Development and internal validation of a dynamic Prealbumin–Immune-Inflammation–NLR Prognostic Score (PINPS) for esophageal squamous cell carcinoma treated with immune checkpoint inhibitors: a retrospective cohort study

Abstract

Background

Immune checkpoint inhibitors improve survival in advanced esophageal squamous cell carcinoma (ESCC), yet many patients derive limited benefit, and identifying those at high risk of poor outcomes remains challenging. Most immune-nutritional composite scores use only pre-treatment data and have rarely been applied to on-treatment risk monitoring. We developed and internally validated the Prealbumin–Immune-Inflammation–Neutrophil-to-Lymphocyte Ratio Prognostic Score (PINPS) and explored a four-quadrant dynamic framework integrating baseline risk with early post-treatment change.

Methods

This single-centre retrospective study enrolled 549 consecutive ESCC patients treated with PD-1/PD-L1 inhibitors between 2019 and 2026. After excluding 24 with incomplete follow-up, 525 were analysed for overall survival (OS). After multiple imputation (m = 20), 12 candidate markers and indices were screened using LASSO-penalised Cox regression with bootstrap internal validation. As a post hoc extension, patients were cross-classified by baseline PINPS risk (high/low) and change in the earliest post-treatment value obtained between days 21 and 90 (decrease/non-decrease). Cytokine correlations (n = 325) and exploratory TBNK immunophenotyping (n = 239) were assessed in available subsets.

Results

During a median follow-up of 10.7 months, 150 deaths occurred. PINPS retained three routine blood-based components: prealbumin, pan-immune-inflammation value, and neutrophil-to-lymphocyte ratio. Its bootstrap-corrected C-index was 0.667 (95% CI 0.615–0.719), improving by 0.145 over clinical variables alone (P < 0.0001). Although continuous ΔPINPS magnitude was not associated with survival (P = 0.864), the four-quadrant framework discriminated OS in landmark analysis (P = 0.0032). Median survival was 18.3 versus 36.8 months in the persistent-high-risk versus baseline-low/decreasing-PINPS groups. Among baseline high-risk patients, 37.7% were reclassified as low risk during treatment. IL-6 was the only cytokine correlated with both baseline PINPS and early PINPS change (all FDR < 0.05), while exploratory TBNK analysis suggested T-cell polarisation.

Conclusions

PINPS provides clinically interpretable survival stratification for ESCC patients receiving PD-1/PD-L1 inhibitors. Its post hoc four-quadrant framework identified a persistent-high-risk subgroup with the worst prognosis and suggested on-treatment risk reclassification in over one-third of baseline high-risk patients. External validation is required before clinical use.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12885-026-16536-8.

Introduction

Esophageal squamous cell carcinoma (ESCC) remains a major global health burden, accounting for more than 540,000 deaths annually [1]. PD-1/PD-L1 inhibitors, validated in KEYNOTE-590 [2], CheckMate-648 [3], and ESCORT-1st [4], have become standard first-line treatment for advanced disease. Objective response rates, however, remain below 50%, and robust tools for identifying patients most likely to derive durable benefit are still lacking.

Blood-based inflammatory and nutritional indices have emerged as accessible candidate biomarkers in this setting. Among single-index markers, neutrophil-to-lymphocyte ratio (NLR; pooled hazard ratio [HR] 2.13) [5], systemic immune-inflammation index (SII; HR 2.33 in a meta-analysis of 2,438 patients) [6, 7], and prognostic nutritional index (PNI; HR 0.53 for high versus low in gastrointestinal cancers) [8] have each demonstrated prognostic value. Lung immune prognostic index (LIPI), among the earliest and most validated blood-based prognostic scores developed specifically for checkpoint inhibitor-treated patients, combines derived NLR with lactate dehydrogenase (LDH) [9]. Each of these tools, however, reflects a single biological axis and is usually assessed only at baseline.

Two gaps remain. First, growing evidence indicates that early treatment-induced shifts in inflammatory markers carry more prognostic weight than static baseline values. Matsumoto and colleagues showed that pairing baseline NLR with the direction of its early change produced a three-tiered classification superior to either measure alone in NSCLC chemoimmunotherapy [10]. This principle has not been tested with composite, multi-dimensional scores. Second, the biological connections linking nutritional decline to systemic inflammation and onward to adaptive immune cell composition have not been systematically characterised in ESCC.

We therefore aimed to develop a clinically applicable immune-nutritional prognostic framework for ESCC patients receiving PD-1/PD-L1 inhibitors and to explore its potential biological correlates. Specifically, we pursued four objectives: (1) derive and internally validate a composite prognostic score, PINPS, by LASSO-penalised Cox regression across candidate immune-nutritional indices; (2) identify cross-dimensional cytokine mediators connecting the nutritional and inflammatory arms of PINPS; (3) explore whether PINPS tracks underlying T-cell polarisation patterns through T-cell/B-cell/NK-cell (TBNK) subset profiling; and (4) test a four-quadrant dynamic monitoring framework combining baseline risk with on-treatment directional change.

Methods

Study design and population

This retrospective cohort study included consecutive patients with pathologically confirmed esophageal squamous cell carcinoma (ESCC) who received at least one cycle of PD-1/PD-L1 inhibitor therapy between January 2019 and January 2026 at Fudan University Shanghai Cancer Center. Eligible patients had baseline laboratory data available within 30 days before the first immune checkpoint inhibitor (ICI) dose. Patients with incomplete or unreliable survival follow-up data were excluded (n = 24; see Fig. 1 for details). The institutional ethics committee approved the study (2603-Exp392); informed consent was waived given the retrospective design.

Fig. 1.

Study flow and biomarker availability. Records from inpatient (n = 1,288) and outpatient (n = 857) sources were deduplicated to 1,227 unique patients. After sequential exclusion of non-esophageal diagnoses (n = 178), confirmed non-squamous histologies (n = 25), patients without PD-1/PD-L1 treatment records (n = 419), second primary malignancies (n = 47), autoimmune disease (n = 8), and haematological disease (n = 1), 549 patients with ESCC were eligible. Of these, 24 lacked valid OS data, leaving 525 for the primary analysis (150 events). PINPS was evaluable in 475 patients (50 lacked baseline prealbumin). Downstream analytical subsets: paired T0–T1 PINPS (n = 382; landmark-eligible n = 365, 105 events), temporal validation (training n = 253, 111 events; testing n = 222, 20 events), cytokine data (n = 325, 87 events), TBNK data (n = 239; complete panel n = 162, 17 events). Exploratory incremental LASSO-Cox analyses assessed cytokine (n = 298*, 77 events) and TBNK (n = 209*, 34 events) contributions beyond PINPS. PFS analysis (n = 417) was pre-specified but not formally reported owing to suboptimal NLP extraction accuracy (see Methods). *Intersection with PINPS components simultaneously non-missing

Data collection and time windows

Clinical and laboratory variables were extracted from the hospital information system.

Baseline (T0) values were the most recent measurement within 30 days before the first ICI dose (D − 30 to D0). Follow-up (T1) values were the earliest measurement between day 21 and day 90 after the first dose. This window was prespecified to capture early on-treatment changes while minimising the influence of acute chemotherapy-related cytopenias within the first 3 weeks.

Immune-nutritional indices

Eleven composite indices were computed from routine blood counts and biochemistry: NLR, platelet-to-lymphocyte ratio (PLR), lymphocyte-to-monocyte ratio (LMR), SII, systemic inflammation response index (SIRI), pan-immune-inflammation value (PIV) (also termed AISI, Aggregate Index of Systemic Inflammation; neutrophils × monocytes × platelets / lymphocytes) [11], LIPI [9], PNI, albumin–bilirubin index (ALBI), albumin-to-alkaline phosphatase ratio (AAPR), and hemoglobin, albumin, lymphocyte, and platelet (HALP). Prealbumin was included as a standalone nutritional marker. Its short half-life (~ 2 days) renders it more responsive to acute nutritional shifts than albumin (~ 21 days) [12]. Biologically, these indices index complementary host axes: NLR reflects the balance between innate (neutrophil-driven) inflammation and adaptive (lymphocyte-mediated) immunity; PIV aggregates neutrophil, monocyte, platelet, and lymphocyte counts into a single measure of systemic immune–inflammatory burden; and prealbumin, as a rapid-turnover hepatic transport protein, captures acute nutritional–inflammatory status.

Cytokine and TBNK flow cytometry

Twelve serum cytokines (interleukin [IL]-1β, IL-2, IL-4, IL-5, IL-6, IL-8, IL-10, IL-12, IL-17, TNF-α, IFN-α, IFN-γ) were measured by multiplex cytometric bead array (BD™ Cytometric Bead Array; BD Biosciences, San Jose, CA) on a BD FACSCanto II flow cytometer (BD Biosciences). TBNK lymphocyte subsets (CD3+, CD4+, CD8+, CD19+, NK cells, plus extended Th1, Th17, Treg, and activation panels) were analysed by multicolour flow cytometry on a BD FACSCanto II using BD Multitest 6-color TBNK reagent (Cat# 337166; BD Biosciences) supplemented with individual monoclonal antibodies from BD Biosciences, Beckman Coulter (Marseille, France), and Invitrogen (Carlsbad, CA). Data were analysed with BD FACSDiva software. Both assays were ordered at the treating physician’s discretion; subset availability is reported in the Results. Cytokine values below the detection limit were set to the assay-specific lower limit of quantification.

Statistical analysis

Primary endpoint and univariate screening

The primary endpoint was overall survival (OS) from first ICI dose to death or last follow-up. Median follow-up was estimated by the reverse Kaplan–Meier method [13]. Follow-up data were administratively censored on 10 March 2026, the date of the most recent follow-up confirmation; the most recently enrolled patient began ICI therapy on 13 January 2026, giving a minimum potential follow-up of 56 days. Univariate Cox regression was performed for all candidate variables; those with P < 0.1 advanced to the next stage.

PINPS construction

Missing values were handled by multiple imputation using chained equations (MICE; m = 20, predictive mean matching) [14]. The principal imputed variables and their missing rates were: prealbumin (9.5%), clinical stage (24.0%), tumour location (20.0%), LIPI (31.9%, reflecting incomplete LDH ordering in routine practice during the earlier study period), and PIV/NLR (1.1%). Missingness was driven primarily by physician ordering patterns, supporting a missing-at-random assumption. The imputation model included all analysis variables plus the outcome (OS status and log-transformed OS time), as recommended for Cox-regression contexts. For the variable-selection imputation, the event indicator and the Nelson–Aalen estimate of the cumulative baseline hazard were included as auxiliary variables (White and Royston, Stat Med 2009;28:1982–1998), so that imputation of high-missingness candidates such as LIPI was conditioned on the observed survival outcome. PINPS coefficients were derived from a biomarker-only Cox model pooled by Rubin’s rules; the clinical-adjusted HRs in Table 2 were pooled separately from a full model including clinical covariates. Cytokine, TBNK, and four-quadrant analyses used complete-case data within their respective available subsets without additional imputation.

Table 2.

Multivariable cox proportional hazards regression for overall survival (n=525, 150 events)

Variable	HR (95% CI)	P value
Prealbumin (T0)	0.994 (0.991–0.997)	0.0001
PIV (×10³, T0)	1.511 (1.247–1.830)	<0.0001
NLR (T0)	0.946 (0.912–0.982)	0.0039
Age	1.016 (0.994–1.039)	0.1450
Sex (Female vs Male)	0.668 (0.391–1.142)	0.1392
Clinical Stage (ordinal)	1.378 (0.984–1.930)	0.0616
Upper/Cervical Location	1.498 (0.998–2.249)	0.0512

The protective direction of NLR (HR < 1) reflects a statistical suppressor effect in the presence of PIV (see Discussion)

HR hazard ratio, CI confidence interval. PIV reported per 10³ unit increment

From the 46 variables reaching univariate P < 0.1, the 11 composite indices plus prealbumin (12 candidates total) were entered into LASSO-penalised Cox regression (α = 1, 10-fold cross-validation, glmnet) [15, 16]. Raw cell counts were excluded to prevent compositional decomposition of composite indices—for example, LASSO simultaneously selecting NLR alongside its constituent neutrophil and lymphocyte counts. The penalty parameter λ was set at lambda.min; lambda.1se produced empty models across all 20 imputed datasets, consistent with the moderate individual effect sizes of these indices. LASSO was run independently on each imputed dataset. Variables selected in ≥ 50% of datasets were refit in a standard Cox model; those reaching P < 0.05 entered the final PINPS formula. Coefficients were pooled by Rubin’s rules; their variability across imputations is reported in Table S1.

Model validation

Discrimination was quantified by Harrell’s C-index [17], pooled across 20 imputed datasets and corrected for optimism with 1,000 bootstrap resamples (rms package). Apparent, pooled, and bootstrap-corrected C-indices are reported separately to allow transparent assessment of overfitting. Time-dependent area under the curve (AUC) was computed at 12, 24, and 36 months (timeROC package) [18] and verified across four imputed datasets. Calibration was assessed by bootstrap-corrected calibration plots. Clinical utility was evaluated by decision curve analysis [19]. The incremental value of PINPS over a clinical-only model (age, sex, clinical stage, tumour location) was estimated by paired-bootstrap ΔC-index with 95% confidence interval (CI) and likelihood ratio test (LRT). The proportional hazards assumption was tested with scaled Schoenfeld residuals; where marginal violations arose, extended Cox models with time-interaction terms confirmed the stability of HR estimates.

Dynamic monitoring framework

Dynamic PINPS monitoring was evaluated in patients with paired T0–T1 data. Continuous ΔPINPS was first assessed in landmark Cox regression (landmark at 42 days, ~ 2 ICI cycles) to establish whether the magnitude of change carried prognostic information. The landmark served to exclude patients who died or were lost before day 42, thereby avoiding immortal-time bias; T1 values obtained after the landmark were treated as fixed covariates, as is standard in landmark analyses where the exposure measurement window extends beyond the landmark time point. Because directional change may retain prognostic information even when continuous change magnitude does not [10], patients were further classified by a four-quadrant framework crossing baseline PINPS risk (high versus low, median cutoff) with the direction of change (decrease versus non-decrease, hereafter termed “improved” and “not improved”). Pairwise comparisons were Holm-corrected. Risk reclassification was defined as a baseline high-risk patient whose T1 PINPS fell below the median. Both an evaluable-denominator rate and an intent-to-monitor rate (imputing non-reclassification for patients without T1 data) are reported. Robustness was tested across four cutoff definitions.

Cross-dimensional correlation and biological validation

Spearman correlations between 14 nutritional/inflammatory indices and 12 cytokines plus 4 TBNK summary panels (224 pairs) were computed; P-values were corrected by the Benjamini–Hochberg false discovery rate [20]. IL-6 prognostic value was assessed by Cox regression (log2-transformed), restricted cubic splines (3 knots at P10/P50/P90) [21], and Kaplan–Meier analysis at a Youden-optimised cutoff.

TBNK analysis was pre-specified as exploratory. Consensus clustering (ConsensusClusterPlus [22]; Monti et al. [23]; 1,000 iterations, 80% resampling, PAM) was applied to patients with complete panels. Spearman correlations between PINPS and individual TBNK parameters were computed in the subset with available data.

Sensitivity analyses

Pre-planned sensitivity analyses comprised: (1) NLR–PIV collinearity assessment, including a two-variable model without NLR; (2) incremental value of advanced lung cancer inflammation index (ALI), geriatric nutritional risk index (GNRI), and monocyte-to-lymphocyte ratio (MLR) beyond PINPS (likelihood ratio test); (3) informative censoring assessment (Fisher’s exact test for mortality in patients with versus without T1 data); (4) adjustment for treatment regimen; and (5) comparison of PINPS-evaluable versus non-evaluable patients (Table S2). Two additional exploratory analyses assessed the incremental prognostic value of specialized biomarkers beyond PINPS: (6) the 12 baseline cytokines were added to the three PINPS components as candidates in a LASSO-Cox regression among 298 patients with simultaneously non-missing cytokine and PINPS data (77 events); (7) five baseline TBNK percentages (CD3+, CD3 + CD4+, CD3 + CD8+, CD3 + CD4−CD8−, CD45+) were similarly evaluated in 209 patients (34 events). In both analyses, ΔC-index relative to a PINPS-only model was computed. Temporal validation was performed by fitting the PINPS model on 2019–2024 patients (n = 253, 111 events) and evaluating risk stratification in 2025–2026 patients (n = 222, 20 events); results are reported in Table S3. Progression-free survival (PFS) was not formally analysed because natural language processing (NLP)-based response extraction achieved 85–89% accuracy, below our pre-specified 90% threshold.

Software and reporting

All analyses used R 4.4.3. Two-sided P < 0.05 was considered significant. Multiple-testing correction used Benjamini–Hochberg where indicated. Reporting followed TRIPOD guidelines [24]; a completed TRIPOD checklist is provided as Additional file 2. Supplementary figures and tables are provided in Additional file 1.

Results

Patient characteristics

Between 2019 and 2026, 549 patients with pathologically confirmed ESCC were enrolled. Twenty-four patients with incomplete follow-up were excluded, leaving 525 eligible for OS analysis (150 events; median follow-up 10.7 months, 95% CI 9.1–14.2). The study flow is shown in Fig. 1; baseline characteristics stratified by PINPS risk appear in Table 1.

Table 1.

Baseline characteristics of 475 PINPS-evaluable patients stratified by PINPS risk group

Variable	Total (N = 475)	PINPS-High (n = 238)	PINPS-Low (n = 237)	P value
Demographics
Age, years	66.1 (59.4–71.3)	68.5 (61.1–72.8)	64.0 (58.6–70.0)	< 0.001
Sex, n = 475				0.139
Female	64 (13.5)	38 (16.0)	26 (11.0)
Male	411 (86.5)	200 (84.0)	211 (89.0)
Clinical characteristics
T stage, n = 410				0.011
T1	18 (4.4)	9 (4.4)	4 (2.0)
T2	43 (10.5)	11 (5.3)	31 (15.2)
T3	121 (29.5)	63 (30.6)	58 (28.4)
T4	74 (18.0)	37 (18.0)	33 (16.2)
Tx	154 (37.6)	84 (40.8)	70 (34.3)
N stage, n = 418				0.048
N0	33 (7.9)	13 (6.2)	20 (9.6)
N1	100 (23.9)	40 (19.1)	60 (28.7)
N2	91 (21.8)	46 (22.0)	45 (21.5)
N3	80 (19.1)	48 (23.0)	32 (15.3)
Nx	114 (27.3)	62 (29.7)	52 (24.9)
M stage, n = 432				0.486
M0	147 (34.0)	69 (31.8)	78 (36.3)
M1	252 (58.3)	129 (59.4)	123 (57.2)
Mx	33 (7.6)	19 (8.8)	14 (6.5)
Clinical stage, n = 363				0.067
I-II	20 (5.5)	5 (2.7)	15 (8.3)
III	91 (25.1)	48 (26.4)	43 (23.8)
IV	252 (69.4)	129 (70.9)	123 (68.0)
Tumour location, n = 376				0.813
Cervical	42 (11.2)	19 (10.0)	23 (12.4)
Upper thoracic	61 (16.2)	29 (15.3)	32 (17.2)
Middle thoracic	141 (37.5)	77 (40.5)	64 (34.4)
Lower thoracic	96 (25.5)	46 (24.2)	50 (26.9)
Middle-lower thoracic	26 (6.9)	13 (6.8)	13 (7.0)
GEJ	10 (2.7)	6 (3.2)	4 (2.2)
Treatment regimen, n = 462				0.159
ICI+Chemo	75 (16.2)	44 (19.0)	31 (13.4)
ICI+Chemo + RT	245 (53.0)	118 (51.1)	127 (55.0)
ICI+Chemo + RT+Surgery	109 (23.6)	49 (21.2)	60 (26.0)
ICI+Chemo+Surgery	33 (7.1)	20 (8.7)	13 (5.6)
ICI drug, n = 462				-
Tislelizumab	156 (33.8)	84 (36.4)	72 (31.2)
Sintilimab	155 (33.5)	74 (32.0)	81 (35.1)
Pembrolizumab	115 (24.9)	57 (24.7)	58 (25.1)
Toripalimab	16 (3.5)	7 (3.0)	9 (3.9)
Camrelizumab	9 (1.9)	3 (1.3)	6 (2.6)
Nivolumab	9 (1.9)	5 (2.2)	4 (1.7)
Others	2 (0.4)	0 (0.0)	0 (0.0)
Nutrition indices
PNI	48.8 (45.5–52.1)	47.5 (43.9–50.3)	50.5 (47.8–53.8)	< 0.001
HALP	29.9 (18.9–43.7)	25.4 (16.6–39.6)	34.6 (22.8–47.4)	< 0.001
AAPR	0.5 (0.4–0.6)	0.5 (0.4–0.6)	0.5 (0.4–0.6)	< 0.001
ALBI	-3.0 (-3.2 to -2.8)	-2.9 (-3.1 to -2.7)	-3.1 (-3.3 to -3.0)	< 0.001
Albumin, g/L	42.7 (40.1–44.9)	40.8 (38.5–43.1)	44.5 (42.2–46.2)	< 0.001
Prealbumin, mg/L	215.0 (174.0-257.0)	177.0 (146.2–198.0)	257.0 (235.0-290.0)	< 0.001
Hemoglobin, g/L	126.0 (116.0-135.0)	121.0 (110.0-130.0)	130.0 (122.0-140.0)	< 0.001
Inflammation indices
NLR	3.6 (2.5–5.4)	3.8 (2.7–5.6)	3.4 (2.5–5.2)	0.096
PLR	176.9 (131.0-257.2)	191.2 (137.8-283.2)	170.0 (127.9–241.0)	0.025
LMR	2.6 (1.8–3.6)	2.3 (1.7–3.2)	2.8 (2.0–4.0)	< 0.001
SII	779.0 (508.7-1256.5)	864.0 (565.7-1461.7)	719.1 (474.5–1120.0)	0.002
SIRI	1.6 (1.0-2.8)	2.0 (1.2–3.2)	1.4 (0.9–2.3)	< 0.001
PIV	361.6 (202.8–625.0)	438.9 (245.6-900.7)	308.0 (165.4-491.1)	< 0.001
Laboratory markers
Neutrophil count, x10^9/L	4.4 (3.4–5.8)	4.6 (3.5-6.0)	4.1 (3.2–5.5)	0.008
Lymphocyte count, x10^9/L	1.2 (0.9–1.7)	1.2 (0.9–1.7)	1.2 (0.9–1.6)	0.863
Monocyte count, x10^9/L	0.5 (0.4–0.6)	0.5 (0.4–0.7)	0.4 (0.3–0.6)	< 0.001
WBC, x10^9/L	6.2 (5.1–7.8)	6.5 (5.2–8.1)	6.1 (4.8–7.4)	0.039
Platelet count, x10^9/L	221.0 (172.0-271.0)	233.5 (182.2–288.0)	205.0 (161.0-247.0)	< 0.001
LDH, U/L	168.0 (149.0-197.0)	166.0 (146.5–199.0)	169.0 (153.0-195.0)	0.488
Total bilirubin, umol/L	8.3 (6.0-11.3)	8.1 (6.2–10.4)	8.5 (6.0-12.1)	0.409
ALP, U/L	86.5 (71.9-102.6)	88.4 (72.3-106.2)	84.0 (71.6-100.5)	0.216
Transferrin, g/L	2.0 (1.8–2.3)	1.9 (1.7–2.2)	2.2 (1.9–2.4)	< 0.001
Immune cell subsets
CD3 + T cells, %	66.8 (56.8–74.6)	66.7 (56.2–74.1)	66.9 (58.1–74.7)	0.573
CD4 + T cells, %	48.1 (39.3–60.1)	47.9 (37.1–60.8)	48.9 (40.9–58.9)	0.360
CD8 + T cells, %	33.4 (23.8–44.2)	34.6 (25.2–47.9)	30.9 (23.5–42.8)	0.081
B cells (CD19+), %	6.6 (3.8–10.6)	6.6 (4.1–10.2)	6.4 (3.7–11.1)	0.743
Treg, %	2.5 (1.2–8.4)	2.6 (1.1–6.8)	2.5 (1.2–8.7)	0.958
Cytokines
IL-1b, n = 298	2.4 (2.4–5.4)	2.4 (2.4–4.4)	2.4 (2.4–5.6)	0.905
IL-2, n = 298	2.4 (2.4–2.7)	2.4 (2.4–2.7)	2.4 (2.4–2.7)	0.737
IL-5, n = 298	2.4 (2.4–3.6)	2.4 (2.4–3.7)	2.4 (2.4–3.6)	0.955
IL-6, n = 298	5.7 (2.4–10.8)	8.2 (4.2–17.3)	3.5 (2.4–7.6)	< 0.001
IL-8, n = 298	3.5 (2.4–16.1)	4.5 (2.4–15.4)	2.9 (2.4–16.8)	0.505
IL-17, n = 298	3.7 (2.4-7.0)	3.6 (2.4–7.1)	3.8 (2.4–6.9)	0.952
IFN-a, n = 298	2.4 (2.4–3.5)	2.4 (2.4–3.6)	2.4 (2.4-3.0)	0.564
IFN-g, n = 298	2.4 (2.4–4.5)	2.4 (2.4–4.4)	2.4 (2.4–4.6)	0.795
TNF-a, n = 298	2.4 (2.4–3.5)	2.4 (2.4–3.8)	2.4 (2.4-3.0)	0.504
Survival outcomes
Follow-up, months	7.3 (3.3–17.0)	5.7 (2.8–13.6)	9.7 (4.4–21.8)	< 0.001
OS events	131 (27.6)	76 (31.9)	55 (23.2)

Continuous variables: median (IQR), compared by Mann-Whitney U test. Categorical variables: n (%), compared by chi-square or Fisher’s exact test

Fifty patients with missing baseline prealbumin were not evaluable for PINPS. Stratification by PINPS is descriptive; differences in PINPS components (prealbumin, PIV, NLR) between risk groups are mathematical consequences of the score formula rather than independent validation. Cytokine data shown for PINPS-evaluable patients with available measurements (n = 298); the full cytokine subset (n = 325) is reported in Table S16. Three additional cytokines (IL-4, IL-10, IL-12) showed negligible distributional variability (median and IQR identical at 2.4 pg/mL in both groups) and are omitted from this table; univariate Cox results are reported in Table S16

PINPS = − 0.007181 × Prealbumin (mg/L) + 0.000355 × PIV (×10⁹/L) + (− 0.05428) × NLR (dimensionless). Cutoff (median): −1.6385

PINPS could be calculated in 475 of 525 patients; the remaining 50 lacked baseline prealbumin. Evaluable and non-evaluable patients were similar in sex, stage, and survival; evaluable patients were modestly younger (standardized mean difference 0.32). As age is included both in the multiple-imputation model and as a covariate in the multivariable Cox model, this difference does not confound the biomarker estimates (Table S2; SA-5). Cytokine data were available for 325; among 525 OS-eligible patients, baseline characteristics were similar between those with (n = 325) and without (n = 200) cytokine data (Table S4). TBNK data were available for 239 OS-eligible patients. Paired T0–T1 PINPS data existed for 382 patients; after the 42-day landmark, 365 remained for dynamic analysis. Of 238 baseline high-risk patients, 195 had follow-up measurements and 43 did not; mortality was similar in both subgroups (32.6% versus 32.0%, Fisher P = 1.0; SA-3).

Part 1: PINPS construction and validation

Variable selection

Of 46 variables reaching univariate P < 0.1 (Figure S1), the 12 candidate variables (11 composite indices plus prealbumin) were entered into LASSO. Six passed the 50% selection threshold: prealbumin (20/20 datasets), PIV (20/20), AAPR (19/20), LIPI (17/20), NLR (11/20), and ALBI (10/20) (Figure S2, Table S1). After Cox refit, three retained significance:

where PIV = neutrophils × monocytes × platelets / lymphocytes. The negative coefficient for NLR is a statistical suppressor effect arising from shared neutrophil and lymphocyte components with PIV; its interpretation and a two-variable alternative model are detailed in the Discussion and in Table S5 (SA-1). These coefficients were derived from the biomarker-only Cox model (pooled by Rubin’s rules); the clinical-adjusted hazard ratios in Table 2 are from a separate full model including age, sex, stage, and tumour location.

Independent prognostic significance

In multivariable Cox regression adjusted for age, sex, clinical stage, and tumour location (n = 525, 150 events), all three PINPS components were independently prognostic (Table 2; Fig. 2): prealbumin HR 0.994 (0.991–0.997, P = 0.0001), PIV HR 1.511 per 10³-unit increment (1.247–1.830, P < 0.0001), and NLR HR 0.946 (0.912–0.982, P = 0.004). The inverse coefficient for NLR reflects a suppressor effect: PIV and NLR share neutrophil and lymphocyte components, so that once PIV captures the dominant inflammatory signal, the residual NLR variance indexes a complementary, directionally opposite association [25, 26]. Removing PIV restored NLR to a conventional risk direction (HR 1.001) with complete loss of significance (P = 0.95; Table S5).

Fig. 2.

Multivariable cox regression forest plot for overall survival (n = 525, 150 events)

Discrimination, calibration, and clinical utility

At the median PINPS cutoff (− 1.6385, computed from the 475 PINPS-evaluable patients; the landmark cohort median was − 1.6348), Kaplan–Meier curves separated clearly (log-rank P = 0.0003; Fig. 3). The PINPS biomarker-only model yielded a bootstrap-corrected C-index of 0.667 (95% CI 0.615–0.719). The MICE-pooled C-index was 0.674, consistent with the bootstrap estimate. Time-dependent AUC was stable across horizons (12-month 0.666, 24-month 0.679, 36-month 0.669) and reproducible across four imputed datasets. Bootstrap-corrected calibration slope was 0.87 (validate, B = 1,000 resamples, complete-case n = 475). A calibration slope of 0.87 indicates mild optimism consistent with an internally validated model, addressed by the bootstrap optimism correction already applied; external validation remains necessary. Calibration curves (calibrate, B = 200) tracked the ideal line at 12 and 24 months; slight overestimation appeared at 36 months, where events were sparse (Fig. 4b).

Fig. 3.

Kaplan–Meier survival curves by PINPS risk group (high versus low, median cutoff − 1.6385). Log-rank P = 0.0003

Fig. 4.

Model validation. a Nomogram for 12-, 24-, and 36-month OS. b Bootstrap calibration curves (B = 200 resamples). c Decision curve analysis. d Time-dependent AUC for biomarker-only, clinical-only, and full models

Decision curve analysis confirmed net benefit across clinically relevant threshold probabilities (Fig. 4c). The paired-bootstrap ΔC over clinical variables was + 0.145 (95% CI 0.073–0.219, P < 0.0001; Table 3). This quantity was computed within bootstrap resamples rather than by simple subtraction of point estimates, which yields 0.122. The biomarker-only model (C-index 0.667) slightly outperformed the full model including clinical covariates (0.659), suggesting that non-prognostic clinical variables introduced noise in this cohort.

Table 3.

Predictive performance comparison of PINPS and clinical models for overall survival

Model	C-index (95% CI)	timeROC AUC @12m (95% CI)	timeROC AUC @24m (95% CI)	timeROC AUC @36m (95% CI)
Clinical only (Age + Stage + Location)	0.545*	0.600 (0.528–0.672)	0.621 (0.544–0.698)	0.604 (0.492–0.716)
PINPS only (Prealbumin + PIV + NLR)	0.667 (0.615–0.719)	0.666 (0.599–0.734)	0.679 (0.607–0.752)	0.669 (0.569–0.768)
Full model (PINPS + Clinical)	0.659 (0.610–0.709)	0.689 (0.622–0.757)	0.711 (0.641–0.782)	0.695 (0.601–0.788)
PINPS = − 0.007181 × Prealbumin (mg/L) + 0.000355 × PIV (×10⁹/L) + (− 0.05428) × NLR (dimensionless).

C-index: bootstrap-corrected (1000 resamples, optimism correction, rms package). 95% CI from bootstrap distribution

Pooled C-index = mean across 20 MICE datasets. Bootstrap-corrected C-index reported in text (0.667) uses 1000 bootstrap resamples with optimism correction (rms package). The apparent C-index from LASSO training was 0.772; the gap of 0.105 reflects the expected degree of optimism in a regularised model fitted to this sample size (525 patients, 150 events). A reduced two-variable model (prealbumin + PIV, omitting NLR) yielded an optimism-corrected C-index of 0.680; the value of 0.682 corresponds to the same two biomarkers with clinical covariates (complete-case, n = 363)

timeROC AUC: Blanche et al. 2013 [18], iid=TRUE. Based on MICE imputation #1 (n = 525)

*MICE pooled C-index; bootstrap correction not performed for clinical-only model

Schoenfeld residual testing showed marginal PH violations for PIV (P = 0.044) and sex (P = 0.011). An extended Cox model with time-interaction terms retained the HR direction of all three PINPS components with < 2% magnitude change. NLR’s P-value widened from 0.006 to 0.078 owing to degrees-of-freedom competition with the interaction terms, but the coefficient remained stable (Table S6).

Part 2: Dynamic PINPS monitoring and risk reclassification

On-treatment changes in inflammatory and nutritional indices

The following dynamic analyses are exploratory extensions of the primary PINPS model developed in Part 1.

All nine inflammation/nutrition indices shifted significantly from T0 to T1 (paired Wilcoxon P < 0.01 each; n = 275–455 depending on index availability; Table S7, Figure S3). Cytokine dynamics (n = 151 paired) and TBNK changes (n = 80, exploratory) are shown in Figures S4–S5 and Table S8.

Continuous ΔPINPS: null result

In landmark analysis (n = 365, 105 events), continuous ΔPINPS showed no association with OS (HR 0.970, 95% CI 0.686–1.373, P = 0.864). There was no evidence of nonlinearity (restricted cubic spline [RCS] P = 0.85), and time-dependent AUC hovered near chance (12-month 0.542; Figure S6, Table S9). The magnitude of PINPS change, when analysed as a continuous variable, was not associated with survival in this cohort (Table S10; individual-index delta results in Table S11; tertile KM in Figure S7).

Four-quadrant framework

We then tested whether the direction of change, crossed with baseline risk, could separate prognostic subgroups. As with the continuous ΔPINPS analysis, the four-quadrant Kaplan–Meier analysis was conducted in the landmark cohort (n = 365), with survival measured from the 42-day landmark to avoid immortal-time bias introduced by the post-baseline group assignment. The four-quadrant framework stratified OS (log-rank P = 0.0032, n = 365; Fig. 5a; Table 4):

Fig. 5.

Four-quadrant dynamic framework (landmark analysis). a Kaplan–Meier curves by baseline risk × direction of change (P = 0.0032, n = 365; survival measured from 42-day landmark). b Sankey diagram of risk reclassification from T0 to T1 to survival outcome

Table 4.

Four-quadrant framework: baseline PINPS risk × direction of on-treatment change (n = 365)

Group	n	Events	Median OS (months)	Mean PINPS (T0)	Mean ΔPINPS
Low+Improved	95	23	36.8	-1.9332	-0.3023
Low + Not Improved	87	24	34.0	-2.1167	0.4012
High+Improved	126	38	19.6	-1.1695	-0.5672
High + Not Improved	57	20	18.3	-1.2903	0.4361

PINPS cutoff (median): −1.6348. Improved: PINPS decreased (T1 < T0); Not Improved: PINPS non-decreased. All survival times are measured from the 42-day landmark, not from treatment initiation

Log-rank P = 0.0032 across all four groups. See Fig. 5a for Kaplan-Meier curves

• Low-risk + improved (i.e., PINPS-decreasing) (n = 95): best prognosis (median OS 36.8 months from landmark)

• Low-risk + not improved (i.e., PINPS non-decreasing) (n = 87): intermediate (34.0 months)

• High-risk + improved (n = 126): intermediate (19.6 months)

• High-risk + not improved (n = 57): worst prognosis (18.3 months)

After Holm correction, the persistent-high group separated from both the Low+Improved (P = 0.013) and Low + Not Improved (P = 0.013) subgroups; intermediate comparisons did not survive correction. Because this framework was constructed after the continuous ΔPINPS analysis yielded a null result, it should be regarded as exploratory.

Risk reclassification

Among 183 baseline high-risk patients in the landmark cohort, 69 (37.7%) crossed below the median cutoff after treatment. When all 238 baseline high-risk patients were considered (imputing non-reclassification for those without follow-up data), the intent-to-monitor reclassification rate was 29.0% (69/238). Reclassified patients showed a trend toward better OS (HR 0.603, 95% CI 0.345–1.053, P = 0.075; Table S12). The Sankey diagram in Fig. 5b illustrates the flow from baseline risk through reclassification to outcome; analogous Sankey diagrams for individual indices (NLR, SII, PNI) are shown in Figure S8.

Sensitivity analyses

The four-quadrant framework was stable across alternative cutoffs: median (P = 0.0032), Youden at 12 months (P = 0.0001), 40th percentile (P < 0.0001), and outlier exclusion (P = 0.0031). Reclassification rates ranged from 33% to 39% (Table S13).

To isolate NLR’s incremental contribution, we compared a two-variable model (prealbumin + PIV + clinical covariates [age, sex, clinical stage, tumour location]) with the three-variable PINPS model (+ NLR), in the complete-case subset with all covariates available (n = 363, 107 events). The addition of NLR was small but statistically significant (LRT P = 0.005); the corresponding optimism-corrected C-indices were 0.682 and 0.695 respectively (rms::validate, B = 500; Monte-Carlo bootstrap variation ≈ ± 0.001). These C-indices exceed the biomarker-only estimate (0.667) because clinical covariates are included in both; the comparison is designed to isolate NLR’s marginal contribution under otherwise identical conditions. Removing PIV entirely restored NLR to a risk direction (HR 1.001, P = 0.95; Table S5), confirming the suppressor mechanism (SA-1). Notably, NLR’s increment reached significance with clinical covariates included (n = 363) but not in the biomarker-only paired comparison (n = 475; see Discussion); we retain the three-variable score because the dynamic four-quadrant framework is built upon it.

Adjusting for treatment regimen changed all component HRs by less than 0.2% (Table S14; SA-4). ALI, GNRI, and MLR added no incremental value beyond PINPS (all LRT P > 0.05; Table S15; SA-2). As a supplementary temporal validation (intended as a signal check rather than formal external validation), the PINPS model trained on 2019–2024 patients (n = 253, 111 events) maintained significant risk stratification in the 2025–2026 testing cohort (n = 222, 20 events; HR 8.15, 95% CI 1.89–35.20, log-rank P = 0.0008; Table S3). All three PINPS coefficients maintained consistent direction in the training subset, although NLR reached only borderline significance (P = 0.051; Table S3 Panel C). Limited events in the testing cohort precluded reliable C-index estimation.

Part 3: IL-6 as a cross-dimensional inflammatory hub

Cross-dimensional correlations

Of 224 nutritional/inflammatory–cytokine/TBNK correlations, 14 survived Benjamini–Hochberg correction and all involved IL-6 (Figure S9); no other cytokine bridged both dimensions. IL-6 correlated inversely with nutritional markers (ρ − 0.31 to − 0.42) and positively with inflammatory indices (ρ + 0.26 to + 0.34). As patients with cytokine data had shorter mean OS than those without (10.4 vs. 15.5 months, P < 0.001; Table S4), these correlations are exploratory.

IL-6 and survival

In 325 patients, log2 IL-6 predicted OS (HR 1.262 per doubling, 95% CI 1.079–1.477, P = 0.004; linear on the log-hazard scale; Table S16, Figures S10–S11) but lost independent significance after adjustment for PINPS (LRT P = 0.69), indicating that PINPS already captures most of IL-6’s prognostic information. A dichotomised analysis (8.17 pg/mL cutoff) was concordant (HR 2.307, 95% CI 1.463–3.640, P = 0.0003; Fig. 6). These findings are exploratory.

Fig. 6.

Kaplan–Meier curves by baseline IL-6 (Youden cutoff 8.17 pg/mL). Log-rank P = 0.0002. Exploratory

Part 4: TBNK immunophenotyping (exploratory)

Consensus clustering of 162 patients with complete TBNK panels (k = 3) identified three subtypes (n = 90/59/13; Fig. 7a) with no OS difference (P = 0.528; Figures S12–S13). Events per variable were too few (≤ 4) for multivariable modelling.

Fig. 7.

TBNK analysis. a Heatmap of 28 baseline TBNK subset percentages (rows) across 162 patients with complete panels (columns), z-score normalised. Hierarchical clustering (Euclidean distance, Ward linkage) applied to both rows and columns. Column annotation bar: OS event status (red = death, blue = alive/censored). Consensus clustering diagnostic plots are provided in Supplementary Materials. b Forest plot of univariate Cox regression for four TBNK parameters reaching P < 0.05 (n = 162; full results for all 28 parameters in Table S17)

Four TBNK variables reached univariate significance (Fig. 7b, Table S17). PINPS correlated positively with Th1 frequency (ρ=+0.239, P = 0.003, n = 154; Table S18, Figure S14) but not Th17 (ρ=−0.129, P = 0.112). These findings remain exploratory.

Incremental value of cytokine and TBNK biomarkers

To assess whether specialized biomarkers provided incremental prognostic value beyond PINPS, exploratory LASSO-Cox analyses were performed in two subsets. Among 298 patients with complete baseline cytokine and PINPS component data (77 events), none of the 12 cytokines were retained by LASSO when the three PINPS components were included as candidates (ΔC-index: −0.001; Table S19; SA-6). Similarly, among 209 patients with complete TBNK and PINPS data (34 events), none of five major T-cell subset markers were selected (ΔC-index: −0.011; Table S20; SA-7). In both analyses, LASSO retained only prealbumin and PIV; NLR was not selected in either subset, consistent with its marginal stability in the full-cohort analysis (55% selection frequency). These results indicate that routine blood-derived indices already capture the core prognostic information for OS in this cohort, and that specialized cytokine panels or flow cytometry assays do not provide independent incremental predictive value.

Discussion

In 549 ESCC patients receiving PD-1/PD-L1 inhibitors, PINPS—a three-biomarker composite spanning nutritional (prealbumin), pan-inflammatory (PIV), and inflammatory-immune (NLR) domains—achieved a bootstrap-corrected C-index of 0.667 and a ΔC of + 0.145 over clinical staging. An exploratory four-quadrant framework that crossed baseline risk with the direction of on-treatment change identified a persistent-high subgroup with the worst outcomes. Over one-third of initially high-risk patients were reclassified during treatment. IL-6 emerged as the sole cytokine linking the nutritional and inflammatory arms of the score. Preliminary TBNK profiling suggested an association between PINPS and Th1 frequency.

Discrimination in context

A C-index of 0.667 warrants frank appraisal. The statistic has documented ceiling effects in survival settings: Hartman et al. showed that it is structurally dominated by comparisons among patients at similar risk, understating discriminative capacity even when prognostic effects are substantial [27]. In practice, 0.66–0.70 is the range reported for most published immune-nutritional composites. PINPS exceeds TNM staging alone (C-index approximately 0.55–0.65 in comparable advanced ESCC populations) [28]. Decision curve analysis, which arguably better reflects clinical decision-making, confirmed net benefit across a range of threshold probabilities. That the biomarker-only model slightly outperformed the full model (0.667 versus 0.659) fits a recognised pattern: adding weakly prognostic covariates to a penalised-regression-derived score introduces noise rather than signal. The universal failure of lambda.1se across all 20 imputed datasets further reflects the moderate individual effect sizes of these indices; no single index exceeded the stricter regularisation threshold, confirming that prognostic value arises from their combination rather than from any single dominant marker.

Comparison with existing tools

Existing indices—NLR [5], SII [7], PNI [8], LIPI [9]—each operate along a single biological axis. PINPS integrates across axes. Prealbumin, with a 2-day half-life, is a sensitive indicator of acute nutritional decline [12]. PIV captures all four peripheral immune cell populations in one metric [11]. NLR contributes a small but statistically verified incremental refinement. The four-quadrant dynamic framework extends the approach of Matsumoto et al. [10], who paired baseline NLR with directional change in a three-tiered classification for NSCLC. PINPS uses a LASSO-derived composite rather than a single ratio and retains all four cells of the baseline × direction matrix rather than collapsing intermediate groups.

Among ESCC-specific composites, the recently reported IIPS (NEU×MON×PLT/[LYM×ALB]) predicted OS (HR 1.864, P = 0.033) in 205 ESCC patients receiving neoadjuvant immunochemotherapy [29], and a nomogram incorporating BMI and dynamic inflammatory markers achieved a PFS C-index of 0.70 in 212 ESCC patients receiving chemoradiotherapy combined with anti-PD-1 therapy [30]. PINPS achieved a comparable bootstrap-corrected C-index (0.667) in a considerably larger cohort (n = 525) and additionally incorporated a nutritional dimension (prealbumin) and a dynamic monitoring framework absent from these comparators.

Importantly, exploratory LASSO-Cox analyses showed that neither 12 baseline cytokines (n = 298, 77 events) nor 5 major TBNK subsets (n = 209, 34 events) provided incremental discrimination beyond the three PINPS components (ΔC-index: −0.001 and − 0.011, respectively; Tables S19–S20). This finding reinforces the clinical accessibility advantage of PINPS: comparable prognostic information is captured by three routine blood tests without requiring specialized cytokine or flow cytometry assays. It also supports the interpretation that the IL-6 and TBNK associations reported in Parts 2 and 3 reflect mechanistic correlates of the same underlying immune-nutritional axis rather than independent prognostic dimensions.

Dynamic monitoring and the direction-versus-magnitude dissociation

A dissociation emerged between continuous ΔPINPS (HR 0.970, P = 0.864) and the categorical four-quadrant stratification (P = 0.0032). This echoes the Matsumoto observation for NLR [10] and points to a threshold-dependent prognostic signal: what matters is whether a patient’s immune-nutritional status crosses a clinically meaningful boundary, not how far it moves. The four-quadrant analysis followed a null continuous result and should be treated as hypothesis-generating.

After Holm correction, only the persistent-high group (High + Not Improved, n = 57) separated from the best-prognosis and low-risk groups; intermediate cells did not differ from each other. The framework’s clinical value is concentrated in identifying the worst-prognosis patients—those who might benefit from early treatment modification or intensified nutritional support—rather than in grading intermediate-risk categories.

NLR suppressor effect

The negative NLR coefficient (HR 0.946) requires transparent discussion. In isolation, higher NLR would paradoxically lower the PINPS score. The mechanism is straightforward: PIV and NLR share neutrophil and lymphocyte components. Once PIV absorbs the bulk of the inflammatory signal, the residual NLR variance captures a complementary, directionally opposite association. Removing PIV from the model restores NLR to a conventional risk direction but with no residual significance (HR 1.001, P = 0.95). This phenomenon—termed net suppression—is well characterised in the methodological literature [25, 26, 31–33]. Paulhus et al. classified it as a situation where the sign of a regression coefficient reverses relative to its zero-order correlation [33]. Tu et al. placed it within the broader family of reversal paradoxes, arguing that the adjusted or unadjusted estimate should be preferred according to the causal structure rather than statistical convention [26]. Critically, Kim demonstrated that suppressor variables genuinely improve predictive accuracy in prediction models and should be retained, whereas they should be excluded from explanatory models seeking unbiased causal estimates [32]. Since PINPS is a prediction tool rather than a causal model, retention of NLR is consistent with this framework. A biomarker-only two-variable model (prealbumin + PIV, omitting NLR) had an optimism-corrected C-index of 0.680 when estimated on its own. In a head-to-head paired comparison restricted to the 475 patients with complete data for both models, three-variable PINPS and the two-variable model discriminated almost identically (0.671 and 0.665, respectively; paired difference + 0.006 in favour of PINPS, 95% CI − 0.015 to + 0.021, not significant); the small numerical differences from the standalone estimates arise because the paired contrast is computed on the common patient set with a shared bootstrap, and the two sets of figures are therefore not directly comparable. The previously cited value of 0.682 corresponds to a model that additionally included clinical covariates (complete-case, n = 363) and is not directly comparable. Within any single specification the addition of NLR did not reduce discrimination (biomarker-only increment ΔC ≈ + 0.006; Supplementary Table S23), and the two-variable model remains a clinically viable alternative for users prioritising interpretive simplicity. PINPS is designed to be evaluated at the composite-score level; users need not interpret individual coefficient signs, much as clinicians using the Framingham Risk Score do not decompose its individual β-weights. NLR was the least stable component (selected in 55% of LASSO imputations versus 100% for prealbumin and PIV), and its biomarker-only marginal discrimination increment of ΔC ≈ + 0.006 should be weighed accordingly. Despite its prevalence in collinear biomarker settings, suppressor effects are rarely identified explicitly in oncology prediction models. To our knowledge, PINPS is among the first immune-nutritional prognostic scores to report and formally validate this phenomenon through symmetric model comparisons.

Biological correlates

IL-6 was the only cytokine bridging the nutritional and inflammatory dimensions of PINPS. Its inverse correlation with prealbumin (ρ=−0.42) and positive correlation with PIV (+ 0.34) fit its known role in hepatic acute-phase protein regulation and cancer-associated cachexia [34–36]. That IL-6 lost independent prognostic significance after PINPS adjustment (LRT P = 0.69) is compatible with IL-6 acting as an upstream mediator of the nutritional–inflammatory axis that PINPS quantifies, rather than contributing additional predictive information. Cross-sectional correlations cannot establish causal directionality.

The TBNK analysis revealed a positive correlation between PINPS and Th1 frequency (ρ=+0.239, P = 0.003), while the PINPS–Th17 association was weak and non-significant. The Th1 finding is counterintuitive: higher inflammatory burden might be expected to suppress anti-tumour Th1 responses. Possible explanations include compensatory Th1 upregulation under inflammatory stress, or non-specific Th1 expansion that does not translate into effective anti-tumour immunity [37]. With only 17 events in the complete-panel TBNK subset, these findings are best treated as hypothesis-generating.

Limitations

Several limitations bear mention. First, this is a single-centre study without external validation; the bootstrap-corrected C-index provides only a preliminary estimate of transportable performance. As a supplementary sensitivity analysis, temporal validation using 2019–2024 patients for training and 2025–2026 patients for testing confirmed that PINPS risk stratification remained effective in recently treated patients (HR 8.15, 95% CI 1.89–35.20, Cox–Wald P = 0.005; Table S3); however, the validation cohort accrued only 20 events during limited follow-up, precluding reliable C-index estimation and underscoring the need for longer-term follow-up or independent external cohorts. Second, the four-quadrant framework was developed post hoc and requires prospective confirmation. Third, although T1 sampling dates were not pre-coded in the locked analysis dataset, we recovered them post hoc from source laboratory records for all 365 landmark-cohort patients, with patient and draw identity confirmed by value matching; sampling clustered early (median 25 days, IQR 23–37; Supplementary Figure S15), and a sensitivity analysis confirmed that the four-quadrant findings were directionally consistent across early (21–42 d) and late (43–90 d) T1 sampling (Supplementary Table S22), though the late stratum was underpowered (n = 55, 11 events). Fourth, the 36-month timeROC AUC (0.669) rests on limited late events and should be interpreted cautiously. Fifth, NLR was the least stable PINPS component (55% LASSO selection), and its reversed coefficient direction demands transparent communication in any future clinical application. Finally, PFS was not analysed because NLP extraction accuracy (85–89%) fell below the pre-specified 90% threshold. The collapse of the penalised model to the null at lambda.1se reflects a modest global signal-to-noise ratio; we therefore adopted lambda.min for the primary model. Data required to compute LIPI were missing in 31.9% of patients; these patients differed systematically from those with complete data (Supplementary Table S21), indicating a missing-at-random rather than a missing-completely-at-random mechanism, which our multiple-imputation model was designed to address.

Conclusions

PINPS, composed of prealbumin, PIV, and NLR, provides survival stratification in ICI-treated ESCC from three routine blood markers that warrants external validation before clinical application. An exploratory four-quadrant framework identifies a persistent-high subgroup with the worst prognosis and captures on-treatment risk reclassification in over one-third of initially high-risk patients; this framework in particular requires prospective confirmation. IL-6 serves as the cytokine bridge between the nutritional and inflammatory components of PINPS, and preliminary TBNK data suggest a link between the score and Th1 frequency. These findings remain exploratory; external validation in independent, ideally multi-centre, cohorts is the necessary next step.

Abbreviations

AAPR

Albumin-to-alkaline phosphatase ratio

ALBI

Albumin–bilirubin index

ALI

Advanced lung cancer inflammation index

AUC

Area under the curve

CI

Confidence interval

DCA

Decision curve analysis

dNLR

Derived neutrophil-to-lymphocyte ratio

ESCC

Esophageal squamous cell carcinoma

FDR

False discovery rate

GNRI

Geriatric nutritional risk index

HR

Hazard ratio

ICI

Immune checkpoint inhibitor

IL-6

Interleukin-6

LASSO

Least absolute shrinkage and selection operator

LDH

Lactate dehydrogenase

LIPI

Lung immune prognostic index

LMR

Lymphocyte-to-monocyte ratio

MICE

Multiple imputation by chained equations

MLR

Monocyte-to-lymphocyte ratio

NLR

Neutrophil-to-lymphocyte ratio

OS

Overall survival

PFS

Progression-free survival

PINPS

Prealbumin–Immune-Inflammation–Neutrophil-to-Lymphocyte Ratio Prognostic Score

PIV

Pan-immune-inflammation value

PLR

Platelet-to-lymphocyte ratio

PNI

Prognostic nutritional index

SII

Systemic immune-inflammation index

SIRI

Systemic inflammation response index

TBNK

T-cell B-cell NK-cell

Th

T-helper

Treg

Regulatory T cell

Authors’ contributions

Huiping Ding and Shengnan Hao are co-first authors. Yiqun Ling, Kuaile Zhao, and Gengsheng He are co-corresponding authors. Huiping Ding: Conceptualization, Data curation, Formal analysis, Software, Visualization, Writing – original draft. Shengnan Hao: Data curation, Investigation, Resources, Validation. Gengsheng He: Methodology, Supervision, Writing – review & editing. Kuaile Zhao: Conceptualization, Project administration, Supervision, Writing – review & editing. Yiqun Ling: Supervision, Writing – review & editing. All authors read and approved the final manuscript.

Funding

This study received no external funding.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Institutional Review Board of Fudan University Shanghai Cancer Center (Ethics Approval No. 2603-Exp392). The requirement for individual informed consent was waived by the Institutional Review Board due to the retrospective nature of the study. This study was conducted in accordance with the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.