Predicting PD-L1 expression in advanced EGFR-mutant lung adenocarcinoma patients using NECT, CECT radiomics and clinical features

Abstract

Objectives

Recent studies have applied machine-learning radiomics to predict EGFR mutations or PD-L1 expression, yet little is known about PD-L1 prediction specifically in EGFR-mutant lung adenocarcinoma, despite evidence linking high PD-L1 expression to TKI resistance. To address this gap, this study integrates CT radiomics features with clinical data to develop and validate interpretable models for predicting PD-L1 expression in advanced EGFR-mutant LUAD and identifying key predictive determinants.

Materials and methods

In this single-center study, we retrospectively collected the pretreatment CT images, including non-contrast enhanced CT (NECT) and contrast enhanced CT (CECT), as well as clinical information from a cohort consisting of 165 EGFR-Mutant LUAD patients consecutively staged at IIIB-IVB between January 2020 and May 2022. After feature extraction and selection, six machine learning algorithms (Logistic regression, Linear discriminate analysis, Supporter vector machine, Gradient boosting, Random forest and adaptive boosting) were employed to construct the models. Model performances were assessed using 5-fold cross-validation and then visualized using coefficient matrices and SHAP (SHapley Additive exPlanations) to interpret the models’ predictions.

Results

A total of 30 radiomics features and 4 clinical features were selected. Models using radiomics features alone achieved the highest area under the curve (AUC) value of 0.91, and the combination of radiomics and clinical features models showed similar predictive performance with an AUC value of 0.90. However, the clinical features models exhibited insufficient predictive ability and poor generalizability (AUC = 0.64). According to the coefficient matrices, the top 5 features associated with PD-L1 ≥ 1% were: cect-wavelet-LHH-gldm-DependenceVariance, nect-wavelet-LHL-glcm-MCC, nect-wavelet-LHL-glrlm-LRLGLE, cect-wavelet-LLH-glcm-Imc2 and cect-wavelet-LLH-firstorder-Skewness, which was consistent with the results form SHAP.

Conclusion

Our study demonstrated that radiomics and clinical features via machine learning method offer a noninvasive approach to predict the PD-L1 expression status in EGFR-mutant LUAD patients. The model shows potential as a noninvasive tool to support immunotherapy decision-making and lays the groundwork for future clinical validation.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12885-025-15514-w.

Introduction

Lung cancer is the most prevalent cancer worldwide, with the highest mortality rate [1]. Approximately 85% of lung cancer are pathologically diagnosed as non-small cell lung cancer (NSCLC), of which lung adenocarcinoma (LUAD) is the most common subtype [2]. Epidermal growth factor receptor (EGFR) mutation is the most common driver gene of NSCLC, particularly in nonsmoking females, accounting for 40%–55% of LUAD in Asian patients [3]. Molecularly targeted therapy, such as epidermal growth factor receptor tyrosine kinase inhibitors (EGFR-TKIs), and immunotherapy, including anti-programmed cell death-1 (PD-1) and anti-programmed cell death ligand-1 (PD-L1), have revolutionized cancer treatment and improved survival for LUAD patients. EGFR-TKIs are the standard first-line treatment for locally advanced or metastatic EGFR-mutant LUAD, but their effectiveness varies from person to person. Previous studies have shown that positive or high PD-L1 expression is associated with de novo resistance to first-, second- and third-generation EGFR TKIs, resulting in shorter progression-free survival and overall survival in inoperable NSCLC patients with EGFR mutation [4–7]. Immunotherapy in combination with antiangiogenic drugs is a potential treatment option for these patients [8, 9].

Gene mutation status can be detected from peripheral blood or tissue samples. However, the expression levels of PD-L1 were only approved to be measured by immunohistochemistry (IHC), which requires surgical or biopsied tumor specimens through invasive procedures, carrying the risk of morbidities or tumor metastasis. Furthermore, the results of IHC are intratumorally heterogeneous and variational from different antibodies and platforms, as well as multiple scoring criteria [10]. Additionally, PD-L1 expression may change as the tumor progresses during EGFR-TKIs therapy [11], and it is impractical and difficult to obtain tumor biopsies multiple times. Therefore, there is an urgent need for an alternative noninvasive method to detect PD-L1 expression, especially when tumor tissues are unavailable or when IHC fails.

Radiomics is a field that analyzes high-throughput images, such as CT and positron emission tomography images, using bioinformatics tools to build clinical models for assessing gene, protein expression, and microenvironment heterogeneity of tumors. Radiomic analyses extract features from images, including tumor shape, size, voxel intensity, and texture, and then identify the features that are correlated with pathological or molecular typing, response to treatment, or clinical outcomes. Radiomics has been widely used in the diagnosis, prediction, and prognosis of NSCLC patients [12]. However, traditional machine learning often lacks model explainability and result interpretation, limiting the broader application of many radiomic studies [13]. Indeed, explainable and interpretable models have the potential to reveal the associations between specific radiomic features and biological behaviors or clinical outcomes, offering new scientific insights necessary for translating radiomics to clinical practice [14].

In this project, we developed an interpretable model to predict PD-L1 expression values ≥ 1% in advanced EGFR-mutant LUAD patients based on radiomic features extracted from pretreatment non-contrast enhanced CT (NECT), contrast enhanced CT (CECT), and clinical information, and aimed to noninvasively predict survival outcomes for these patients before they received EGFR-TKI treatments.

Methods

Patients

In this single-center observational study, a cohort of consecutive patients with lung cancer was retrospectively recruited from The Second Affiliated Hospital of Chongqing Medical University (Chongqing, China) between January 2020 and May 2022. The inclusion criteria were as follows: (1) aged 18 years and above, (2) staged IIIB to IVB according to the 8th edition of the AJCC/UICC TNM staging system, (3) histologically or cytologically confirmed primary LUAD, (4) next-generation sequencing proven EGFR mutations (including exon 19 deletions, the L858R point mutation and other rare mutations), and (5) pretreatment CECT and NECT images. Exclusion criteria were (1) clinical characteristics as gender, age, and clinical stage were missing; (2) previous operative treatment; (3) lack of sufficient left-over tissue for IHC; and (4) target lesion < 1 cm in size or difficult to delineate on CT images. The authors declare that this study was conducted in accordance with the Helsinki Declaration as revised in 2013.

Clinical data, including demographic characteristics (gender, age, smoking status, family malignant tumor history), tumor TMN stage and clinical stage, and laboratory results, including baseline tumor markers and inflammatory indexes, were extracted from the Electronic Medical Records System using a standardized data collection form. Tumor markers consisted of baseline carcinoembryonic antigen, cytokeratin fragment 19, and neuron-specific enolase. Inflammatory indexes included white blood cell count and absolute neutrophil and lymphocyte counts. Peripheral blood next-generation sequencing was performed for all patients after the diagnosis of LUAD. Subtypes of EGFR mutations and gene expression of other oncogenic driver mutations, such as MET amplification, TP53, PIK3CA, and ERBB2, were collected from the next-generation sequencing analysis. The study was reviewed and approved by the Ethics Committees of the Second Hospital of Chongqing Medical University. Informed consent was obtained from each patient prior to data collection.

PD-L1 expression by IHC

PD-L1 expression by tumor cells was assessed on paraffin sections or tissue using Agilent 22C3 antibody performed on the platform of Dako Link 48 in IHC assays. PD-L1 expression was interpreted by the tumor proportion score, which is defined as the percentage of neoplastic cells with membranous staining relative to all neoplastic cells and was classified as negative (< 1% or absence of expression) and positive (≥ 1%).

CT images acquisition and processing

All patients underwent chest NECT and CECT patients using Multi-detector CT scanners (Aquilion ONE, Toshiba; Somatom Definition Force, Siemens). The scanning protocol was as follows: a tube voltage of 120 kV, adaptive tube current, and a layer thickness of 1.25 mm. The contrast agent was iohexol solution, the total dose was 150–300 mg/kg iodine, and the injection flow rate was 2.5-3.0 ml/s. NECT is a CT image performed without the administration of contrast agents, primarily used to evaluate tissue density and structural features. While the CECT image is performed after intravenous administration of contrast media to enhance visualization of vascular structures and lesion characteristics.

Resampling was performed according to the methodology described in our previous study. In brief, CT image pixel intensities were first converted from radiodensity values to Hounsfield Units based on scan metadata. Thereafter, all tumor masks and corresponding CT datasets were uniformly resampled to achieve consistent spatial resolution, setting slice thickness and in-plane pixel spacing to 1 mm and [1.0, 1.0] mm, respectively. Slice dimensions were adjusted accordingly, and interpolation was employed to reconstruct the resampled volumes. Finally, intensity values were clipped to the range of − 1000 to 400 HU and normalized using a min–max scaling approach, linearly mapping all voxel values to the [0, 1] interval.

The region of interest (ROI) was delineated to encompass the entire primary pulmonary tumor lesions on both NECT and CECT images. The ROIs were manually and independently delineated on each slice using ITK-SNAP by two experienced radiologists with more than 5 years of clinical experience in diagnostic imaging, and subsequently verified by an expert radiologist with more than 10 years of experience. The PD-L1 status was blinded to all radiologists involved in the study. Reproducibility was evaluated by two radiologists using 30 randomly selected chest CT scans, with an intraclass correlation coefficient greater than 0.75 considered indicative of satisfactory interobserver agreement.

Features extraction and features selection for CT images and clinical information

Radiomics feature extraction (involving LoG and Wavelet) from the original images and ROI images was performed using the radiomics and SimpleITk Python packages. For feature selection, we employed recursive feature elimination (RFE) in the training set to select the top 30 radiomics features and the top 4 clinical features for training our machine learning models. In RFE, we used logistic regression as the external estimator to assign weights to each feature. Initially, the evaluator was trained on the full set of features, determining the significance of each. Subsequently, features with minimal importance were systematically removed. This iterative process continued until we achieved the target number of features.

Modeling of radiomics or clinical features to predict the PD-L1 expression status

To model radiomics or clinical features for predicting PD-L1 expression status, we used various machine learning algorithms, including logistic regression, Linear discriminate analysis (LDA), supporter vector machine (SVM), gradient boosting, random forest (RF) and AdaBoost. All modeling was performed in Python using the scikit-learn library. Continuous variables were z-score transformed, and categorical variables were one-hot encoded. We used the area under the curve (AUC) of the receiver operating characteristic curves when tuning the models, and evaluated their performance based on balanced accuracy and precision. We employed repeated stratified 5-fold cross-validation to evaluate model performance. Hyperparameters tuning was conducted using Optuna with 1000 iterations. Finally, coefficient matrices and Shapley Additive Explanations (SHAP) were used to interpret the model.

Results

Demographic characteristics

A total of 236 patients met the inclusion criteria, 71 were excluded, and ultimately, 165 patients were included in the study (see Fig. 1). Among them, 96 (58.3%) had PD-L1 expression values greater than 1%. Half of the patients (55.2%) were female, with a median age of 64.6 years, ranging from 24.0 to 88.0. Among the cohort, 59 (35.8%) were smokers (current or former), and the majority (88.5%) were without a family history. Twenty-one (12.7%) patients in the cohort were diagnosed with stage IIIB or IIIC, and the others (87.3%) were staged IV. Sixty-six (40.0%) had EGFR exon 19 deletions, 86 (52.1%) had the L858R point mutation, and 13 (7.9%) had other rare mutations. Nine (5.5%) patients were with MET amplification, and 27 (16.4%) patients were with TP53 mutation. Demographic and clinical characteristics are shown in Table 1.

Fig. 1.

Flowchart illustrating patient selection, inclusion, and exclusion process for this study. A total of 1,755 patients with advanced lung cancer were initially screened. Among them, 236 patients met the inclusion criteria, and 71 were subsequently excluded due to predefined exclusion factors. Finally, 165 patients were included in the study cohort

Table 1.

Patient baseline demographic and disease characteristics

Demographic characteristics	Total (n = 165)	PD-L1 ≥ 1% (n = 96)	PD-L1<1% (n = 69)
Gender
Male	74 (44.8%)	49 (51.0%)	25 (36.2%)
Female	91 (55.2%)	47 (49.0%)	44 (63.8%)
Age, median (range)	64.6, (24.0–88.0)	63.0, (24.0–84.0)	66.0, (37.0–88.0)
Smoking status
Never	106 (64.2%)	57 (59.4%)	49 (71.0%)
Current	44 (26.7%)	31 (32.3%)	13 (18.8%)
Former	15 (9.1%)	8 (8.3%)	7 (10.1%)
Family malignant tumor history
Without	146 (88.5%)	86 (89.6%)	60 (87.0%)
With	19 (11.5%)	10 (10.4%)	9 (13.0%)
Clinical Stages
III	21 (12.7%)	13 (13.5%)	8 (11.6%)
IV	144 (87.3%)	83 (86.5%)	61 (88.4%)
Tumor Marker
CEA (ng/mL)	115, (0.8–1000)	96.1, (0.8–1000.0)	142.9, (1.0-1000.0)
CYFRA21-1 (ng/mL)	10.9, (0.5-103.7)	11.1, (0.5-103.7)	10.6, (1.0-77.5)
NSE (ng/mL)	19.5, (4.0-79.5)	20.2, (4.0-79.5)	18.4, (7.8–72.3)
Inflammatory index
WBC (*109)	7.7, (1.2–64.7)	7.6, (2.9–34.4)	7.9, (1.2–64.7)
ANC (*109)	6.1, (0.2–66.9)	6.2, (1.5–66.9)	6.0, (0.2–62.5)
ALC (*109)	1.3, (0.3–3.6)	1.3, (0.3–3.6)	1.2, (0.3–3.1)
EGFR mutation
19 del	66 (40.0%)	34 (35.4%)	32 (46.4%)
L858R	86 (52.1%)	56 (58.3%)	30 (43.5%)
Others	13 (7.9%)	6 (6.2%)	7 (10.1%)
MET amplification	9 (5.5%)	7 (7.3%)	2 (2.9%)
TP53 mutation	27 (16.4%)	20 (20.8%)	7 (10.1%)
PIK3CA mutation	9 (5.5%)	6 (3.6%)	3 (4.3%)
ERBB2 mutation	3 (1.8%)	2 (1.2%)	1 (1.4%)

WBC White blood cell count, ANC Absolute neutrophil count, ALC Absolute lymphocyte count, CEA Carcinoembryonic antigen, CYFRA21-1 Cytokeratin-19 fragment, NSE Neuron-specific enolase

Feature extraction and selection

For radiomics features, the feature extraction step yielded 1702 features (851 for CECT and 851 for NECT) from the original and ROI images. After radiomics feature selection, we included 30 features for further analysis: (1) 15 features from CECT, including 4 original shape features (first-order statistics, N = 1 and texture, N = 3) and 11 wavelet-based features (first-order statistics, N = 2 and texture, N = 9); (2) 15 features from NECT, including 2 original shape features(shape, N = 1 and texture, N = 1) and 13 wavelet-based features (first-order statistics, N = 4 and texture, N = 9) (see Table 2). For clinical features, the RFE algorithm selected the best 4 features: smoking status, EGFR mutations, MET mutations, and TP53 mutations (see Fig. 2).

Table 2.

The CT image features incorporated into radiomics models

CECT	Original	Firstorder	Kurtosis
	original	gldm	Large Dependence Low Gray Level Emphasis
	original	glszm	Small Area Emphasis
	original	ngtdm	Strength
	wavelet-LLH	firstorder	Skewness
	wavelet-LLH	glcm	Imc2
	wavelet-LLH	glrlm	Short Run Low Gray Level Emphasis
	wavelet-LHH	gldm	Dependence Variance
	wavelet-LHH	glrlm	Long Run Low Gray Level Emphasis
	wavelet-HLL	gldm	Dependence Variance
	wavelet-HLL	glszm	Size Zone Non-Uniformity Normalized
	wavelet-HLL	glszm	Small Area Emphasis
	wavelet-HLL	glszm	Small Area Low Gray Level Emphasis
	wavelet-HLH	firstorder	Skewness
	wavelet-HHH	glcm	Inverse Variance
NECT	original	shape	Flatness
	original	glszm	Gray Level Non-Uniformity Normalized
	wavelet-LLH	firstorder	Kurtosis
	wavelet-LLH	firstorder	Skewness
	wavelet-LLH	glrlm	Short Run Low Gray Level Emphasis
	wavelet-LLH	ngtdm	Strength
	wavelet-LHL	firstorder	Minimum
	wavelet-LHL	glcm	MCC
	wavelet-LHL	gldm	Dependence Variance
	wavelet-LHL	glrlm	Long Run Low Gray Level Emphasis
	wavelet-LHH	glrlm	Long Run High Gray Level Emphasis
	wavelet-HLH	gldm	Small Dependence High Gray Level Emphasis
	wavelet-HLH	glrlm	Short Run High Gray Level Emphasis
	wavelet-HHL	firstorder	Median
	wavelet-HHH	glcm	MCC

Fig. 2.

Feature importance as determined by Recursive Feature Elimination (RFE) with logistic regression. The y-axis represents the features of the dataset, and the x-axis denotes their respective importance scores. Features are ranked from the most to the least important, illustrating their significance in predicting the target variable

Model performances

After feature selection, we initially employed radiomics features, clinical features, and the combination of clinical and image features separately to construct models. The baseline AUC of the six untuned models were performed using 5-fold cross-validation (Table 3); and their balanced accuracy and precision are shown in the Supplementary Figure S1, S2, S3.

Table 3.

The baseline AUC of the three untuned models

	Radiomics	Clinical	Combined
Logistic regression	0.91 (0.04)	0.61 (0.07)	0.90 (0.05)
Linear discriminate analysis	0.91 (0.04)	0.62 (0.08)	0.90 (0.05)
Supporter vector machine	0.82 (0.06)	0.61 (0.08)	0.89 (0.05)
Gradient boosting	0.68 (0.09)	0.64 (0.06)	0.68 (0.08)
Random forest	0.65 (0.09)	0.64 (0.06)	0.63 (0.09)
Adaboost	0.66 (0.11)	0.63 (0.08)	0.68 (0.10)

Data are shown in mean (SD)

When using radiomics features alone to construct models (see Fig. 3), both Logistic Regression and LDA models exhibited comparable optimal performance with an AUC value of 0.91 (logistic regression 95% CI: 0.90, 0.93; LDA 95% CI: 0.89, 0.93), followed by SVM model (AUC = 0.82, 95% CI: 0.80–0.85), Gradient boosting (AUC = 0.68, 95% CI: 0.66–0.70), RF (AUC = 0.65, 95% CI: 0.62–0.69), and gradient boosting (AUC = 0.66, 95% CI: 0.61–0.70). When constructing models solely based on clinical features (see Fig. 4), the Gradient boosting model (AUC = 0.64, 95%CI: 0.62, 0.66) and RF (AUC = 0.64, 95%CI: 0.61, 0.67) performed the best, followed by Adaboost (AUC = 0.63, 95% CI: 0.60, 0.66), LDA (AUC = 0.62, 95%CI: 0.59, 0.65), Logistic regression (AUC = 0.61, 95%CI: 0.58, 0.64) and SVM (AUC = 0.61, 95%CI: 0.58, 0.64). Obviously, the AUC values were all less than 0.7, suggesting that clinical features alone are inadequate for predicting PD-L1 status. When constructing models by combining radiomics features with clinical features (see Fig. 5), the Logistic Regression model and LDA both achieved an AUC value of 0.9 (logistic regression 95%CI: 0.88, 0.92; LDA 95% CI: 0.88, 0.92), followed by SVM (AUC = 0.89, 95%CI: 0.87, 0.91), Gradient boosting (AUC = 0.68, 0.65, 0.71) and Adaboost (AUC = 0.68, 0.64, 0.72), and RF (AUC = 0.63, 0.59, 0.67).

Fig. 3.

Receiver Operating Characteristic (ROC) curves for six different models built using radiomic features. The x-axis represents the false positive rate (1-Specificity) while the y-axis represents the true positive rate (Sensitivity). Curves are evaluated on the held-out test set (n = 33)

Fig. 4.

ROC curves for six different models built using clinical features. The x-axis represents the false positive rate (1-Specificity) while the y-axis represents the true positive rate (Sensitivity). Curves are evaluated on the held-out test set (n = 33)

Fig. 5.

ROC curves for six different models built using the combination of clinical and radiomic features. The x-axis represents the false positive rate (1-Specificity) while the y-axis represents the true positive rate (Sensitivity). Curves are evaluated on the held-out test set (n = 33)

The results indicate that combining radiomics features with clinical feature models showed a similar performance to radiomics models. Conversely, the clinical feature models yielded the poorest performance. Therefore, constructing a model solely based on radiomics features is the optimal choice. Next, we selected radiomics features to establish our models and employed Optuna to tune hyperparameters for SVM, RF, and adaboost models, and all performances of the models were improved (0.91 for SVM, 0.74 for adaboost and 0.67 for random forest). Although the AUC value of the SVM model improved after tuning the hyperparameters, it still did not outperform the Logistic Regression model (DeLong test p-value = 0.75). Considering that Logistic Regression is simpler and offers higher explainability, it is regarded as the optimal choice. The calibration curve of the logistic regression model is shown in Supplementary Figure S4.

Models’ explainability

We visualize the Logistic Regression model using coefficient matrices and SHAP techniques. According to coefficient matrices (see Fig. 6), the top 5 features positively collected with PD-L1 ≥ 1% were: cect-wavelet-LHH-gldm-DependenceVariance, nect-wavelet-LHL-glcm-MCC, nect-wavelet-LHL-glrlm-LongRunLowGrayLevelEmphasis, cect- wavelet-LLH-glcm-Imc2 and cect-wavelet-LLH-firstorder-Skewness. The top 5 features positively collected with PD-L1 < 1% were: nect-wavelet-LHL-gldm-DependenceVariance, nect-wavelet-HLH-gldm-SmallDependenceHighGrayLevelEmphasis, cect-original-ngtdm-Strength, nect-wavelet-LLH-glrlm-ShortRunLowGrayLevelEmphasis, and nect-original-glszm-Gray LevelNon-UniformityNormalized. According to SHAP (see Fig. 7), the top 4 features positively collected with PD-L1 ≥ 1% were: cect-wavelet-LHH-gldm-DependenceVariance, cect-wavelet-LLH-firstorder-Skewness, nect-wavelet-LHL-glrlm-LongRunLowGrayLevelEmphasis, and nect-wavelet-LHH-glrlm-LongRunHighGrayLevelEmphasis. The top 5 features positively collected with PDL1 < 1% were: cect-wavelet-HLL-glszm-SmallAreaEmphasis, cect-original firstorder-Kurtosis, nect-wavelet-LHL-firstorder-Minimum, nect-wavelet-LLH-firstorder-Skewness, and cect-wavelet-HLL-glszm-SizeZoneNonUniformityNormalized.

Fig. 6.

Coefficient values of features from the logistic regression model. The y-axis shows the feature names, while the x-axis denotes the coefficient values. Positive coefficients indicate a direct relationship between the feature and the probability of patients being PD-L1 positive (labeled as 1), meaning an increase in the feature value raises the likelihood of the model predicting a PD-L1 positive outcome. Negative coefficients suggest an inverse relationship, where an increase in the feature reduces the prediction probability

Fig. 7.

SHAP explanations for the Radiomics Logistic regression model. The lower plot illustrates the feature breakdown, with the red bar indicating an upward trend, while the blue bar represents a decrease in the average response for each feature. The larger the absolute value, the greater the impact on the model’s output

Discussion

Although EGFR-TKIs are the standard therapy for patients with EGFR-sensitive mutations, approximately 5–10% of patients still have no response to EGFR-TKIs due to primary resistance [15]. Previous studies have demonstrated that in EGFR-mutant NSCLC cells, PD-L1, through activating the TGF-β/Smad canonical signaling pathway, mediates the induction of epithelial-mesenchymal transition and contributes to primary resistance to EGFR-TKIs [16]. PD-L1 expression has been identified as a prognostic biomarker for these patients [17]. Moreover, recent research has reported distinct genomic characteristics between PD-L1–negative and PD-L1–positive EGFR-mutant patients, and demonstrated that combining EGFR-TKIs with chemotherapy can prolong median PFS in PD-L1–positive cases [18]. Therefore, our study aimed to predict PD-L1 expression in EGFR-mutated LUAD patients based on CT images and clinical information, providing a noninvasive, reliable, and reproducible means to anticipate their response to EGFR-TKIs, thereby providing evidence for clinical decision-making. Imaging-based analyses can capture abundant information from the entire tumor lesion and can be performed at multiple time points for treatment monitoring, thus avoiding spatial and temporal tumoral heterogeneity. Moreover, it can be carried out when a biopsy is not feasible, for example in frail patients, or when the sample is inadequate after routine diagnostic and molecular studies. Our research findings also demonstrate that models constructed based on CT image features can effectively predict PD-L1 expression status.

Several previous studies have employed deep learning or radiomics techniques to construct models for predicting the PD-L1 expression status in early stage or advanced NSCLC patients based on CT or PET/CT images, achieving AUC values ranging from 0.73 to 0.89 [19–22]. Only one study attempted to establish CT-based PD-L1 expression prediction models in EGFR-mutant NSCLC [23], but it lacked results interpretation, which is crucial for identifying CT features associated with the PD-L1 expression status. Our models outperformed previous radiomics models in predicting PD-L1 ≥ 1%.

Our research observed that image features associated with positive PD-L1 expression were predominantly textural features. In comparison to original texture features, wavelet-transformed features provided richer information on the intratumor heterogeneity than the original image [24]. These features, derived from GLCM, GLDM and GLSZM, predominantly delineate patterns or spatial distributions of voxel intensities, capturing texture heterogeneity and correlating with tumor heterogeneity [25]. The variables “DependenceVariance”, “MCC”, and “IMC2” all elucidate the relationships between different gray-level pixels within the CT images. A positive skewness signifies that the peak of the pixel distribution is skewed to the left with the tail extending toward the right side, and the majority of pixel values are lower. “LRLGLE” represented the means of the distribution of the long homogeneous runs with high gray levels following wavelet transformation.

Many previous radiomics models incorporated substantial information in the form of wavelet-transformed GLCM, GLRLM, and GLSZM features, which is in line with the model developed in this study. Jiang et al. found 9 that wavelet-based CT features, including wavelet-HHH-gldm-Dependence Variance and wavelet-HHL-glszm-GrayLevelNonUniformity, were connected with PD-L1 expression [26]. Shi et al. selected four original features, including Shape Maximum 2D Diameter Slice, and 5 wavelet-based features, including wavelet-HLL-glrlm-LRLGLE and wavelet-LHH-glcm-MCC, to predict PDL-1 expression status [27]. Bracci et al. found that glrlm-LRLGE was related to PD-L1 status in locally advanced or metastatic NSCLC patients, which was consistent with our results [28].

Our study identified that EGFR mutation type, MET amplification, TP53 mutation and smoking history were correlated with positive PD-L1 expression, in line with previous research findings. It has been verified that the EGFR L858R subtype tends to exhibit high PD-L1 expression and higher tumor mutational burden [29]; meanwhile, emerging data suggest that immunotherapy is more promising in patients with L858R mutations compared to those with 19DEL mutations [30]. Similarly, c-MET amplification is one of the mechanisms for EGFR-TKI resistance, upregulating PD-L1 expression and promoting the immune escape ability of EGFR-mutant NSCLC cells [31]. Both the Gene Expression Omnibus and The Cancer Genome Atlas databases demonstrated noticeably increased PD-L1 mRNA expression in the TP53 mutation subgroup compared with other gene mutations [32]. Calles et al. reported that smoking may lead to an elevated number of mutations, such as TP53, which in turn results in the upregulation of PD-L1 expression [33]. Overall, patients with L858R mutation, MET amplification, TP53 mutation, and a smoking history tend to have a high tumor mutational burden level and PD-L1 expression.

However, there are several limitations to our study. First, the PD-L1 expression status was measured through biopsy rather than surgical specimens due to the inoperable stage of the enrolled patients with locally advanced or metastatic disease. Therefore, the obtained PD-L1 expression status may not completely reflect the true PD-L1 expression of the entire tumor. Second, as a retrospective study based on electronic medical records, information bias and selection bias are inevitable. In this study, the proportion of positive PD-L1 expression was higher than that reported, which may be due to the small sample size. Third, clinical outcome data, including progression-free survival and overall survival, were not included in the analysis, and the potential role of the radiomic model in predicting therapy response was not explored. Finally, the main limitation of this study is the absence of external validation, which restricts the generalizability of the findings. Validating the model in an independent cohort will be essential to confirm its robustness and enhance its applicability in broader clinical settings. These aspects will be addressed in our future studies.

Interestingly, previous research has shown that the discriminative ability of CT-based features is superior to PET imaging in predicting PD-L1 expression status [26]. In routine clinical practice, enhanced CT, rather than PET/CT, is the predominant method utilized for assessing the progression of NSCLC. This indicates that efforts should be made to enhance the effectiveness of enhanced CT-based radiomic models in predicting PD-L1 expression status. Our study adds further evidence to this field. Most current studies have focused on predicting PD-L1 expression based on single-modality imaging, such as CT or PET/CT, or through the combination of PET/CT and metabolomics. Our study suggests that genetic mutations and clinical characteristics also play important roles in predicting PD-L1 expression, indicating that future research could place greater emphasis on multi-omics approaches for integrated prediction of PD-L1 expression. Furthermore, our study demonstrates that radiomic features can predict PD-L1 expression in patients with EGFR-mutation. Since CT scans are routinely performed during treatment, this approach holds promise for noninvasive, dynamic prediction of PD-L1 expression in the future, and larger sample sizes and multiple-center studies will be needed to validate our results.

Conclusion

Our study developed interpretable machine learning models to predict the PD-L1 expression status in advanced EGFR-mutant NSCLC patients, with the radiomic model displaying excellent performance. Machine learning can serve as an adjunctive tool for estimating PD-L1 expression status when a PD-L1 biopsy fails or when the medical condition progresses, and it may assist in guiding immunotherapy decisions and could contribute to treatment planning, pending further prospective validation.

Abbreviations

AUC

Area under the curve

CT

Computed tomography

CECT

Contrast enhanced CT

EGFR

Epidermal growth factor receptor

EGFR-TKIs

Epidermal growth factor receptor tyrosine kinase inhibitors

IHC

Immunohistochemistry

LDA

Linear discriminate analysis

LUAD

Lung adenocarcinoma

NECT

Non-contrast enhanced CT

NSCLC

Non-small cell lung cancer

PD-1

Programmed cell death-1

PD-L1

Programmed cell death ligand-1

ROI

Region of interest

RFE

Recursive feature elimination

RF

Random forest

SHAP

SHapley Additive exPlanations

SVM

Supporter vector machine

Authors’ contributions

J.J.L collected clinical data and wrote the manuscript, Z.Z.Y analyzed the data and designed the figures, L.Z.L, F.T, T.H.Z and C.Z also collected clinical data. Y.P and Z.Z.Y provided critical comments and suggestions, Z.M.Z and B.X.T revised the manuscript. X.Y.Z designed the study and finalized the manuscript. All authors read and approved the final version of the manuscript.

Funding

This work was supported by grants from the Natural Science Foundation of Chongqing (CSTB2024NSCQ-MSX0414) to Xiaoyue Zhang, and Wu Jieping Medical Foundation (No. 320.6750.2020-01-40) to Benxu Tan.

Data availability

All relevant code and data of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

The study was reviewed and approved by the Ethics Committees of the Second Hospital of Chongqing Medical University. Informed consent was obtained from each patient prior to data collection.

Consent for publication

All authors agree with the content of the manuscript.

Competing interests

The authors declare no competing interests.