cohort sets according to the available IINS-category comparisons. The IINS 1-versus-0 analysis included five analytic cohorts from three articles—Song et al. (2022), the training and validation cohorts of Jiang et al. (2022), and the training and validation cohorts of He et al. (2022)—comprising 2,402 unique participants. He et al. (2024) categorized participants as IINS 0, 1, and 2–3 and therefore contributed only to the IINS 1-versus-0 comparison. The IINS 2-versus-0 and IINS 3-versus-0 analyses each included three analytic cohorts from two articles—Song et al. (2022) and the training and validation cohorts of Jiang et al. (2022)—comprising 1,967 participants. Compared with IINS 0, higher categories were associated with progressively poorer OS (IINS 1 versus 0: HR = 2.18, 95% CI: 1.76–2.69, p < 0.001, Figure 3A; IINS 2 versus 0: HR = 3.41, 95% CI: 2.61–4.45, p < 0.001, Figure 3B; and IINS 3 versus 0: HR = 5.46, 95% CI: 3.82–7.82, p < 0.001, Figure 3C).

Figure 3.

Graded association between baseline inflammation-immunity-nutrition score and overall survival. Panel A included five analytic cohorts from three articles and 2,402 participants; Panels B and C each included three analytic cohorts from two articles and 1,967 participants. Forest plots compare overall survival among patients with different inflammation-immunity-nutrition score (IINS) categories: (A) IINS 1 versus 0 (HR = 2.18, 95% CI: 1.76–2.69, I² = 0%, Cochran’s Q test p = 0.812); (B) IINS 2 versus 0 (HR = 3.41, 95% CI: 2.61–4.45, I² = 0%, Cochran’s Q test p = 0.715); and (C) IINS 3 versus 0 (HR = 5.46, 95% CI: 3.82–7.82, I² = 0%, Cochran’s Q test p = 0.382). CI, confidence interval; HR, hazard ratio; IINS, inflammation-immunity-nutrition score; OS, overall survival; T, training cohort; V, validation cohort.

3.3. Pretreatment IINS and progression-free survival

The high-versus-low PFS analysis included four cohorts from four articles, comprising 1,164 unique participants: Wang et al. (2025) (n = 200), Zhang et al. (2022) (n = 101), Liang et al. (2022) (n = 144), and Li et al. (2021) (n = 719). The pooled hazard ratio indicated that elevated IINS was associated with shorter PFS (HR = 1.95, 95% CI: 1.65–2.31, p < 0.001; Figure 4A). Heterogeneity was minimal (I² = 0%, Cochran’s Q test p = 0.641), supporting use of a fixed-effect model.

Figure 4.

Association between baseline inflammation-immunity-nutrition score and progression-free survival. The analysis included four cohorts from four articles and 1,164 unique participants. (A) Forest plot of the association between elevated inflammation-immunity-nutrition score (IINS) and progression-free survival (PFS). The pooled estimate was calculated using a fixed-effect inverse-variance model (HR = 1.95, 95% CI: 1.65–2.31; I² = 0%, Cochran’s Q test p = 0.641). (B) Leave-one-out sensitivity analysis. CI, confidence interval; HR, hazard ratio; IINS, inflammation-immunity-nutrition score; PFS, progression-free survival.

A leave-one-out sensitivity analysis showed that removal of any single cohort had little influence on the pooled estimate (Figure 4B). Begg’s and Egger’s tests did not reach statistical significance (Begg p = 0.308; Egger p = 0.340). Nevertheless, with only four contributing cohorts, these tests had very limited power and cannot be interpreted as demonstrating the absence of publication bias.

The graded PFS analyses were based on a different, non-overlapping set of cohorts from those included in the high-versus-low PFS analysis. They included three analytic cohorts from two articles—Song et al. (2022) (n = 701) and the training and validation cohorts of Jiang et al. (2022) (n = 775 and n = 491, respectively)—comprising 1,967 unique participants. These cohorts provided separate estimates for IINS categories 1, 2, and 3 relative to IINS 0. Thus, the larger denominator in the graded PFS analysis reflects a different set of larger cohorts rather than an expansion of the four-cohort binary PFS analysis. Higher IINS categories were associated with progressively poorer PFS (IINS 1 versus 0: HR = 2.20, 95% CI: 1.18–4.11, p = 0.013, Figure 5A; IINS 2 versus 0: HR = 3.17, 95% CI: 1.76–5.72, p < 0.001, Figure 5B; and IINS 3 versus 0: HR = 3.44, 95% CI: 2.46–4.82, p < 0.001, Figure 5C).

Figure 5.

Graded association between baseline inflammation-immunity-nutrition score and progression-free survival. Each comparison included three analytic cohorts from two articles and 1,967 unique participants: Song et al. (2022) and the training and validation cohorts of Jiang et al. (2022). These cohorts were distinct from those included in the high-versus-low PFS analysis. DerSimonian–Laird random-effects models were used for: (A) IINS 1 versus 0 (HR = 2.20, 95% CI: 1.18–4.11; I² = 79.8%, Cochran’s Q test p = 0.007); (B) IINS 2 versus 0 (HR = 3.17, 95% CI: 1.76–5.72; I² = 63.7%, Cochran’s Q test p = 0.064); and (C) IINS 3 versus 0 (HR = 3.44, 95% CI: 2.46–4.82; I² = 0%, Cochran’s Q test p = 0.429). CI, confidence interval; HR, hazard ratio; IINS, inflammation-immunity-nutrition score; PFS, progression-free survival; T, training cohort; V, validation cohort.

3.4. Subgroup analysis according to cancer type

A post hoc exploratory cancer-type subgroup analysis was conducted using the eight analytic cohorts and 1,983 participants included in the primary high-versus-low OS analysis. The other-cancers subgroup comprised two independent articles and two cohorts—Wang et al. (2025) (breast cancer, n = 200) and Li et al. (2025) (colorectal cancer, n = 719)—and yielded an HR of 2.85 (95% CI: 2.20–3.67; I² = 0%). The non-small cell lung cancer estimate combined two non-overlapping cohorts from one article, Cheng et al. (2025) (training, n = 302; validation, n = 82), and should therefore be regarded as a study-specific within-publication summary rather than a pooled estimate across independent articles (HR = 1.40, 95% CI: 1.17–1.69; I² = 11.6%). Similarly, the extranodal natural killer/T-cell lymphoma estimate combined the training and validation cohorts from the single He et al. (2024) article (combined n = 435) and was study-specific (HR = 1.56, 95% CI: 1.19–2.05; I² = 36.6%). The hepatocellular carcinoma subgroup comprised two independent articles and two cohorts—Zhang et al. (2022) (n = 101) and Liang et al. (2022) (n = 144)—with an HR of 2.99 (95% CI: 1.45–6.15; I² = 0%) (Figure S1). The test for subgroup differences was recalculated using Cochran’s between-subgroup Q statistic under the fixed-effect inverse-variance framework (Qbetween = 22.4, df = 3, p < 0.001). Despite this statistical result, the subgroup analysis was post hoc, two subgroups were derived from only one article each, and the number of independent articles was small; the apparent differences in effect magnitude should therefore be interpreted cautiously and as hypothesis-generating.

4. Discussion

In this meta-analysis, elevated IINS was significantly associated with poorer OS and PFS outcomes in patients with cancer. A stepwise association was also observed, with increasing IINS scores corresponding to progressively higher risks of adverse survival outcomes. These findings suggest that IINS may provide clinically relevant prognostic information by integrating systemic inflammation, immune competence, and nutritional status. Nevertheless, the current evidence should be interpreted as demonstrating an association rather than establishing causality or confirming the superiority of IINS over other established prognostic indices.

The biological plausibility of the observed association is supported by the individual components of IINS. hsCRP is an acute-phase protein produced predominantly by hepatocytes in response to inflammatory signalling. Elevated circulating hsCRP may reflect a systemic pro-inflammatory state associated with tumour progression, angiogenesis, immune evasion, and metastatic dissemination [32–34]. In contrast, lymphocytes contribute substantially to antitumor immune surveillance. Reduced circulating lymphocyte counts may indicate impaired host immune competence and diminished capacity to generate effective antitumor responses, whereas greater circulating or intratumoral lymphocyte abundance has generally been associated with more favourable outcomes in several malignancies [35,36]. Serum albumin reflects not only nutritional status but also systemic inflammation, hepatic synthetic function, and physiological reserve. Hypoalbuminemia is common in patients with advanced cancer and may be associated with cachexia, impaired treatment tolerance, and unfavourable survival outcomes [37,38]. By combining these three components, IINS provides a composite representation of the interaction between systemic inflammation, immune function, and nutritional reserve.

From a clinical perspective, IINS has several potentially favourable characteristics. Its components are routinely measured, inexpensive, and widely available, allowing the score to be calculated without specialized testing. These features support further evaluation of IINS as a potential adjunct to established prognostic systems. However, the present meta-analysis demonstrates an association with prognosis and does not establish that IINS improves discrimination, reclassification, treatment selection, or clinical outcomes beyond tumour stage, performance status, molecular markers, or existing inflammation- and nutrition-based indices. Its clinical role should therefore remain investigational until prospective comparative studies demonstrate incremental value and clinical utility.

IINS should also be considered in the context of other inflammation- and nutrition-based indices. The prognostic nutritional index primarily integrates albumin and lymphocyte count, the systemic immune-inflammation index incorporates platelet, neutrophil, and lymphocyte counts, the systemic inflammation response index combines neutrophil, monocyte, and lymphocyte counts, and the modified Glasgow prognostic score is based on C-reactive protein and albumin. These indices reflect overlapping but non-identical biological dimensions. IINS is conceptually distinct because it combines an inflammatory marker, an immune-cell measure, and a nutritional marker within a single scoring framework. Nevertheless, this broader composition should not be interpreted as evidence of inherent superiority. Direct head-to-head comparisons remain limited, and differences in study populations, endpoints, cutoff definitions, and statistical models make indirect comparisons unreliable.

Some included studies in hepatocellular carcinoma reported that IINS showed favourable prognostic discrimination compared with conventional clinical variables and several established inflammatory or nutritional indices [29,30]. These observations support further evaluation of IINS, particularly in HCC, where systemic inflammation, liver function, nutritional status, and immune competence may all influence prognosis. However, comparative findings were concentrated in HCC cohorts and cannot yet be generalized to other malignancies. Larger prospective studies incorporating formal model-comparison approaches, such as discrimination, calibration, reclassification, and decision-curve analyses, are needed to determine whether IINS provides meaningful incremental prognostic value beyond existing clinical models.

The clinical heterogeneity of the included studies should also be considered. The patient populations differed in cancer type, disease stage, treatment modality, IINS definition, and survival endpoint. Some cohorts included patients undergoing surgery, whereas others included patients receiving systemic therapy, including immune checkpoint inhibition. Consequently, the pooled association represents an overall prognostic relationship across heterogeneous oncology settings rather than a treatment-specific predictive effect. In particular, the present findings do not establish that IINS predicts benefit from a particular intervention. Future studies should distinguish clearly between prognostic utility, which reflects outcomes independent of treatment, and predictive utility, which indicates differential benefit from a specific treatment.

Although the biological components of IINS may plausibly influence radiotherapy outcomes, radiotherapy-related information was insufficiently characterized in the studies included in this meta-analysis. Systemic inflammation, reduced lymphocyte reserve, and poor nutritional status may theoretically affect tumour radiosensitivity, immune-mediated responses to irradiation, normal-tissue repair, and treatment tolerance [39–43]. However, these considerations are hypothesis-generating and are not directly supported by the present pooled results.

Dedicated studies in radiotherapy-treated populations are therefore needed to determine whether pretreatment IINS is associated with local control, progression, survival, or radiation-related toxicity. Such studies should prospectively report radiotherapy indication, target volume, dose and fractionation, treatment technique, concurrent systemic therapy, treatment interruptions, and acute and late adverse events. Future investigations should also prioritize prospective, multicenter, and multiethnic validation, standardization of IINS scoring definitions and cutoff values, and direct comparison with established inflammatory, nutritional, clinical, and molecular prognostic models.

Several limitations should be acknowledged. First, the available evidence was predominantly derived from retrospective observational studies and was therefore susceptible to selection bias and residual or unmeasured confounding. Although the most fully adjusted HRs were preferentially extracted, covariate adjustment varied substantially across studies, which may have affected the comparability of the pooled estimates. Second, some articles contributed separate training and validation cohorts. These cohorts comprised non-overlapping participants and were therefore treated as separate analytic units; however, cohorts from the same publication shared elements of study design, eligibility criteria, and reporting, and residual within-publication methodological dependence could not be excluded. A multilevel or cluster-robust adjustment was not performed because only a small number of articles contributed multiple cohorts. Third, the evidence covered a limited range of malignancies and was entirely derived from Chinese populations. The cancer-type subgroup analysis was post hoc and included few independent articles; notably, the non-small cell lung cancer and extranodal natural killer/T-cell lymphoma estimates were each derived from training and validation cohorts within a single article. Accordingly, differences between cancer types should be considered exploratory and hypothesis-generating. Differences in IINS definitions and cutoff thresholds, treatment settings, and progression-related endpoint definitions further limited comparability and generalizability. The available evidence was insufficient for reliable subgroup or meta-regression analyses according to disease stage, treatment modality, or IINS cutoff definition. Finally, Begg’s and Egger’s tests were based on only eight analytic cohorts for OS and four cohorts for PFS and consequently had limited ability to detect publication bias; non-significant test results do not exclude its presence.

Despite these limitations, this meta-analysis provides a quantitative synthesis of the currently available evidence and supports a consistent association between elevated IINS and unfavourable survival outcomes. Further prospective studies are warranted to establish standardized cutoff values and clarify the incremental and treatment-specific prognostic value of IINS.

5. Conclusion

This meta-analysis indicates that elevated IINS is associated with poorer overall and progression-related survival outcomes across the cancer populations studied. The observed graded associations support further investigation of IINS as an accessible prognostic marker but do not establish causality, incremental prognostic value over existing indices, or utility for treatment decision-making. Prospective, multicentre, multiethnic studies using standardized IINS definitions and direct comparative model evaluation are required before IINS can be considered for routine clinical implementation.

Funding Statement

This work was supported by the National Natural Science Foundation of China (Nos. 81903000), the Basic Research Foundation of Luzhou People’s Government - Southwest Medical University Plan (Nos. 2024LZXNYDJ024), Natural Science Foundation of Sichuan Province (Nos. 2023NSFSC1846), Sichuan Medical Science and Technology Innovation Research Association.2024 Scientific Research Project (Nos. YCH-KY-YCZD2024-009).

Disclosure statement

No potential conflict of interest was reported by the authors.

Data availability statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.