Predicting PD-L1⁺ CD8⁻ status in non-small cell lung cancer tissue from clinical indicators using machine learning

Abstract

Background:

Programmed death-ligand 1 (PD-L1) expression and CD8-positive (CD8⁺) T-cell infiltration in tumor tissue are associated with prognosis in non-small cell lung cancer (NSCLC). However, the prognostic value of combined PD-L1/CD8 immune phenotyping in surgically treated NSCLC and the feasibility of predicting high-risk immune phenotypes using routine clinical indicators remain unclear. This study aimed to evaluate the prognostic significance of PD-L1/CD8 status and to develop machine learning models for predicting PD-L1-positive/CD8-negative (PD-L1⁺ CD8⁻) status.

Methods:

This study included 844 patients with NSCLC who underwent surgical resection at Wuhan Union Hospital and Renmin Hospital of Wuhan University between March 2012 and November 2022. PD-L1 expression and CD8⁺ T-cell infiltration were assessed by immunohistochemistry, and preoperative clinical data were collected. Kaplan–Meier analysis, stratified survival analysis, univariable and multivariable Cox proportional hazards regression models were used to evaluate the association between PD-L1/CD8 status and overall survival (OS). Patients from Wuhan Union Hospital were divided into training, test, and internal validation sets, whereas patients from Renmin Hospital of Wuhan University were used as an external validation set. Multiple machine learning models were developed to predict PD-L1⁺ CD8⁻ status. Feature importance and SHapley Additive exPlanations (SHAP) analysis were used to interpret model predictions.

Results:

Based on PD‑L1/CD8 status, tumors from the 844 patients were classified into four immune phenotypes. Kaplan–Meier analyses revealed significant differences in OS among the four PD-L1/CD8 subgroups, with the PD-L1⁺ CD8⁻ subgroup displaying the poorest survival. This pattern was also observed in several stratified analyses. Multivariable Cox regression analysis further demonstrated that PD-L1⁺ CD8⁻ status was independently associated with worse OS compared with PD-L1⁺ CD8⁺ status (hazard ratio = 3.261, 95% confidence interval: 1.310–8.116, P = 0.011). On this basis, machine learning models were developed to predict PD-L1⁺ CD8⁻ status. Among the evaluated models, the stacking model showed the best overall performance, with area under the curve values of 0.998, 0.700, 0.853, and 0.783 in the training, test, internal validation, and external validation sets, respectively. Feature importance and SHAP analyses suggested that PD-L1⁺ CD8⁻ status was associated with a composite pattern involving hematologic, biochemical, inflammatory, and coagulation-related indicators.

Conclusions:

PD-L1/CD8-based phenotyping identified prognostically distinct subgroups in surgically treated NSCLC, with PD-L1⁺ CD8⁻ status defining a high-risk phenotype associated with poor survival. A machine learning model based on clinical indicators may enable efficient, noninvasive identification of PD-L1⁺ CD8⁻ status and help guide treatment decisions.

Introduction

Lung cancer remains one of the most common malignancies and the leading cause of cancer-related death worldwide.^[1] Histologically, lung cancer is broadly classified into small cell lung cancer (SCLC) and non-small cell lung cancer (NSCLC), with NSCLC accounting for the majority of cases.^[2] For patients with resectable NSCLC, surgical resection remains a cornerstone of curative-intent treatment. However, postoperative outcomes remain heterogeneous, even among patients with similar clinicopathological stages, highlighting the need for improved biomarkers and risk stratification strategies.^[3] Many patients develop resistance to the current treatments, leading to disease progression.^[4] Programmed death-ligand 1 (PD-L1) expression is widely used as a biomarker for immune checkpoint inhibitor therapy, but its prognostic and predictive value remains imperfect because PD-L1 expression is dynamically regulated, spatially heterogeneous, and influenced by both tumor-intrinsic and immune-mediated mechanisms.^[5–10] In addition to PD-L1 expression, CD8 positive (CD8⁺) T-cell infiltration reflects local antitumor immune activity and may provide prognostic value.^[11] Thus, a better understanding of the combined roles of PD-L1 expression and CD8⁺ T-cell infiltration in tumor tissue may have important clinical implications.

Clinical indicators provide biologically interpretable foundations for predictive models by quantifying physiological status, disease progression, and therapeutic responses. For example, routine clinical parameters such as leukocyte count, hemoglobin, and serum albumin, combined with demographic features including age and body mass index, have been used to predict immune checkpoint inhibitor responses across 21 cancer types, and showed predictive performance exceeding that of several established biomarkers, including PD-L1 immunostaining, in selected validation settings.^[12] Prediction models based on routinely available clinical and biochemical indicators have been applied in disease risk assessment; for example, longitudinal changes in fibrosis markers have been associated with the risk of cirrhosis and hepatocellular carcinoma in patients with non-alcoholic fatty liver disease.^[13] These studies underscore the importance of clinical indicators, particularly when integrated with advanced analytic methods such as machine learning, for personalized medical decision-making.

Machine learning has emerged as a transformative tool for medical prediction.^[14–16] A recently developed pathology foundation model trained on large-scale pathology images showed strong performance across multiple cancer diagnostic and prognostic tasks.^[14] In lung adenocarcinoma, integrated single-cell and bulk transcriptomic analyses, together with machine learning-based modeling, have been used to identify lactate metabolism-related genes and construct a prognostic risk stratification model.^[17] To facilitate clinical translation, interpretable methods such as least absolute shrinkage and selection operator (LASSO) regression and decision tree visualization can explicitly quantify the contribution of biomarkers.^[18] Together, these advances highlight the pivotal role of machine learning in modern oncology research.

In this study, we evaluated PD-L1/CD8 status in surgically resected NSCLC tissues using immunohistochemistry (IHC) and examined its association with overall survival (OS). We further developed prediction models to identify PD-L1-positive/CD8-negative (PD-L1⁺ CD8⁻) status, which appeared to be most closely related to poor prognosis.

Methods

Study design and data sources

This study included 844 patients with NSCLC who underwent surgical treatment at the Department of Thoracic Surgery of Wuhan Union Hospital or Renmin Hospital of Wuhan University between March 2012 and November 2022 [Figure 1]. The inclusion criteria were as follows: (1) pathologically confirmed newly diagnosed NSCLC; (2) sufficient tumor tissue available for PD-L1/CD8 IHC evaluation; (3) no preoperative anti-tumor treatment; (4) age ≥18 years; and (5) complete key clinical data. The exclusion criteria were as follows: (1) history of other active malignancies or synchronous multiple primary tumors; (2) metastatic lung tumors or uncertain diagnosis of primary NSCLC; (3) inadequate tissue quality or unsuccessful/unevaluable IHC staining; (4) receipt of neoadjuvant anti-tumor therapy; and (5) missing key clinicopathological data. All 844 eligible patients were included in the PD-L1/CD8 classification analysis, baseline characteristic analysis, and machine learning model development. The study was approved by the Ethics Committees of Wuhan Union Hospital (No. S363) and Renmin Hospital of Wuhan University (No. WDRY2023-K004). The requirement for informed consent was waived because of the retrospective design and use of de-identified samples, in accordance with the Declaration of Helsinki.

Figure 1.

Flowchart of patient enrollment and study design for 1254 patients with surgically treated NSCLC. This flowchart shows the retrospective screening of 1254 patients with NSCLC who underwent surgical resection at Wuhan Union Hospital and Renmin Hospital of Wuhan University between March 2012 and November 2022. Patients with multiple exclusion criteria were counted according to the primary reason for exclusion. A total of 410 patients were excluded according to predefined criteria. Survival analysis was performed in 298 patients with available follow-up data. For machine-learning analyses, the Wuhan Union Hospital cohort (n = 546) was divided into training, test, and internal validation sets, whereas the Renmin Hospital of Wuhan University cohort (n = 298) served as the external validation set. IHC: Immunohistochemistry; NSCLC: Non-small cell lung cancer; PD-L1: Programmed death-ligand 1.

For survival analyses, patients were followed up every 1–3 months until death or the last follow-up date of March 13, 2025, whichever occurred first. OS was defined as the time from the date of surgery to death from any cause; patients who were alive at the last follow-up were censored on that date. Follow-up data were available for 298 patients, among whom 86 all-cause deaths were recorded. Among the 844 enrolled patients, 546 (64.69% [546/844]) were lost to follow-up because they could not be contacted or declined to provide follow-up information.

Demographic data, preoperative laboratory tests, radiologic findings, pathology results, tumor stage, and tumor grade were retrospectively extracted from hospital records. Preoperative laboratory tests included complete blood count, coagulation profile, and renal and liver function tests. To ensure that PD-L1 and CD8 expression reflected the baseline tumor immune status, only patients who had not received any preoperative anti-tumor therapy were included. Postoperative anti-tumor treatments, including chemotherapy, radiotherapy, immunotherapy, and targeted therapy, were recorded and considered in survival analyses. Consequently, most patients were in the early or intermediate stages of NSCLC, as shown in the baseline characteristics [Table 1]. All patients were staged as M0 (no metastasis); therefore, M stage was not included in subsequent analyses.

Table 1.

Clinicopathological characteristics according to PD-L1/CD8 status in patients with non-small cell lung cancer.

Characteristics	Group (N = 844)				Test statistics	P values
	PD-L1+ CD8+(N = 116)	PD-L1+ CD8−(N = 138)	PD-L1− CD8−(N = 374)	PD-L1− CD8+(N = 216)
Sex					3.316*	0.345
Male, n (%)	77 (66.38)	79 (57.25)	215 (57.49)	124 (57.41)
Female, n (%)	39 (33.62)	59 (42.75)	159 (42.51)	92 (42.59)
Age (years), median (IQR)	62.0 (55.0, 67.0)	60.0 (54.0, 67.0)	61.0 (55.0, 66.0)	60.0 (54.3, 67.8)	0.895†	0.827
Smoking					6.130*	0.105
Yes, n (%)	61 (52.59)	60 (43.48)	151 (40.37)	101 (46.76)
No, n (%)	55 (47.41)	78 (56.52)	223 (59.63)	115 (53.24)
White blood cell count (×109/L), median (IQR)	6.1 (5.2, 7.6)	6.1 (5.0, 7.5)	5.5 (4.6, 7.0)	5.8 (4.8, 7.0)	8.417†	0.038
Red blood cell count (×1012/L), median (IQR)	4.2 (3.9, 4.5)	4.3 (3.9, 4.7)	4.2 (3.9, 4.6)	4.3 (3.9, 4.7)	4.053†	0.256
Hemoglobin (g/L), median (IQR)	130.0 (121.0, 140.0)	131.0 (122.3, 144.3)	130.0 (120.0, 143.0)	133.0 (123.0, 144.0)	4.413†	0.220
Hematocrit (%), median (IQR)	38.5 (36.0, 41.8)	38.6 (36.5, 42.3)	38.3 (35.7, 41.6)	39.1 (36.4, 42.5)	4.658†	0.199
MCV (fL), median (IQR)	91.4 (87.3, 96.7)	91.4 (86.9, 94.7)	91.2 (86.8, 95.9)	91.7 (86.2, 96.3)	0.771†	0.856
MCH (pg), median (IQR)	30.6 (29.4, 32.5)	30.5 (29.1, 31.7)	31.0 (29.5, 32.2)	31.0 (29.6, 32.0)	4.021†	0.259
MCHC (g/L), median (IQR)	335.0 (326.6, 344.9)	335.6 (327.4, 343.7)	336.3 (326.4, 346.4)	336.0 (321.4, 350.0)	0.526†	0.913
Platelet count (×109/L), median (IQR)	216.0 (181.0, 265.0)	205.5 (165.0, 244.0)	205.0 (167.0, 247.0)	194.0 (163.5, 240.0)	7.078†	0.069
Neutrophil percentage (%), median (IQR)	61.7 (55.5, 67.4)	61.3 (54.1, 66.4)	58.7 (52.3, 66.3)	60.2 (52.9, 68.5)	5.166†	0.160
Lymphocyte percentage (%), median (IQR)	26.2 (21.1, 32.0)	27.9 (22.5, 34.4)	30.7 (23.0, 36.2)	27.6 (21.8, 36.1)	10.020†	0.018
Monocyte percentage (%), median (IQR)	8.1 (6.9, 9.5)	7.6 (6.5, 9.1)	7.6 (6.4, 9.0)	7.5 (6.2, 9.3)	5.866†	0.118
Eosinophil percentage (%), median (IQR)	2.4 (1.3, 3.6)	2.1 (1.1, 3.4)	1.9 (1.3, 2.8)	2.1 (1.3, 3.5)	5.630†	0.131
Basophilic granulocyte percentage (%), median (IQR)	0.4 (0.2, 0.7)	0.4 (0.3, 0.6)	0.4 (0.3, 0.6)	0.3 (0.2, 0.6)	4.461†	0.216
Neutrophil count (×109/L), median (IQR)	3.7 (3.0, 4.8)	3.6 (2.8, 4.5)	3.2 (2.5, 4.3)	3.4 (2.7, 4.5)	10.576†	0.014
Lymphocyte count (×109/L), median (IQR)	1.6 (1.2, 2.0)	1.6 (1.3, 2.1)	1.6 (1.3, 2.0)	1.6 (1.3, 2.0)	0.878†	0.831
NLR, median (IQR)	2.4 (1.7, 3.2)	2.2 (1.6, 2.8)	1.9 (1.5, 2.9)	2.2 (1.5, 3.1)	8.633†	0.035
Monocyte count (×109/L), median (IQR)	0.5 (0.4, 0.6)	0.5 (0.3, 0.6)	0.4 (0.3, 0.5)	0.5 (0.4, 0.6)	19.395†	<0.001
Eosinophil count (×109/L), median (IQR)	0.1 (0.1, 0.2)	0.1 (0.1, 0.2)	0.1 (0.1, 0.2)	0.1 (0.1, 0.2)	8.770†	0.033
Basophilic granulocyte count (×109/L), median (IQR)	0.03 (0.00, 0.04)	0.02 (0.00, 0.04)	0.02 (0.00, 0.03)	0.02 (0.00, 0.04)	7.588†	0.055
Alanine aminotransferase (U/L), median (IQR)	19.0 (15.0, 26.0)	22.0 (16.0, 28.0)	19.0 (14.0, 25.0)	20.0 (15.0, 27.0)	7.067†	0.070
Aspartate aminotransferase (U/L), median (IQR)	20.0 (17.0, 26.0)	20.0 (16.0, 27.0)	21.0 (15.0, 28.0)	19.0 (15.2, 27.0)	1.567†	0.667
Alkaline phosphatase (U/L), median (IQR)	75.5 (62.0, 91.4)	72.0 (60.9, 88.0)	71.2 (58.0, 88.0)	74.0 (58.6, 86.7)	1.437†	0.697
Gamma-glutamyl transpeptidase (U/L), median (IQR)	24.0 (16.0, 33.0)	21.0 (15.0, 33.0)	20.0 (14.0, 30.0)	23.0 (15.0, 30.0)	5.628†	0.131
Total bilirubin (μmol/L), median (IQR)	12.1 (9.0, 15.8)	12.2 (8.9, 16.7)	12.4 (9.3, 15.8)	12.4 (9.1, 15.5)	0.438†	0.932
Direct bilirubin (μmol/L), median (IQR)	4.0 (3.0, 5.3)	3.8 (2.9, 5.4)	3.8 (2.8, 5.1)	3.7 (2.7, 5.2)	1.540†	0.673
Total protein (g/L), mean ± SD	64.29 ± 5.54	66.21 ± 5.82	65.34 ± 6.43	65.16 ± 7.03	1.819‡	0.142
Albumin (g/L), median (IQR)	39.2 (36.6, 42.8)	40.7 (37.8, 43.7)	40.3 (37.7, 43.5)	40.5 (38.2, 43.0)	8.177†	0.042
Globulin (g/L), median (IQR)	24.8 (21.9, 27.1)	25.2 (21.8, 28.7)	24.7 (22.1, 27.8)	23.9 (21.5, 27.4)	4.571†	0.206
Albumin to globulin ratio, median (IQR)	1.6 (1.4, 1.8)	1.6 (1.4, 1.9)	1.6 (1.4, 1.8)	1.7 (1.5, 1.9)	7.233†	0.065
Blood urea nitrogen (mmol/L), median (IQR)	5.3 (4.3, 6.3)	5.7 (4.5, 6.8)	5.3 (4.4, 6.6)	5.4 (4.5, 6.5)	3.034†	0.386
Creatinine (μmol/L), median (IQR)	67.0 (57.8,78.3)	67.0 (55.8, 79.0)	63.5 (55.0, 77.3)	64.9 (55.0, 76.2)	2.430†	0.488
Uric acid (μmol/L), mean ± SD	322.50 ± 103.55	306.76 ± 82.58	303.40 ± 96.50	304.42 ± 100.06	1.167‡	0.321
Total carbon dioxide (mmol/L), median (IQR)	25.7 (23.9, 27.0)	25.4 (23.9, 27.7)	25.6 (23.6, 27.2)	25.4 (23.9, 27.6)	0.338†	0.953
Fibrinogen (g/L), median (IQR)	3.2 (2.6, 4.0)	3.1 (2.5, 3.9)	2.9 (2.5, 3.5)	2.8 (2.4, 3.4)	17.184†	0.001
Activated partial thromboplastin time (s), median (IQR)	32.9 (27.9, 37.1)	33.5 (29.0, 37.5)	33.3 (28.3, 37.2)	31.4 (27.3, 35.3)	11.365†	0.010
International normalized ratio, median (IQR)	0.99 (0.9, 1.1)	0.99 (0.9, 1.1)	0.98 (0.9, 1.1)	0.97 (0.9, 1.0)	5.297†	0.151
Prothrombin time (s), median (IQR)	12.3 (11.2, 13.1)	12.2 (11.3, 13.1)	12.3 (11.2, 13.2)	11.9 (11.0, 12.9)	4.659†	0.199
Thrombin time (s), median (IQR)	16.9 (16.0, 17.8)	17.1 (16.4, 17.9)	17.2 (16.3, 18.0)	17.3 (16.4, 18.0)	4.028†	0.258
Size					12.473*	0.052
≤1 cm, n (%)	7 (6.03)	17 (12.32)	24 (6.42)	17 (7.87)
>1 cm and <3 cm, n (%)	60 (51.72)	64 (46.38)	203 (54.28)	130 (60.19)
≥3 cm, n (%)	49 (42.24)	55 (39.86)	140 (37.43)	64 (29.63)
Unknown, n (%)	0 (0)	2 (1.45)	7 (1.87)	5 (2.31)
Pathological type					18.489*	0.005
Adenocarcinoma, n (%)	74 (63.79)	99 (71.74)	294 (78.61)	179 (82.87)
Squamous cell carcinoma, n (%)	38 (32.76)	34 (24.64)	69 (18.45)	32 (14.81)
Others, n (%)	4 (3.45)	3 (2.17)	9 (2.41)	5 (2.31)
Unknown, n (%)	0 (0)	2 (1.45)	2 (0.53)	0 (0)
Clinical T category					43.545*	<0.001
T1, n (%)	60 (51.72)	81 (58.70)	218 (58.29)	139 (64.35)
T2, n (%)	34 (29.31)	29 (21.01)	120 (32.09)	62 (28.70)
T3, n (%)	14 (12.07)	23 (16.67)	17 (4.55)	7 (3.24)
T4, n (%)	8 (6.90)	3 (2.17)	11 (2.94)	3 (1.39)
Unknown, n (%)	0 (0)	2 (1.45)	8 (2.14)	5 (2.31)
Clinical N category					–§	0.047
N0, n (%)	81 (69.83)	100 (72.46)	267 (71.39)	178 (82.41)
N1, n (%)	14 (12.07)	14 (10.14)	41 (10.96)	15 (6.94)
N2, n (%)	20 (17.24)	21 (15.22)	57 (15.24)	18 (8.33)
N3, n (%)	1 (0.86)	1 (0.72)	1 (0.27)	0 (0)
Unknown, n (%)	0 (0)	2 (1.45)	8 (2.14)	5 (2.31)
Stage					–§	0.002
IA, n (%)	49 (42.24)	66 (47.83)	177 (47.33)	124 (57.41)
IB, n (%)	12 (10.34)	7 (5.07)	51 (13.64)	28 (12.96)
ⅡA, n (%)	9 (7.76)	12 (8.70)	24 (6.42)	10 (4.63)
ⅡB, n (%)	17 (14.66)	20 (14.49)	41 (10.96)	18 (8.33)
ⅢA, n (%)	16 (13.79)	17 (12.32)	54 (14.44)	16 (7.41)
ⅢB, n (%)	8 (6.90)	7 (5.07)	9 (2.41)	4 (1.85)
ⅢC, n (%)	1 (0.86)	1 (0.72)	0 (0)	0 (0)
Unknown, n (%)	4 (3.45)	8 (5.80)	18 (4.81)	16 (7.41)
EGFR					1.695*	0.638
Wild-type, n (%)	8 (6.90)	17 (12.32)	60 (16.04)	15 (6.94)
Mutant, n (%)	10 (8.62)	14 (10.14)	40 (10.70)	13 (6.02)
Unknown, n (%)	98 (84.48)	107 (77.54)	274 (73.26)	188 (87.04)
ALK					–§	0.482
Wild-type, n (%)	6 (5.17)	8 (5.80)	31 (8.29)	12 (5.55)
Mutant, n (%)	0 (0)	1 (0.72)	1 (0.27)	0 (0)
Unknown, n (%)	110 (94.83)	129 (93.48)	342 (91.44)	204 (94.44)

^*χ² value; ^†H value; ^‡F value; ^§Fisher–Freeman–Halton exact test. Unknown categories were excluded from statistical comparisons. ALK: Anaplastic lymphoma kinase; EGFR: Epidermal growth factor receptor; IQR: Interquartile range; MCH: Mean corpuscular hemoglobin; MCHC: Mean corpuscular hemoglobin concentration; MCV: Mean corpuscular volume; NLR: Neutrophil-to-lymphocyte ratio; PD-L1: Programmed death-ligand 1; SD: Standard deviation; –: Not available.

IHC examination

All unstained formalin-fixed paraffin-embedded (FFPE) tissue sections were stored at 4°C in sealed, light-protected containers until immunohistochemical staining.^[19] Before analysis, the structural integrity, surface discoloration, and transparency changes were assessed. Microscopic examination was performed to evaluate cellular morphology, preservation of tissue architecture, and uniformity of cell density. Samples that failed to meet the IHC quality criteria, including those with compromised tissue architecture, inconsistent cellular distribution, fragmentation, or inadequate adhesion, were excluded. Sections (5-µm thick) from FFPE blocks were subjected to IHC. After antigen retrieval, sections were incubated overnight at 4°C with anti-human PD-L1 rabbit monoclonal antibody (13684S, dilution 1:200; Cell Signaling Technology, Danvers, MA, USA) and anti-human CD8 mouse monoclonal antibody (66868-1-Ig, dilution 1:4000; Proteintech, Wuhan, China). PD-L1⁺ and CD8-positive (CD8⁺) cells were quantified using a digital microscope. A cutoff of 5% was used to define positivity [Figure 2]. Patients were classified into four groups according to PD-L1/CD8 status: PD-L1⁺ CD8⁻, PD-L1⁻ CD8⁻, PD-L1⁺ CD8⁺, and PD-L1⁻ CD8⁺.

Figure 2.

PD-L1/CD8 status in tumor tissue. Immunohistochemical staining was performed using anti-human PD-L1 rabbit monoclonal antibody and anti-human CD8 mouse monoclonal antibody to detect membranous PD-L1 expression and CD8⁺ T-cell infiltration. Representative staining patterns showed marked heterogeneity and were classified into four groups: PD-L1⁺ CD8⁺, PD-L1⁺ CD8⁻, PD-L1⁻ CD8⁻, and PD-L1⁻ CD8⁺. Scale bars, 50 μm. PD-L1: Programmed death-ligand 1.

Machine learning classifiers

Patients from Wuhan Union Hospital (n = 546) were used for model development and randomly divided into a training set (n = 382), a test set (n = 82), and an internal validation set (n = 82) at a ratio of 7:1.5:1.5 using a computer-generated randomization procedure. After splitting, the distributions of age and sex were checked across the three subsets to ensure approximate balance. The training set was used for model fitting and hyperparameter tuning, the test set was used for preliminary performance evaluation, and the internal validation set was used for further assessment of model stability. Patients from Renmin Hospital of Wuhan University (n = 298) served as an independent external validation set and were not involved in model training, hyperparameter tuning, or model selection. This allocation was based on commonly used machine learning practices requiring a sufficiently large training set for complex models and adequate sample size in the evaluation subsets. After preprocessing, three categorical and 39 continuous predictors were retained. Five machine learning algorithms were used to predict PD-L1⁺ CD8⁻ status: logistic regression, decision tree, random forest, extreme gradient boosting (XGB), and light gradient boosting machine (LightGBM). A stacking ensemble model was then constructed using these base learners. Logistic regression is a generalized linear model that maps linear predictions to probabilities via a sigmoid function and is suitable for binary classification tasks such as disease diagnosis or mortality prediction.^[20] XGB is based on a gradient boosting framework that optimizes the loss function using second-order Taylor expansion and controls model complexity through L1/L2 regularization.^[21] It can automatically handle missing values and incorporate cross-validation to reduce overfitting.^[21] Decision trees recursively partition the feature space using criteria such as the Gini index or information gain and are particularly useful when interpretable decision paths are desired.^[22] Random forest reduces variance by aggregating predictions from multiple bootstrapped decision trees and performs well in multimodal data integration.^[23] LightGBM uses a histogram-based algorithm and a leaf-wise growth strategy to improve computational efficiency, making it well suited to large-scale medical datasets.^[24] These algorithms are widely used in medical machine learning and were therefore selected for the present study. XGB, LightGBM, and decision trees were configured to account for class imbalance by assigning greater weight to the minority class. Missing values were imputed using model-based prediction. Hyperparameters were optimized using grid search, and the model with the highest area under the receiver operating characteristic curve (AUC) during tuning was selected as the optimal model. Receiver operating characteristic (ROC) curves were then used to evaluate model discrimination. Model performance was evaluated using AUC, sensitivity, specificity, accuracy, precision, recall, and F1 score. A stacking model was constructed using the above base learners. By integrating the strengths of multiple algorithms, stacking reduces the overfitting risk of any single model^[25] and leverages complementary errors across models to improve predictive stability and generalizability, particularly in complex datasets with limited sample sizes.^[26]

Feature importance analysis was performed to evaluate the contribution of individual clinical variables to the prediction of PD-L1⁺ CD8⁻ status in each machine learning model. For tree-based models, including decision tree, random forest, lightGBM, and XGB, feature importance was calculated based on the contribution of each variable to model splitting and prediction improvement. For logistic regression, feature importance was assessed based on the absolute magnitude of the standardized regression coefficients. For the stacking model, feature importance was derived from the relative contribution of input features to the final ensemble prediction. The importance scores were normalized and ranked within each algorithm.

SHapley Additive exPlanations (SHAP), a game-theoretic approach for interpreting machine learning models, was used to explain model outputs and quantify the contribution of each feature.^[27] SHAP provides both global feature importance and case-level interpretability by showing how each feature contributes to an individual prediction. Data preprocessing, model development, and statistical analyses were performed in Python (version 3.8; Python Software Foundation, Wilmington, DE, USA).

Statistical methods

The normality of continuous variables was assessed using the Kolmogorov–Smirnov test. Normally distributed continuous variables are presented as means with standard deviations (SDs), whereas non-normally distributed continuous variables are presented as medians with interquartile ranges (IQRs). Categorical variables are presented as frequencies and percentages. Comparisons among the four PD-L1/CD8 subgroups were performed using one-way analysis of variance (ANOVA) for normally distributed continuous variables with homogeneous variances, Welch’s ANOVA for normally distributed continuous variables with unequal variances, the Kruskal–Wallis H test for non-normally distributed continuous variables. For categorical variables, the Pearson chi-squared (χ²) test was used when its assumptions were satisfied; otherwise, the Fisher–Freeman–Halton exact test with Monte Carlo estimation was used. For variables showing a significant overall difference, subsequent post-hoc pairwise comparisons were performed according to the corresponding overall test: Bonferroni-adjusted pairwise comparisons after one-way ANOVA, the Games–Howell test after Welch’s ANOVA, and Mann–Whitney U test with Bonferroni correction after the Kruskal–Wallis H test. For post-hoc analyses requiring Bonferroni correction, adjusted P values were calculated by multiplying the raw P values by six, corresponding to the six pairwise comparisons among the four PD-L1/CD8 subgroups, with values capped at 1.00. Comparisons between the PD-L1⁺ CD8⁻ group and the non-PD-L1⁺ CD8⁻ group were performed using Student’s independent-samples t test for normally distributed continuous variables with homogeneous variances, Welch’s t test for normally distributed continuous variables with unequal variances, and the Mann–Whitney U test for non-normally distributed continuous variables. Categorical variables were compared using the Pearson χ² test or Fisher’s exact test, as appropriate. All tests were two-sided. For overall comparisons among the four subgroups, statistical significance was set at P <0.05. For post-hoc pairwise comparisons, Bonferroni-adjusted P <0.05 was considered significant.

Survival curves were estimated using the Kaplan–Meier method and compared with the log-rank test. Factors associated with OS were first assessed using univariable Cox proportional hazards regression analysis. Variables with P <0.05 in the univariable analysis were included in the multivariable Cox proportional hazards regression model. Sex, age, and treatment variables were additionally included because of their clinical relevance and potential influence on prognosis. Hazard ratios (HRs) and 95% confidence intervals (CIs) were reported. Data analyses were performed in Python (version 3.8; Python Software Foundation, Wilmington, DE, USA). P <0.05 was considered statistically significant.

Results

PD-L1/CD8 status and patient survival

A total of 844 patients with NSCLC who underwent surgical resection were included in this study. Patients were categorized into four groups according to PD-L1 and CD8 status. Baseline clinicopathological characteristics are shown in Table 1. The largest subgroup was PD-L1⁻ CD8⁻ (44.31% [374/844]), followed by PD-L1⁻ CD8⁺ (25.60% [216/844]), PD-L1⁺ CD8⁻ (16.35% [138/844]) and PD-L1⁺ CD8⁺ (13.74% [116/844]). No significant differences were observed among the four groups with respect to sex, age, smoking status, tumor size, epidermal growth factor receptor (EGFR) status, or anaplastic lymphoma kinase (ALK) status. Descriptively, adenocarcinoma was most frequent in the PD-L1⁻ CD8⁺ group (82.87% [179/216]), whereas squamous cell carcinoma was most common in the PD-L1⁺ CD8⁺ group (32.76% [38/116]; χ² = 18.489, P = 0.005). The distributions of clinical T category, clinical N category, and tumor stage differed significantly among the four PD-L1/CD8 subgroups (all P <0.05). Post-hoc pairwise comparisons with Bonferroni correction showed that the PD-L1⁺ CD8^– subgroup had a significantly higher proportion of T3 disease than the PD-L1^– CD8^– and PD-L1^– CD8⁺ subgroups (both adjusted P <0.001), but a significantly lower proportion of stage IA disease and higher proportions of stage IIB and III disease compared with the PD-L1^– CD8⁺ subgroup (adjusted P <0.05). Among the four PD-L1/CD8 subgroups, significant overall differences were observed in white blood cell count, lymphocyte percentage, neutrophil count, neutrophil-to-lymphocyte ratio (NLR), monocyte count, albumin, eosinophil count, fibrinogen, and activated partial thromboplastin time (APTT) (all P <0.05) [Table 1].

Follow-up data were available for 298 patients, among whom 86 all-cause deaths were recorded. The median follow-up time was 56.0 months (95% CI: 46.3–76.4 months). Kaplan–Meier analysis showed significant differences in OS among the four PD-L1/CD8 subgroups (P = 0.001) [Figure 3A]. Because pathological type, tumor stage, clinical T category, and clinical N category could strongly affect OS, we further evaluated the association between PD-L1/CD8 status and OS in stratified analyses. Although the PD-L1⁺ CD8⁻ subgroup showed the shortest OS among stage I patients, the difference did not reach statistical significance among the four groups (P = 0.063) [Figure 3B]. In contrast, a significant difference in OS was observed among the four PD-L1/CD8 subgroups in patients with stage II–IV disease (P = 0.014) [Figure 3C]. The PD-L1⁺ CD8⁻ group also had the worst prognosis among patients with adenocarcinoma (P = 0.001) [Figure 3D]. In squamous cell carcinoma, survival in the PD-L1⁺ CD8⁻ group continued to decline, although the difference was not statistically significant (P = 0.085) [Figure 3E]. A significant difference was also observed in patients with other pathological types, although the sample size was small (P = 0.044) [Figure 3F]. At the T1 stage, OS differed significantly among the four PD-L1/CD8 subgroups (P = 0.001) [Figure 3G]. At T2–T4 stages, no significant differences were observed among the four groups [Figure 3H–I]. In analyses stratified by clinical N category, significant differences in OS were found among the four PD-L1/CD8 subgroups in patients with N0 disease and those with N1–3 disease (P = 0.032 and P = 0.033, respectively) [Figure 3J–K]. Overall, both the main and stratified analyses indicated that PD-L1⁺ CD8⁻ status was associated with the shortest OS among the four groups. When each PD-L1/CD8 subgroup was compared with the remaining patients, no significant OS difference was observed for PD-L1⁺ CD8⁺ versus non-PD-L1⁺ CD8⁺ [Figure 3L]. PD-L1⁺ CD8⁻ status showed significantly worse OS compared with non-PD-L1⁺ CD8⁻ status (P <0.001) [Figure 3M]. No significant OS difference was observed for PD-L1⁻ CD8⁻vs. non-PD-L1⁻ CD8⁻ [Figure 3N], whereas a significant OS difference was observed for PD-L1⁻ CD8⁺vs. non-PD-L1⁻ CD8⁺ (P = 0.018) [Figure 3O].

Figure 3.

Kaplan–Meier analysis of OS according to PD-L1/CD8 status. (A) OS according to PD-L1/CD8 status in the survival analysis cohort. (B–C) OS in subgroups stratified by tumor stage. (D–F) OS in subgroups stratified by pathological type. (G–I) OS across clinical T categories. (J–K) OS according to clinical N category. (L–O) Comparisons between each PD-L1/CD8 subgroup and the remaining patients: PD-L1⁺ CD8⁺vs. non-PD-L1⁺ CD8⁺; PD-L1⁺ CD8⁻vs. non-PD-L1⁺ CD8⁻; PD-L1⁻ CD8⁻vs. non-PD-L1⁻ CD8⁻; PD-L1⁻ CD8⁺vs. non-PD-L1⁻ CD8⁺. Statistical significance was assessed using the two-sided log-rank test, with P <0.05 considered statistically significant. OS: Overall survival; PD-L1: Programmed death-ligand 1.

Factors associated with patient survival

To further examine the association between PD-L1/CD8 status and OS, we performed univariable and multivariable Cox regression analyses. ALK was not included in the Cox regression analysis because all patients with definitive follow-up outcomes had wild-type ALK. Univariable Cox regression identified several factors significantly associated with OS [Table 2], which included albumin (χ² = 4.597, HR = 0.953, 95% CI: 0.912–0.996, P = 0.032), fibrinogen (χ² = 4.661, HR = 1.084, 95% CI: 1.007–1.166, P = 0.031), and APTT (χ² = 7.124, HR = 1.056, 95% CI: 1.015–1.100, P = 0.008). Compared with PD-L1⁺ CD8⁺ status, PD-L1⁺ CD8⁻ status was associated with poorer OS (χ² = 4.020, HR = 1.970, 95% CI: 1.015–3.825, P = 0.045). Among pathological types, the “others” category was associated with poorer OS than adenocarcinoma (χ² = 9.772, HR = 3.831, 95% CI: 1.650–8.892, P = 0.002), whereas no significant difference was observed between squamous cell carcinoma and adenocarcinoma. With respect to clinical T category, T2 was associated with poorer OS (χ² = 5.655, HR = 1.748, 95% CI: 1.103–2.768, P = 0.017) compared to T1. For clinical N category, both N1 and N2–3 were associated with poorer OS (χ² = 8.620, HR = 2.604, 95% CI: 1.375–4.933, P = 0.003 and χ² = 31.304, HR = 3.938, 95% CI: 2.436–6.365, P <0.001) compared to N0. Regarding tumor stage, stage ⅡB and ⅢA were associated with poorer OS (χ² = 6.970, HR = 2.486, 95% CI: 1.264–4.888, P = 0.008 and χ² = 28.037, HR = 4.552, 95% CI: 2.597–7.977, P <0.001) compared to IA. Chemotherapy was significantly associated with worse OS (χ² = 10.433, HR = 2.101, 95% CI: 1.339–3.298, P = 0.001).

Table 2.

Univariable Cox proportional hazards regression analysis of overall survival in patients with non-small cell lung cancer

Characteristics	Wald χ2	HR	95% CI	P-value
Sex
Male		Ref
Female	0.413	0.866	0.559–1.342	0.520
Age (years)	0.077	0.997	0.974–1.020	0.781
Smoking
No		Ref
Yes	0.912	1.229	0.805–1.876	0.339
White blood cell count (×109/L)	1.459	1.044	0.974–1.119	0.227
Red blood cell count (×1012/L)	0.794	0.837	0.566–1.238	0.373
Hemoglobin (g/L)	0.006	0.999	0.987–1.013	0.938
Hematocrit (%)	0.736	0.979	0.932–1.028	0.391
MCV (fL)	0.001	1.000	0.983–1.018	0.972
MCH (pg)	1.309	1.042	0.971–1.118	0.253
MCHC (g/L)	0.387	1.002	0.996–1.007	0.534
Platelet count (×109/L)	0.557	1.001	0.998–1.004	0.456
Neutrophil percentage (%)	0.542	1.007	0.988–1.026	0.462
Lymphocyte percentage (%)	1.724	0.986	0.966–1.007	0.189
Monocyte percentage (%)	0.904	1.014	0.986–1.042	0.342
Eosinophil percentage (%)	0.061	0.984	0.864–1.120	0.805
Basophilic granulocyte percentage (%)	0.591	0.736	0.337–1.607	0.442
Neutrophil count (×109/L)	1.390	1.046	0.971–1.127	0.236
Lymphocyte count (×109/L)	0.707	0.850	0.582–1.242	0.400
NLR	2.876	1.028	0.996–1.061	0.090
Monocyte count (×109/L)	2.295	1.241	0.939–1.640	0.130
Eosinophil count (×109/L)	0.034	1.194	0.181–7.873	0.854
Basophilic granulocyte count (×109/L)	0.444	0.013	0.000–4312.827	0.505
Alanine aminotransferase (U/L)	1.452	0.989	0.972–1.007	0.228
Aspartate aminotransferase (U/L)	0.005	0.999	0.982–1.017	0.943
Alkaline phosphatase (U/L)	1.201	1.003	0.998–1.008	0.273
Gamma-glutamyl transpeptidase (U/L)	0.101	0.999	0.992–1.006	0.750
Total bilirubin (μmol/L)	1.706	0.971	0.929–1.015	0.191
Direct bilirubin (μmol/L)	0.925	1.046	0.954–1.146	0.336
Total protein (g/L)	1.520	0.976	0.938–1.015	0.218
Albumin (g/L)	4.597	0.953	0.912–0.996	0.032
Globulin (g/L)	0.251	1.013	0.962–1.068	0.617
Albumin to globulin ratio	1.252	0.710	0.390–1.293	0.263
Blood urea nitrogen (mmol/L)	0.035	0.995	0.942–1.050	0.852
Creatinine (μmol/L)	0.016	1.000	0.994–1.007	0.901
Uric acid (μmol/L)	<0.001	1.000	0.998–1.002	0.988
Total carbon dioxide (mmol/L)	0.590	1.033	0.952–1.120	0.442
Fibrinogen (g/L)	4.661	1.084	1.007–1.166	0.031
Activated partial thromboplastin time (s)	7.124	1.056	1.015–1.100	0.008
International normalized ratio	0.079	0.744	0.094–5.862	0.779
Prothrombin time (s)	0.787	1.063	0.929–1.216	0.375
Thrombin time (s)	0.955	0.924	0.788–1.083	0.328
PD-L1/CD8 status
PD-L1+ CD8+		Ref
PD-L1+ CD8−	4.020	1.970	1.015–3.825	0.045
PD-L1− CD8−	0.305	0.844	0.462–1.542	0.581
PD-L1− CD8+	2.826	0.535	0.258–1.109	0.093
Size
≤1 cm		Ref
>1 cm and <3 cm	0.003	0.970	0.297–3.169	0.960
≥3 cm	1.004	1.816	0.565–5.832	0.316
Pathological type
Adenocarcinoma		Ref
Squamous cell carcinoma	0.260	1.139	0.691–1.876	0.610
Others	9.772	3.831	1.650–8.892	0.002
Clinical T category
T1		Ref
T2	5.655	1.748	1.103–2.768	0.017
T3–4	2.599	1.785	0.882–3.610	0.107
Clinical N category
N0		Ref
N1	8.620	2.604	1.375–4.933	0.003
N2–3	31.304	3.938	2.436–6.365	<0.001
Stage
IA		Ref
IB	0.015	1.048	0.487–2.258	0.904
IIA	1.346	1.698	0.694–4.156	0.246
IIB	6.970	2.486	1.264–4.888	0.008
IIIA	28.037	4.552	2.597–7.977	<0.001
IIIB–C	3.179	2.621	0.909–7.562	0.075
EGFR
Wild-type		Ref
Mutant	1.195	0.724	0.406–1.291	0.274
Use of chemotherapy	10.433	2.101	1.339–3.298	0.001
Use of radiotherapy	1.171	1.649	0.666–4.079	0.279
Use of immunotherapy	1.593	0.281	0.039–2.020	0.207
Use of targeted therapy	1.484	0.535	0.195–1.464	0.223

CI: Confidence interval; EGFR: Epidermal growth factor receptor; HR: Hazard ratio; MCH: Mean corpuscular hemoglobin; MCHC: Mean corpuscular hemoglobin concentration; MCV: Mean corpuscular volume; NLR: Neutrophil-to-lymphocyte ratio; PD-L1: Programmed death-ligand 1; Ref: Reference group; –: Not available.

Variables with statistical significance in the univariable analysis were entered into the multivariable analysis. In addition, sex, age, and treatment variables were included in the multivariable model because of their well-established clinical relevance and potential influence on OS. Multivariable Cox regression analysis [Table 3] identified albumin (χ² = 6.791, HR = 0.920, 95% CI: 0.864–0.980, P = 0.009) as an independent predictor of OS. Among PD-L1/CD8 subgroups, PD-L1⁺ CD8⁻ status remained significantly associated with worse OS than PD-L1⁺ CD8⁺ status (χ² = 6.452, HR = 3.261, 95% CI: 1.310–8.116, P = 0.011). Pathological type also had prognostic significance, with the “others” pathological type category showing poorer OS than adenocarcinoma (χ² = 8.358, HR = 6.537, 95% CI: 1.831–23.347, P = 0.004).

Table 3.

Multivariable Cox proportional hazards regression analysis of overall survival in patients with non-small cell lung cancer.

Characteristics	Wald χ2	Adjusted HR	95% CI	P values
Sex
Male		Ref
Female	0.062	1.072	0.619–1.858	0.803
Age (years)	0.814	1.014	0.984–1.044	0.367
Albumin (g/L)	6.791	0.920	0.864–0.980	0.009
Fibrinogen (g/L)	0.037	1.012	0.894–1.146	0.848
Activated partial thromboplastin time (s)	1.111	1.028	0.977–1.081	0.292
PD-L1/CD8 status
PD-L1+ CD8+		Ref
PD-L1+ CD8−	6.452	3.261	1.310–8.116	0.011
PD-L1− CD8−	0.017	0.946	0.407–2.198	0.897
PD-L1− CD8+	0.044	1.105	0.435–2.806	0.833
Pathological type
Adenocarcinoma		Ref
Squamous cell carcinoma	0.191	1.162	0.593–2.275	0.662
Others	8.358	6.537	1.831–23.347	0.004
Clinical T category
T1		Ref
T2	<0.001	1.007	0.446–2.277	0.986
T3–4	3.010	0.103	0.008–1.342	0.083
Clinical N category
N0		Ref
N1	0.192	0.620	0.073–5.260	0.661
N2–3	0.598	0.268	0.010–7.535	0.440
Stage
IA		Ref
IB	0.071	1.181	0.347–4.017	0.790
IIA	0.003	0.962	0.252–3.661	0.954
IIB	2.158	5.832	0.555–61.307	0.142
IIIA	3.497	25.350	0.856–750.266	0.061
IIIB	2.769	140.113	0.414–47363.418	0.096
Use of chemotherapy	0.098	1.097	0.615–1.958	0.755
Use of radiotherapy	1.088	1.744	0.614–4.957	0.297
Use of immunotherapy	2.506	0.190	0.024–1.485	0.113
Use of targeted therapy	0.885	0.550	0.159–1.909	0.347

CI: Confidence interval; HR: Hazard ratio; PD-L1: Programmed death-ligand 1; Ref: Reference group.

Baseline characteristics between PD-L1⁺ CD8⁻ and non-PD-L1⁺ CD8⁻ cases

After confirming that PD-L1⁺ CD8⁻ status was associated with worse OS, we divided patients into PD-L1⁺ CD8⁻ and non-PD-L1⁺ CD8⁻ groups and compared their baseline characteristics [Table 4]. Several clinicopathological variables differed significantly between the PD-L1⁺ CD8⁻ group (n = 138) and the non-PD-L1⁺ CD8⁻ group (n = 706). The distribution of tumor size differed significantly between the two groups (χ² = 6.905, P = 0.032); 55 patients in the PD-L1⁺ CD8⁻ group had tumors ≥3 cm compared with 253 patients in the non-PD-L1⁺ CD8⁻ group (39.86% [55/138] vs. 35.84% [253/706]). The distribution of clinical T category also differed significantly between the two groups (χ² = 24.269, P <0.001). In addition, clinical stage distribution differed significantly between the two groups (P = 0.031), and stage IIB accounted for a higher proportion in the PD-L1⁺ CD8⁻ group (14.49% [20/138]). Total protein was significantly higher in the PD-L1⁺ CD8⁻ group than in the non-PD-L1⁺ CD8⁻ group (median, 66.8 g/L [IQR, 62.3–70.2 g/L] vs. 64.8 g/L [IQR, 61.2–69.1 g/L]; Z = −2.084, P = 0.037). Fibrinogen was also significantly higher in the PD-L1⁺ CD8⁻ group (median, 3.1 g/L [IQR, 2.5–3.9 g/L] vs. 2.9 g/L [IQR, 2.5–3.6 g/L]; Z = −2.519, P = 0.012).

Table 4.

Comparison of clinicopathological characteristics between the PD-L1⁺ CD8⁻ group and the non-PD-L1⁺ CD8⁻ group in patients with non-small cell lung cancer

Characteristics	Group (N = 844)		Test statistics	P values
	Non-PD-L1+ CD8−(N = 706)	PD-L1+ CD8−(N = 138)
Sex			0.134*	0.714
Male, n (%)	416 (58.92)	79 (57.25)
Female, n (%)	290 (41.08)	59 (42.75)
Age (years), median (IQR)	61.0 (55.0, 66.0)	60.0 (54.0, 67.0)	–0.408†	0.683
Smoking			0.034*	0.853
Yes, n (%)	313 (44.33)	60 (43.48)
No, n (%)	393 (55.67)	78 (56.52)
White blood cell count (×109/L), median (IQR)	5.7 (4.7, 7.1)	6.1 (5.0, 7.5)	–1.420†	0.156
Red blood cell count (×1012/L), median (IQR)	4.3 (3.9, 4.6)	4.3 (3.9, 4.7)	–1.045†	0.296
Hemoglobin (g/L), median (IQR)	131.0 (121.1, 143.0)	131.0 (122.3, 144.3)	–0.107†	0.915
Hematocrit (%), mean ± SD	38.73 ± 4.44	38.87 ± 4.05	−0.332‡	0.740
MCV (fL), median (IQR)	91.4 (86.7, 96.1)	91.4 (86.9, 94.7)	–0.846†	0.398
MCH (pg), median (IQR),	30.9 (29.5, 32.2)	30.5 (29.1, 31.7)	–1.978†	0.048
MCHC (g/L), median (IQR)	336.0 (325.2, 346.9)	335.6 (327.4, 343.7)	–0.278†	0.781
Platelet count (×109/L), median (IQR)	204.0 (167.0, 247.8)	205.5 (165.0, 244.0)	–0.092†	0.926
Neutrophil percentage (%), median (IQR)	59.7 (52.9, 67.2)	61.3 (54.1, 66.4)	–0.521†	0.602
Lymphocyte percentage (%), mean ± SD	28.69 ± 9.71	28.05 ± 9.53	0.697‡	0.486
Monocyte percentage (%), median (IQR)	7.7 (6.5, 9.2)	7.6 (6.5, 9.1)	–0.112†	0.911
Eosinophil percentage (%), median (IQR)	2.0 (1.3, 3.2)	2.1 (1.1, 3.4)	–0.095†	0.924
Basophilic granulocyte percentage (%), median (IQR)	0.4 (0.2, 0.6)	0.4 (0.3, 0.6)	–1.232†	0.218
Neutrophil count (×109/L), median (IQR)	3.4 (2.6, 4.5)	3.6 (2.8, 4.5)	–1.472†	0.141
Lymphocyte count (×109/L), median (IQR)	1.6 (1.3, 2.0)	1.6 (1.3, 2.1)	–0.603†	0.546
NLR, median (IQR)	2.1 (1.5, 3.0)	2.2 (1.6, 2.8)	–0.707†	0.480
Monocyte count (×109/L), median (IQR)	0.4 (0.4, 0.6)	0.5 (0.3, 0.6)	–0.945†	0.345
Eosinophil count (×109/L), median (IQR)	0.1 (0.1, 0.2)	0.1 (0.1, 0.2)	–0.532†	0.595
Basophilic granulocyte count (×109/L), median (IQR)	0.02 (0.00, 0.03)	0.02 (0.00, 0.04)	–1.793†	0.073
Alanine aminotransferase (U/L), median (IQR)	19.0 (15.0, 26.0)	22.0 (16.0, 28.0)	–1.449†	0.147
Aspartate aminotransferase (U/L), median (IQR)	20.0 (16.0, 27.3)	20.0 (16.0, 27.0)	–0.074†	0.941
Alkaline phosphatase (U/L), median (IQR)	73.0 (58.8, 87.9)	72.0 (60.9, 88.0)	–0.110†	0.912
Gamma-glutamyl transpeptidase (U/L), median (IQR)	22.0 (14.0, 31.0)	21.0 (15.0, 33.0)	–0.295†	0.768
Total bilirubin (μmol/L), median (IQR)	12.4 (9.2, 15.6)	12.2 (8.9, 16.7)	–0.223†	0.823
Direct bilirubin (μmol/L), median (IQR)	3.8 (2.8, 5.2)	3.8 (2.9, 5.4)	–0.677†	0.498
Total protein (g/L), median (IQR)	64.8 (61.2, 69.1)	66.8 (62.3, 70.2)	–2.084†	0.037
Albumin (g/L), median (IQR)	40.2 (37.7, 43.2)	40.7 (37.8, 43.7)	–0.862†	0.389
Globulin (g/L), median (IQR)	24.5 (21.9, 27.5)	25.2 (21.8, 28.7)	–1.447†	0.148
Albumin to globulin ratio, median (IQR)	1.6 (1.4, 1.9)	1.6 (1.4, 1.9)	–0.475†	0.635
Blood urea nitrogen (mmol/L), median (IQR)	5.3 (4.4, 6.5)	5.7 (4.5, 6.8)	–1.147†	0.252
Creatinine (μmol/L), median (IQR)	64.8 (55.1, 77.0)	67.0 (55.8, 79.0)	–1.165†	0.244
Uric acid (μmol/L), median (IQR)	301.0 (237.0, 364.2)	301.2 (249.5, 369.1)	–0.355†	0.723
Total carbon dioxide (mmol/L), median (IQR)	25.5 (23.7, 27.1)	25.4 (23.9, 27.7)	–0.322†	0.748
Fibrinogen (g/L), median (IQR)	2.9 (2.5, 3.6)	3.1 (2.5, 3.9)	–2.519†	0.012
Activated partial thromboplastin time (s), median (IQR)	32.5 (28.0, 36.8)	33.5 (29.0, 37.5)	–1.495†	0.135
International normalized ratio, median (IQR)	1.0 (0.9, 1.0)	1.0 (0.9, 1.1)	–1.506†	0.132
Prothrombin time (s), median (IQR)	12.2 (11.1, 13.1)	12.1 (11.3, 13.1)	–0.325†	0.745
Thrombin time (s), median (IQR)	17.1 (16.3, 18.0)	17.1 (16.4, 17.9)	–0.179†	0.858
Size			6.905*	0.032
≤1 cm, n (%)	48 (6.80)	17 (12.32)
>1 cm and <3 cm, n (%)	393 (55.67)	64 (46.38)
≥3 cm, n (%)	253 (35.84)	55 (39.86)
Unknown, n (%)	12 (1.70)	2 (1.45)
Pathological type			1.941*	0.379
Adenocarcinoma, n (%)	547 (77.48)	99 (71.74)
Squamous cell carcinoma, n (%)	139 (19.69)	34 (24.64)
Others, n (%)	18 (2.55)	3 (2.17)
Unknown, n (%)	2 (0.28)	2 (1.45)
Clinical T category			24.269*	<0.001
T1, n (%)	417 (59.07)	81 (58.70)
T2, n (%)	216 (30.59)	29 (21.01)
T3, n (%)	38 (5.38)	23 (16.67)
T4, n (%)	22 (3.12)	3 (2.17)
Unknown, n (%)	13 (1.84)	2 (1.45)
Clinical N category			–§	0.636
N0, n (%)	526 (74.50)	100 (72.46)
N1, n (%)	70 (9.91)	14 (10.14)
N2, n (%)	95 (13.46)	21 (15.22)
N3, n (%)	2 (0.28)	1 (0.72)
Unknown, n (%)	13 (1.84)	2 (1.45)
Stage			–§	0.031
IA, n (%)	350 (49.58)	66 (47.83)
IB, n (%)	91 (12.89)	7 (5.07)
IIA, n (%)	43 (6.09)	12 (8.70)
IIB, n (%)	76 (10.76)	20 (14.49)
IIIA, n (%)	86 (12.18)	17 (12.32)
IIIB, n (%)	21 (2.97)	7 (5.07)
IIIC, n (%)	1 (0.14)	1 (0.72)
Unknown, n (%)	38 (5.38)	8 (5.80)
EGFR			0.042*	0.838
Wild-type, n (%)	83 (11.76)	17 (12.32)
Mutant, n (%)	63 (8.92)	14 (10.14)
Unknown, n (%)	560 (79.32)	107 (77.54)
ALK			–§	0.284
Wild-type, n (%)	49 (6.94)	8 (5.80)
Mutant, n (%)	1 (0.14)	1 (0.72)
Unknown, n (%)	656 (92.92)	129 (93.48)

^*χ² value; ^†U value; ^‡t value; ^§Fisher’s exact test. Unknown categories were excluded from statistical comparisons. ALK: Anaplastic lymphoma kinase; EGFR: Epidermal growth factor receptor; IQR: Interquartile range; MCH: Mean corpuscular hemoglobin; MCHC: Mean corpuscular hemoglobin concentration; MCV: Mean corpuscular volume; NLR: Neutrophil-to-lymphocyte ratio; PD-L1: Programmed death-ligand 1; SD: Standard deviation; –: Not available.

Training, testing, internal validation, and external validation of machine learning models for prediction

After establishing the clinical significance of PD-L1⁺ CD8⁻ status for OS, we developed machine learning models to predict this phenotype in NSCLC. Patients from Wuhan Union Hospital (n = 546) were used for model development and randomly divided into training, test, and internal validation sets at a ratio of 7:1.5:1.5 using a computer-generated randomization procedure, resulting in 382, 82, and 82 patients in the three subsets, respectively. Age and sex distributions were checked after splitting to ensure approximate balance. Patients from Renmin Hospital of Wuhan University (n = 298) served as an independent external validation set and were not involved in model training, hyperparameter tuning, or model selection. We then evaluated the discriminative performance of multiple models across the four datasets.

In the training set [Figure 4A], the decision tree model showed modest performance (AUC = 0.789, 95% CI: 0.723–0.844, specificity = 80%, sensitivity = 68%), whereas logistic regression showed relatively limited performance (AUC = 0.651, 95% CI: 0.584–0.717, specificity = 51%, sensitivity = 72%). The lightGBM and random forest model further improved predictive accuracy (AUC = 0.987, 95% CI: 0.975–0.996, specificity = 96%, sensitivity = 94% and AUC = 0.997, 95% CI: 0.993–0.999, specificity = 100%, sensitivity = 95%, respectively). XGB and the stacking model showed excellent discriminative ability in the training set (AUC = 0.999, 95% CI: 0.997–1.000, specificity = 100%, sensitivity = 98% and AUC = 0.998, 95% CI: 0.995–1.000, specificity = 99%, sensitivity = 98%, respectively).

Figure 4.

Discriminative performance of machine learning algorithms for PD-L1⁺ CD8⁻ status in the training (A), test (B), internal validation (C), and external validation (D) sets. Each panel shows the trade-off between sensitivity and specificity for the respective models. Colored ROC curves represent the performance of individual models, and the white dashed line indicates the performance expected from a random classifier. AUC values, 95% CIs, specificity, and sensitivity are provided to quantify the discriminative ability of each model. AUC: Area under the receiver operating characteristic curve; CI: Confidence interval; LightGBM: Light gradient boosting machine; PD-L1: Programmed death-ligand 1; ROC: Receiver operating characteristic; XGB: Extreme gradient boosting.

In the test set [Figure 4B], the decision tree model and logistic regression exhibited a modest decline in performance (AUC = 0.609, 95% CI: 0.452–0.750, specificity = 77%, sensitivity = 41% and AUC = 0.542, 95% CI: 0.381–0.698, specificity = 66%, sensitivity = 53%, respectively). In contrast, the performance of the lightGBM, random forest, and XGB models decreased more substantially (AUC = 0.649, 95% CI: 0.477–0.804, specificity = 75%, sensitivity = 53%; AUC = 0.729, 95% CI: 0.592–0.848, specificity = 68%, sensitivity = 71%; and AUC = 0.735, 95% CI: 0.594–0.858, specificity = 60%, sensitivity = 82%, respectively).The stacking model also showed reduced performance in the test set (AUC = 0.700, 95% CI: 0.549–0.834, specificity = 68%, sensitivity = 76%).

In the internal validation set [Figure 4C], model performance again varied across algorithms. The lightGBM model and logistic regression showed modest performance (AUC = 0.697, 95% CI: 0.469–0.885, specificity = 64%, sensitivity = 86% and AUC = 0.630, 95% CI: 0.369–0.836, specificity = 57%, sensitivity = 86%, respectively). By contrast, the decision tree, random forest, XGB, and stacking models all maintained robust performance (AUC = 0.859, 95% CI: 0.696–0.987, specificity = 69%, sensitivity = 86%; AUC = 0.836, 95% CI: 0.700–0.933, specificity = 81%, sensitivity = 86%; AUC = 0.842, 95% CI: 0.684–0.963, specificity = 79%, sensitivity = 86%; and AUC = 0.853, 95% CI: 0.684–0.971, specificity = 85%, sensitivity = 86%, respectively).

In the external validation set [Figure 4D], the decision tree, lightGBM and logistic regression showed modest performance (AUC = 0.546, 95% CI: 0.459–0.631, specificity = 88%, sensitivity = 22%; AUC = 0.630, 95% CI: 0.518–0.731, specificity = 84%, sensitivity = 44% and AUC = 0.453, 95% CI: 0.340–0.564, specificity = 84%, sensitivity = 22%, respectively). In contrast, the random forest, XGB, and stacking model maintained relatively robust performance (AUC = 0.721, 95% CI: 0.606–0.818, specificity = 79%, sensitivity = 64%; AUC = 0.707, 95% CI: 0.618–0.794, specificity = 71%, sensitivity = 69%; and AUC = 0.783, 95% CI: 0.667–0.886, specificity = 98%, sensitivity = 56% respectively). Across the training, test, and internal validation sets shown in Supplementary Figure 1, http://links.lww.com/CM9/C993, the stacking model showed favorable overall performance, with strong accuracy, precision, recall, and F1 scores.

Most informative parameters of machine learning models to predict PD-L1⁺ CD8⁻ status

Feature importance analysis showed that the contribution of individual variables varied across algorithms, although several hematologic, biochemical, and coagulation-related indicators were repeatedly identified as influential predictors of PD-L1⁺ CD8⁻ status [Supplementary Figure 2, http://links.lww.com/CM9/C993]. Across models, several variables showed relative consistency in feature importance, including platelet count, monocyte count, basophil-related indices, direct bilirubin, alkaline phosphatase (ALP), albumin, hemoglobin/red blood cell count, and coagulation-related parameters. These findings suggest that prediction of PD-L1⁺ CD8⁻ status may rely on a composite pattern involving hematologic, biochemical, inflammatory, and coagulation-related alterations rather than on a single dominant predictor.

SHAP force plots were used to illustrate sample-level explanations of model predictions. In these representative examples, individual feature values contributed differently to the predicted probability of PD-L1⁺ CD8⁻ status, with red features pushing the model output toward a higher predicted probability and blue features pushing the output toward a lower predicted probability [Figure 5A–E]. In the decision tree model, uric acid, eosinophil count, and prothrombin time contributed positively to the prediction, whereas blood urea nitrogen (BUN), neutrophil count, and aspartate aminotransferase (AST) contributed negatively [Figure 5A]. In the lightGBM model, basophilic granulocyte count, ALP, NLR, neutrophil count, AST, and BUN contributed positively to the predicted probability [Figure 5B]. In the logistic regression model, APTT, ALP, mean corpuscular hemoglobin concentration, creatinine, and age contributed positively, whereas globulin, hemoglobin, uric acid, lymphocyte percentage, gamma-glutamyl transpeptidase, and total carbon dioxide contributed negatively [Figure 5C]. In the random forest model, BUN, AST, age, basophilic granulocyte count, and albumin-to-globulin ratio contributed negatively to the predicted probability [Figure 5D]. In the XGB model, ALP, AST, mean corpuscular hemoglobin concentration, and NLR contributed positively, whereas gamma-glutamyl transpeptidase, mean corpuscular hemoglobin, and prothrombin time contributed negatively [Figure 5E]. SHAP summary plots were further used to evaluate the global contribution of variables across samples. Neutrophil count ranked among the most influential predictors in most models, except logistic regression [Figure 5F]. Together, the SHAP analyses suggested that PD-L1⁺ CD8⁻ prediction was driven by a combination of hematologic, biochemical, inflammatory, and coagulation-related indicators rather than by a single dominant variable.

Figure 5.

SHAP force plots and summary plots for five machine learning models. (A–E) SHAP force plots for the decision tree, lightGBM, logistic regression, random forest, and XGB models, respectively. Red features indicate variables that increased the predicted probability of PD-L1⁺ CD8⁻ status, whereas blue features indicate variables that decreased the predicted probability. (F) SHAP summary plots showing feature importance across the five models. A/G: Albumin-to-globulin ratio; ALP: Alkaline phosphatase; APTT: Activated partial thromboplastin time; AST: Aspartate aminotransferase; BUN: Blood urea nitrogen; CRE: Creatinine; GGT: Gamma-glutamyl transpeptidase; LightGBM: Light gradient boosting machine; MCH: Mean corpuscular hemoglobin; MCHC: Mean corpuscular hemoglobin concentration; MCV: Mean corpuscular volume; NLR: Neutrophil-to-lymphocyte ratio; PD-L1: Programmed death-ligand 1; PT: Prothrombin time; RBC: Red blood cell count; SHAP: Shapley additive explanations; TCO₂: Total carbon dioxide; TT: Thrombin time; UA: Uric acid; WBC: White blood cell count; XGB: Extreme gradient boosting.

Discussion

The tumor immune microenvironment is increasingly recognized as an important determinant of patient outcome in tumors.^[28,29] In NSCLC, PD-L1 expression and CD8⁺ T-cell infiltration represent two complementary aspects of the tumor immune microenvironment: immune evasion and host antitumor response.^[30,31] In the present study, we found that combined assessment of PD-L1 and CD8 status provided prognostic stratification in surgically treated NSCLC. In particular, PD-L1⁺ CD8⁻ status emerged as a high-risk immune phenotype associated with shorter OS. Based on this finding, we further developed and validated machine learning models using preoperative clinical characteristics to predict PD-L1⁺ CD8⁻ status in tumor tissue.

PD-L1 expression alone has substantial biological and clinical complexity. It is dynamically regulated, spatially heterogeneous, and influenced by both intrinsic oncogenic signaling and adaptive immune responses.^[32,33] Furthermore, PD-L1 is expressed not only on the tumor cell membrane but also in the cytoplasm, exosomes, and even the nucleus.^[34,35] Membranous PD-L1 (mPD-L1) is the canonical form that binds PD-1 on cytotoxic T lymphocytes and suppresses antitumor immunity. By contrast, CD8⁺ T-cell infiltration more directly reflects the presence of an active local cytotoxic immune response and has been associated with favorable survival in tumors.^[36,37] Importantly, these two markers are biologically linked rather than independent: CD8⁺ T cells secrete interferon-γ, which can induce PD-L1 upregulation in tumor cells.^[38] Therefore, evaluating PD-L1 expression together with CD8⁺ T-cell infiltration may provide a more complete and biologically meaningful assessment of the tumor immune microenvironment than either marker alone.

In our study, we evaluated PD-L1 expression and CD8⁺ T-cell infiltration together and found that patients with PD-L1⁺ CD8⁻ status had the worst prognosis. Similar findings have been reported in esophageal cancer, suggesting that these four PD-L1/CD8-defined subgroups may represent biologically distinct immune states with different prognostic implications.^[29] On the one hand, high PD-L1 expression indicates preserved or enhanced tumor immune-evasive capacity, which can suppress cytotoxic T-cell function through the PD-1/PD-L1 axis.^[39] On the other hand, low CD8⁺ T-cell infiltration suggests insufficient local antitumor immune surveillance, implying that the tumor microenvironment lacks an effective effector T-cell population capable of recognizing and eliminating residual malignant cells.^[40] This combination may characterize an immune-suppressed or immune-excluded state, in which tumor cells not only evade immune attack but also face limited immune pressure within the tumor bed. Mechanistically, the adverse outcome associated with the PD-L1⁺ CD8⁻ status may be explained by coordinated immune-evasive and immune-exclusion programs within the tumor microenvironment. One important mechanism is T-cell exclusion driven by transforming growth factor-β (TGF-β). TGF-β can promote extracellular matrix deposition and stromal remodeling, thereby creating physical and functional barriers that limit CD8⁺ T-cell trafficking into the tumor parenchyma and contribute to low intratumoral CD8⁺ T-cell infiltration.^[41,42] At the same time, TGF-β may reinforce immune suppression by sustaining T-cell dysfunction and cooperating with PD-L1-associated inhibitory signaling, thereby weakening local antitumor immunity.^[43,44] Cancer-associated fibroblasts (CAFs) may further strengthen this process. CAFs can remodel the extracellular matrix and stromal architecture, thereby restricting T-cell trafficking into the tumor bed and contributing to low intratumoral CD8⁺ T-cell infiltration.^[45,46] In addition, CAF-derived cytokine networks, including TGF-β and interleukin-6 (IL-6), together with metabolic reprogramming within the microenvironment, may suppress CD8⁺ T-cell activation and effector function even when T cells are present.^[46,47] Emerging evidence also suggests that CAFs, or CAF-dominated microenvironments, may participate in PD-L1-associated immunosuppression, thereby linking stromal remodeling with checkpoint-mediated immune escape.^[48,49] Taken together, these mechanisms suggest that PD-L1⁺ CD8⁻ status may represent an immune-excluded, immunosuppressive phenotype characterized by both preserved checkpoint-mediated immune evasion and insufficient effective cytotoxic T-cell surveillance. These findings are clinically relevant because prognosis remains heterogeneous even after curative-intent surgery, and conventional clinicopathological variables do not fully capture underlying tumor biology. Our results suggest that PD-L1/CD8-based immune phenotyping may complement standard pathological assessment and help identify patients at higher risk of poor postoperative outcomes. In particular, patients with PD-L1⁺ CD8⁻ status may warrant closer surveillance and more individualized postoperative management.

After defining the biological significance of PD-L1⁺ CD8⁻ status, we collected preoperative clinical characteristics and developed machine learning models to screen this phenotype. To our knowledge, few studies have evaluated whether readily available clinical characteristics can be used to infer PD-L1⁺ CD8⁻ status in surgically treated NSCLC. Because these variables are available before surgery and before tissue-based immune phenotyping is completed, they may support earlier risk assessment. Previous studies have used spatial immune profiling, computed tomography (CT)-based artificial intelligence, radiomics, histopathology images, and molecular signatures to evaluate or predict PD-L1 expression, CD8⁺ T-cell infiltration, or immunotherapy-related outcomes.^{[50–52,57,61,62]} Although such datasets are information-rich, they are often costly and difficult to implement in routine practice. By contrast, clinical indicators are inexpensive, standardized, easy to obtain, and directly reflect the physiological state of the patient.^[53–56] Machine learning offers several advantages over traditional statistical methods, including the ability to handle high-dimensional data, identify nonlinear relationships, and perform automated feature selection.^[57] In the present study, multiple machine learning models were used to predict PD-L1⁺ CD8⁻ status, and a stacking model showed the best overall performance, with AUCs of 0.998, 0.700, 0.853 and 0.783 in the training, test, internal validation, and external validation sets, respectively. By combining multiple algorithms, stacking models may improve predictive robustness by integrating complementary information from different base learners; however, they may still overfit when the sample size is limited or when model complexity is high.^[58,59] Notably, XGB and the stacking model achieved high discrimination in the training set, which likely reflects near-complete fitting of the training data rather than perfect generalizable performance. Therefore, we interpreted model performance primarily based on the test, internal validation, and external validation sets. Although previous studies have predicted PD-L1 expression from hematoxylin-eosin-stained histopathology slides using weakly supervised deep learning and multiple-instance learning approaches,^[60] such approaches remain invasive because they still require tissue acquisition. In contrast, our model is based on routinely available clinical markers and may therefore provide a more practical alternative. Radiomics^[51,61] and pathomics^[62] have also been used to predict PD-L1 expression and CD8⁺ T-cell infiltration, but these studies generally considered PD-L1 and CD8 separately rather than jointly. Most prior work has focused on the prognostic or therapeutic relevance of PD-L1 or CD8 alone, whereas the tumor microenvironment is inherently multifactorial and likely requires a more integrated analytic approach.

Feature-importance analyses showed that the ranking of predictors varied across algorithms, which is expected because different models quantify variable contributions in different ways. More importantly, several variables showed relative consistency across models, including platelet count, monocyte count, basophil-related indices, albumin, hemoglobin/red blood cell count, direct bilirubin, ALP, and coagulation-related parameters. This pattern suggests that PD-L1⁺ CD8⁻ status is unlikely to be captured by a single biomarker and may instead reflect a composite host–tumor state involving systemic inflammation, myeloid immune regulation, nutritional status, hepatic metabolic alterations, and coagulation abnormalities. Platelet-related markers have been linked to PD-L1 biology in lung cancer: a recent study reported that lymphocyte-to-neutrophil ratio and platelet-to-neutrophil ratio had predictive value for PD-L1 expression in lung cancer.^[63] Basophils are one of the main sources of IL-4 in the bone marrow.^[64] Through its effects on granulocyte–monocyte progenitors, IL-4 can drive monocytes toward an immunosuppressive macrophage phenotype that expresses PD-L1 and inhibits the antitumor activity of CD8⁺ T cells.^[64] Albumin and albumin–bilirubin-related markers have likewise been associated with outcomes in NSCLC, suggesting that host nutritional and inflammatory status may be relevant to immune phenotypes.^[65] In surgically treated NSCLC, platelet- and fibrinogen-based inflammatory/coagulation markers have also been linked to postoperative survival, which is consistent with the recurrent contribution of coagulation-related variables in our models.^[66,67]

Future studies may explore whether immune-restorative strategies aimed at enhancing CD8⁺ T-cell activation or trafficking could be relevant for this high-risk immune phenotype. Potential approaches include cytokine-based enhancement of T-cell proliferation and activation, such as IL-2,^[68,69] IL-7,^[70,71] and IL-15^[72]; immune adjuvants such as OX40 agonists^[73] and adoptive T-cell therapies such as chimeric antigen receptor T-cell therapy.^[74] In future clinical research settings, routinely available patient data could be entered into the stacking model to estimate the probability of PD-L1⁺ CD8⁻ status, thereby supporting risk assessment and postoperative management.

Several limitations should be acknowledged. First, the sample size was modest, and the retrospective design makes selection bias and residual confounding unavoidable. Second, although external validation was performed, both cohorts were derived from hospitals within the same region, and additional multicenter validation is required to establish the generalizability of the model. Third, the identified predictors should be interpreted as contributors to model performance rather than as mechanistic determinants of PD-L1⁺ CD8⁻ status, and the biological basis of these associations warrants further investigation.

Conclusions

In conclusion, PD-L1/CD8-based immune phenotyping identified prognostically distinct subgroups in surgically treated NSCLC, with PD-L1⁺ CD8⁻ status defining a high-risk phenotype associated with poor survival. A machine learning model based on routine preoperative clinical characteristics showed promising performance for identifying this phenotype and may provide a practical tool for preoperative risk stratification in resectable NSCLC.

Acknowledgments

We are grateful to Dr. Bo Huang (Wuhan Union Hospital) for useful discussions throughout this study.

Funding

This work was supported by the National Key Research and Development Program of China (Nos. 2022YFF1203300 and 2023YFC3043505), the National Natural Science Foundation of China (No. 82330003), the Natural Science Foundation of Hubei Province (Nos. 2022CFB238 and 2026AFB247), and the Hubei Provincial Key Research and Development Program (No. 2022BCA016).

Conflicts of interest

None.

Supplementary Material

cm9-139-2950-s001.docx ^{(227.6KB, docx)}