Sage Journals: Discover world-class research

Abstract

Objective

To develop and validate a CT-based radiomics model for predicting sequential immunotherapy response after concurrent radiochemotherapy (CCRT) in patients with unresectable stage III non-small cell lung cancer (NSCLC).

Methods

The study retrospectively included 71 patients who received sequential immunotherapy after concurrent chemoradiotherapy (CCRT) between January 2019 and December 2022, while prospectively including 17 additional patients between January 2023 and July 2023. The study documented each patient's tumor response and prognosis within two months of completing treatment. Patients were then categorized based on their treatment response, resulting in the identification of two distinct groups: treatment-responsive group and treatment-insensitive group. First, ITK-SNAP software was used to delineate the primary tumor lesions in the lung window and define a region of interest (ROI). Second, features were extracted using Python (version 3.6) and filtered using Least absolute shrinkage and selection operator regression. Third, radiological models were built using six machine learning algorithms: logistic regression (LR), discriminant analysis (DA), neural network (NN), random forest (RF), support vector machine (SVM) and K-Nearest Neighbour (KNN). Finally, Kaplan-Meier survival analysis was performed for high- and low-risk patients predicted by radiomic modeling.

Results

Based on the performance of radiomics models constructed by various machine learning algorithms in the prospective validation set, the LR with the highest AUC value (AUC: 90.00%) was finally selected, which also performed well in the independent test set (AUC: 84.96%). Risk stratification of patients based on the radiomic model constructed by LR was excellent for PFS (P = 0.001) and OS (P = 0.019) in the training set, PFS (P = 0.010) and OS (P = 0.028) in the prospective validation set, and PFS (P = 0.014) and OS (P = 0.041) in the test set.

Conclusion

Pretreatment CT-based radiomics model accurately and efficiently predicts treatment response and risk stratification in patients with unresectable stage III NSCLC treated with concurrent chemoradiotherapy and sequential programmed death-1 inhibitor therapy. Prior to prospective data collection, the study was registered with the China Clinical Trial Registry under the trial registration name: Prediction of concurrent chemoradiotherapy efficacy and its related molecular signaling pathway by medical artificial intelligence model based on CT of lung cancer, with the registration number: ChiCTR2100053175 (https://www.chictr.org.cn/showproj.html?proj = 136872).

Keywords

CCRT (concurrent radiochemotherapy)CT (computed tomography)immunotherapy NSCLC (non-small cell lung cancer)radiomics

Introduction

The incidence and mortality of lung cancer remain high.¹ From a pathological perspective, lung cancer is categorized into two main subtypes: non-small cell lung cancer (NSCLC) and small cell lung cancer. NSCLC constitutes approximately 85% of all cases. About 30% of patients with NSCLC present with stage III disease at initial diagnosis,² with most ineligible for surgery. Unlike stage IV NSCLC, stage III NSCLC represents a highly heterogeneous group of patients with wide variations in tumor size and lymph node involvement, leading to significant differences in treatment modalities and prognosis.³ However, there is hope for a cure in this group of patients without distant metastases, and selecting an effective treatment option is a significant challenge for all oncologists. For patients with inoperable stage III NSCLC, the traditional standard of care involves curative radiotherapy and chemotherapy.⁴ Several studies have demonstrated that concurrent radiotherapy (CCRT) is superior to sequential radiotherapy in such patients,^5,6 with a 5-year progression-free survival (PFS) rate of only 18.3% and a 5-year overall survival (OS) rate of 32.1%.⁷ This indicates that CCRT alone is insufficient to address the treatment requirements of patients with inoperable stage III NSCLC. The advent of immunotherapy has led to large phase III trials, such as PACIFIC and GEMSTONE-301, which successfully established the standard of care for immune-consolidating therapy following radical radiotherapy and chemotherapy in inoperable stage III lung cancer.^8–12 However, all patients are unsuitable for this treatment regimen, and in clinical practice, some patients may not respond to sequential programmed death-1 (PD-1) inhibitor therapy following CCRT. Besides, the sensitivity of patients to the treatment can significantly influence their prognosis. Previous research has shown that patients’ response to therapy is critical to their PFS and OS outcomes.^13,14 Therefore, new biomarkers are urgently needed to predict the effectiveness of sequential PD-1 inhibitor therapy after CCRT and to identify suitable candidates.

In conventional medicine, imaging is an important part of the monitoring of the efficacy of treatments in cancer patients. As technology advances, artificial intelligence is increasingly applied to cancer medical imaging.^15,16 However, much of the current research focuses on the therapeutic sensitivity of single radiotherapy, chemotherapy, or immunotherapy.^17–22 Currently, no model exists to predict the treatment response in unresectable stage III NSCLC patients who receive CCRT followed by sequential PD-1 inhibitor therapy. In model selection for small-to-medium sample sizes, traditional machine learning algorithms (eg, logistic regression) can effectively prevent overfitting through regularization. In contrast, deep learning algorithms are prone to overfitting due to their high parameter complexity. Furthermore, deep learning has inherent black-box limitations: implicit layer features lack biologically interpretable mappings, and training requires Graphics Processing Unit clusters, which is computationally intensive. Conversely, traditional machine learning algorithms provide explicit feature weights to support clinical decisions while maintaining computational efficiency.

The aim of this study is to develop and validate radiomics-based machine learning models that predict the response of inoperable stage III NSCLC patients to sequential PD-1 inhibitors following CCRT, using preprocessed CT scans. These models aim to assist clinicians in identifying suitable candidates for this treatment strategy.

Materials and Methods

Patients

The training set consisted of 38 cases collected retrospectively from Hospital A (January 2019 to December 2022); the prospective validation set included 17 cases collected prospectively from Hospitals A, B, C, and D (January to July 2023); and the test set comprised 33 cases retrospectively collected from Hospitals B, C, and D (January 2019 to December 2022). The primary endpoint of this study was the tumor progression within two months after completing CCRT, followed by sequential PD-1 inhibitor therapy, while the secondary endpoints were the PFS and OS in patients. Due to the retrospective nature of data collection and the anonymization protocol approved by the ethics committee, verbal consent was deemed appropriate. Verbal consent was formally documented by research staff through audio recordings and written notes, confirming participants’ agreement to: use of their de-identified medical data for this study and future research; analysis of treatment outcomes without personal identifiers. For prospective data, written consent was obtained, specifying identical terms.

Study Design and Treatment

Figure 1 presents the experimental flow chart. Inclusion criteria: (1) NSCLC confirmed by pathological diagnosis; (2) no indication for surgery; (3) stage III NSCLC according to the eighth edition of the Tumor-Node-Metastasis classification²³; (4) no previous specialized cancer treatments (such as neoadjuvant chemotherapy, radiotherapy, targeted therapy, interventional therapy, or immunotherapy) before this study; (5) patients who received standard chest intensity-modulated radiotherapy (radiotherapy dose: 60-66 Gy/ 30-33 fractions) combined with a synchronous chemotherapy regimen (for squamous cell carcinoma, the chemotherapy regimen consisted of weekly paclitaxel combined with cisplatin; chemotherapy regimen for non-squamous cell carcinoma involved pemetrexed combined with cisplatin administered once every three weeks)^24,25; (6) after completing CCRT, patients received at least 2 cycles of a PD-1 inhibitor (Tislelizumab or Sintilimab 200 mg/3 weeks)^26,27; (7) patients underwent CT-enhanced imaging of the chest and imaging histology of other sites obtained one month before CCRT treatment and two months after immunotherapy. The exclusion criteria included the presence of other malignancies, incomplete or poor-quality CT images, and inadequate follow-up data. In this study, PFS was measured from the commencement of CCRT to either the first recurrence/ metastasis of lung cancer or the last follow-up on December 31, 2023. OS was characterized as the interval from the commencement of CCRT to death or the final follow-up date (December 31, 2023).

Figure 1.

Experimental Flowchart.

Patient efficacy was assessed blindly within two months after treatment completion, according to Immune-related Response Criteria based solely on CT images and follow-up records, excluding model-predicted outcomes.²⁸ The treatment results were divided into a treatment-sensitive group (immune complete response/ immune partial response) and a treatment-resistant group (immune stable disease/ immune unconfirmed progressive disease/ immune confirmed progressive disease). To standardize the collected contrast-enhanced chest CT images, three-dimensional reconstruction was performed using MATLAB 2014b (https://ww2.mathworks.cn/), with the slice thickness adjusted to 1 mm. A radiotherapist with ten years of experience in cancer treatment delineated the primary tumor lesions on the CT images layer by layer (Window width −1600 Hounsfield units [HU] and window height −600 HU) using ITK-SNAP software (version 3.8.0; http://www.itksnap.org) following the guidelines of the International Commission on Radiation Units 83 report.²⁹ After completing the annotation for each layer, the corresponding region of interest (ROI) was designated.

Statistical Analysis

The radiomics features were extracted from the ROI of patients. Initially, the data were standardized for radiomics features using the Z-score method. Thereafter, Lasso regression with 5-fold cross-validation folds was used to select relevant features. Logistic regression (LR), Discriminant Analysis (DA), Neural Network (NN), Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN) were applied to construct radiomics models using the selected features. The selection of tools for the construction of radiomics models was based on their performance in terms of area under the curve (AUC) values in the training set. Accuracy, precision, recall, specificity, F1-score, receiver operating characteristic (ROC), AUC, Clinical Decision Curve Analysis (DCA), and Calibration curve were used to evaluate the model's predictive performance. Patients were categorized into high-risk and low-risk groups using predicted probabilities and cutoff values from the radiomics model, and their PFS and OS were analyzed using the Kaplan-Meier method. The predictive model was developed using Python software (version 3.6; https://github.com/radiomics/pyradiomics). R software (version 4.3.0; http://www.R-project.org) and SPSS software (version 26.0; IBM Corporation, Armonk, New York, USA) were used for all statistical analyses. Additionally, PyRadiomics 3.0 was employed as the feature extraction framework, strictly adhering to its default parameters. The reporting of this study adheres to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines.³⁰

Results

88 patients were eligible for inclusion in this study. The training set retrospectively collected data containing 38 Hospital A patients treated between January 2019 and December 2022. The prospective validation set collected 4 Hospital B patients, 4 Hospital C patients, 1 Hospital D patient, and 8 Hospital A patients treated between January and July 2023. The test set collected 15 Hospital B patients, 11 Hospital C patients, 7 Hospital D patients treated between January 2019 and December 2022. The results presented in Table 1 indicate that no statistically significant differences were observed between the two patient groups regarding treatment response, immunotherapy drugs, gender, ethnicity, age, smoking history, pathological type, T stage and N stage (P > 0.05). From each patient's ROI, 1834 radiomic features were extracted, encompassing shape, texture, first-order, and other radiomic features. Following Lasso regression screening, six radiomic features were identified as associated with treatment efficacy (Table 2 and Figure 2). Supplementary Table 1 presents the weights and clinical significance of the six screened radiomics features.

Figure 2.

Lasso Regression Screening for Radiomics Features. (a) Lasso Coefficient Profile of Radiomics Features. (b) Cross Curve Validation.

Table 1.

Subgroup Difference Table.

Variables	Total (n = 88)	Training set (n = 38)	Prospective validation set (n = 17)	Test set (n = 33)	P
Gender, n(%)
Female	18 (20.45)	9 (23.68)	2 (11.76)	7 (21.21)	0.593
Male	70 (79.55)	29 (76.32)	15 (88.24)	26 (78.79)
Ethnic Groups, n(%)
Miao	25 (28.41)	13 (34.21)	5 (29.41)	7 (21.21)	0.052
Dong	19 (21.59)	11 (28.95)	4 (23.53)	4 (12.12)
Han	40 (45.45)	11 (28.95)	7 (41.18)	22 (66.67)
Others	4 (4.55)	3 (7.89)	1 (5.88)	0 (0.00)
Age, n(%)					0.321
≤60	54 (61.36)	26 (68.42)	8 (47.06)	20 (60.61)
＞60	34 (38.64)	12 (31.58)	9 (52.94)	13 (39.39)
Smoking History, n(%)					0.285
No	36 (40.91)	19 (50.00)	5 (29.41)	12 (36.36)
Yes	52 (59.09)	19 (50.00)	12 (70.59)	21 (63.64)
Histology, n(%)					0.180
LUSC	53 (60.23)	24 (63.16)	13 (76.47)	16 (48.48)
LUAD	34 (38.64)	13 (34.21)	4 (23.53)	17 (51.52)
Other	1 (1.14)	1 (2.63)	0 (0.00)	0 (0.00)
Efficacy, n(%)					0.537
Sensitivity	43 (48.86)	19 (50.00)	10 (58.82)	14 (42.42)
Resistant	45 (51.14)	19 (50.00)	7 (41.18)	19 (57.58)
PD-1 Inhibitor, n(%)					0.397
Tislelizumab	49 (55.68)	22 (57.89)	7 (41.18)	20 (60.61)
Sintilimab	39 (44.32)	16 (42.11)	10 (58.82)	13 (39.39)
T staging, n(%)					0.642
T1	5 (5.68)	2 (5.26)	0 (0.00)	3 (9.09)
T2	20 (22.73)	8 (21.05)	5 (29.41)	7 (21.21)
T3	16 (18.18)	5 (13.16)	5 (29.41)	6 (18.18)
T4	47 (53.41)	23 (60.53)	7 (41.18)	17 (51.52)
N staging, n(%)					0.051
N1	9 (10.23)	6 (15.79)	0 (0.00)	3 (9.09)
N2	33 (37.50)	9 (23.68)	11 (64.71)	13 (39.39)
N3	46 (52.27)	23 (60.53)	6 (35.29)	17 (51.52)
TNM staging, n(%)					0.037
IIIA	17 (19.32)	10 (26.32)	2 (11.76)	5 (15.15)
IIIB	42 (47.73)	11 (28.95)	12 (70.59)	19 (57.58)
IIIC	29 (32.95)	17 (44.74)	3 (17.65)	9 (27.27)

Abbreviations: LUSC, Lung squamous cell carcinoma; LUAD, Lung adenocarcinoma; PD-1 Inhibitor, Programmed Death-1 Inhibitor; T staging, Tumor staging; N staging, Node staging; TNM, Tumor-Node-Metastasis staging.

Table 2.

Statistical Analysis of Selected Radiomics Features.

	Treatment sensitive (N = 43)	Treatment resistant (N = 45)	Overall (N = 88)	P
Feature 1:
exponential_firstorder_Interquartile-Range
Mean (SD)	0.270 (1.29)	−0.270 (0.472)	0.000000000000000387 (1.01)	0.0398
Median [Min, Max]	−0.405 [−0.925, 4.30]	−0.481 [−0.796, 1.41]	−0.468 [−0.925, 4.30]
Feature 2:
exponential_ngtdm_Complexity
Mean (SD)	0.324 (1.35)	−0.324 (0.137)	0.000000000000000291 (1.01)	0.00937
Median [Min, Max]	−0.268 [−0.445, 4.71]	−0.289 [−0.445, 0.188]	−0.268 [−0.445, 4.71]
Feature 3:
lbp.3D.k_glszm_SmallAreaLowGray-LevelEmphasis
Mean (SD)	0.368 (0.786)	−0.368 (1.07)	0.00000000000000106 (1.01)	0.00234
Median [Min, Max]	0.331 [−1.35, 2.48]	−0.296 [−2.57, 1.89]	0.0481 [−2.57, 2.48]
Feature 4:
lbp.3D.m2_glcm_Correlation
Mean (SD)	−0.300 (0.874)	0.300 (1.05)	0.000000000000000524 (1.01)	0.0184
Median [Min, Max]	−0.503 [−2.34, 2.12]	0.207 [−1.32, 3.30]	−0.200 [−2.34, 3.30]
Feature 5:
wavelet.HHH_glszm_SizeZoneNon-UniformityNormalized
Mean (SD)	0.227 (1.09)	−0.227 (0.870)	0.00000000000000284 (1.01)	0.104
Median [Min, Max]	−0.0139 [−1.85, 2.96]	−0.320 [−2.85, 1.69]	−0.132 [−2.85, 2.96]
Feature 6:
waveletLHH_firstorder_Skewness
Mean (SD)	0.319 (1.29)	−0.319 (0.423)	0.000000000000000449 (1.01)	0.011
Median [Min, Max]	−0.00131 [−1.31, 5.02]	−0.327 [−1.19, 0.541]	−0.175 [−1.31, 5.02]

Abbreviations: SD, standard deviation; Min, minimum; Max, maximum.

The DeLong test (Table 3) comparing ROC performance across algorithms in the prospective validation set showed no significant differences between KNN, LR, DA, RF, and SVM (P＞ 0.05), with the neural network (NN) exhibiting slightly lower performance. Given the limitations of the small sample size, the logistic regression model was ultimately selected for its optimal balance between interpretability and stability, and it demonstrated the highest AUC value in the prospective validation set. Figure 3 shows the ROC curves for the validation and test sets. As detailed in Table 4, the radiomics model developed via LR demonstrated strong predictive consistency across all datasets: training set (AUC = 0.895, 95% CI 0.79–1), prospective validation set (AUC = 0.900, 95% CI 0.75–1), and test set (AUC = 0.850, 95% CI 0.72–0.98). The model accuracy, precision, recall, specificity, and F1-score were all satisfactory. Given the limited sample size, we performed calibration-curve analysis to assess the reliability of the model's predicted probabilities (Figure 4) and clinical-decision-curve analysis to evaluate real-world utility (Figure 5).

Figure 3.

ROC Curve. (a) ROC Curve for Validation Set. (b) ROC Curve for Test Set. KNN, K-Nearest Neighbor; LR, Logistic Regression; DA, Discriminant Analysis; NN, Neural Network; RF, Random Forest; SVM, Support Vector Machine.

Figure 4.

Calibration Plot. (a) Training Set Calibration Plot. (b) Prospective Validation Set Calibration Plot. (c) Test Set Calibration Plot.

Figure 5.

Analysis of Clinical Decision Curves Comparing Radiomics Model with TNM Stage in the Test Set.

Table 3.

DeLong Test p-Value Table.

Variables	KNN	LR	DA	NN	RF	SVM
KNN		0.102	0.528	0.274	0.601	0.801
LR	0.102		0.188	0.047	0.087	0.203
DA	0.528	0.188		0.107	0.168	0.65
NN	0.274	0.047	0.107		0.757	0.254
RF	0.601	0.087	0.168	0.757		0.534
SVM	0.801	0.203	0.65	0.254	0.534

Abbreviations: KNN, K-Nearest Neighbor; LR, Logistic Regression; DA, Discriminant Analysis; NN, Neural Network; RF, Random Forest; SVM, Support Vector Machine.

Table 4.

Predictive Performance of Models Constructed Using Logistic Regression.

Statistical value	Training set	Prospective validation set	Test set
Accuracy	84.21%	82.35%	84.21%
Precision	80.95%	83.33%	92.31%
Recall	89.47%	71.43%	63.16%
Specificity	73.68%	90.00%	92.86%
F1-score	85.00%	76.92%	75.00%
AUC	89.47%	90.00%	84.96%

Abbreviations: AUC, area under the curve.

In the entire cohort, 43 patients exhibited sensitivity to CCRT followed by sequential immunotherapy, whereas 45 patients demonstrated treatment resistance to this treatment regimen. Based on the most recent follow-up, a total of 60 patients were diagnosed with tumor progression, with a median PFS of 8.5 months (95% confidence interval [CI]: 7.0-11.0). Among these, 45 patients died, with a median OS of 12.0 months (95% CI: 10.0-17.0). Using a cutoff value of 0.602 derived from the training set and the predictive probabilities of the radiomics model to categorize patients into high- and low-risk groups, we observed significant differences in PFS and OS between risk groups: Figure 6a-b (training set), 6c-d (prospective validation set), and 6e-f (test set).

Figure 6.

Survival Analysis Curve. (a)The Training Set Revealed a Significant Difference in PFS Between Patients in Low- and High-Risk Groups. (b) The Training Set Exhibited a Significant Difference in OS Between Patients in Low- and High-Risk Groups. (c) The Prospective Validation Set Revealed a Significant Difference in PFS Between Patients in Low- and High-Risk Groups. (d) The Prospective Validation Set Indicated a Significant Difference in OS Between Patients in Low- and High-Risk Groups. (e)The Test Set Revealed a Significant Difference in PFS Between Patients in Low- and High-Risk Groups. (f) The Test Set Exhibited a Significant Difference in OS Between Patients in Low- and High-Risk Groups.

Discussion

Our study demonstrated that the machine learning radiomics model based on pretreatment CT performed well in the training, validation and test sets to predict treatment response and survival risk stratification in patients with unresectable stage III NSCLC receiving CCRT followed by sequential PD-1 inhibitors (P < 0.05). This suggests that pretreatment CT examination can non-invasively and effectively predict patient response to sequential PD-1 inhibitor therapy following CCRT without the need for hematology tests and biopsy. This is the first study to evaluate a machine-learning radiomics model based on pretreatment CT for predicting treatment efficacy in patients with unresectable stage III NSCLC treated with sequential PD-1 inhibitors following CCRT.

Studies have indicated that approximately 30% of NSCLC patients are diagnosed with stage III at the time of their initial visit,³¹ with the majority being ineligible for surgical intervention. For patients with unresectable stage III NSCLC, CCRT is the primary treatment option,⁴ with a 5-year PFS rate of just 18.3% and an OS rate of 32.1%.³² The PACIFIC trial series’ successive publications helped clinicians recognize immunotherapy's safety and effectiveness as a consolidation treatment after CCRT.^8–11 Chemotherapeutic agents not only kill cancer cells but also act synergistically with immunotherapy by inducing immunogenic cell death, increasing cancer cell antigen expression, and regulating gene expression.^33–35 Radiotherapy and immunotherapy also work synergistically. Radiation therapy enhances tumor immunotherapy by facilitating the release of tumor-associated antigens, optimizing the tumor immune microenvironment, and inducing hypoxia within tumor cells; immune checkpoint inhibitors (ICI) can eliminate radiation resistance induced by radiotherapy.^36,37 Some studies have demonstrated that ICI can not only attack cancer cells by activating T cells but also regulate the tumor microenvironment by normalizing tumor blood vessels, thereby increasing radiation sensitivity.^38,39 Consequently, combining radiotherapy, chemotherapy, and immunotherapy can produce stronger anti-tumor effects. Based on these research findings, durvalumab has been approved worldwide as the standard treatment for patients with stage III NSCLC who are ineligible for surgical resection.²⁵

However, due to individual variability, only certain patients may benefit from sequential PD-1 inhibitor therapy following CCRT, and if patients irresponsive to this regimen receive this therapy, it may result in both physical and financial losses, which can have a negative impact on the patient. Therefore, identifying sensitive patients who may benefit from sequential PD-1 inhibitors following CCRT will have important implications for individualized treatment. It has been proven through studies that radiomics is a non-invasive and accessible approach for predicting how a patient will respond to treatment.^40,41 However, given that immunotherapy is a novel treatment option, a significant number of studies have largely focused on using radiomics to predict the effectiveness of single radiotherapy, chemotherapy, immunotherapy, or dual therapy.^42–45

Due to the support of medical insurance policies in China, most patients requiring PD-1 inhibitors opt for tislelizumab and sindilimab, which have demonstrated comparable efficacy and safety to other PD-1 inhibitors.^46–51 Therefore, all patients included in this study used these two immunotherapy drugs, and statistical analysis revealed no significant difference in efficacy between the two (P > 0.05). We screened six radiomic features and detailed their clinical implications in Supplementary Table 1. These features collectively suggest that increased intratumoral textural complexity is associated with enhanced treatment sensitivity.⁵² The three most critical radiomic features are as follows: 1. “waveletLHH_firstorder_Skewness” : This feature correlates with density asymmetry in tumor gray-scale distribution. Predominance of low-density areas (eg, necrosis, edema) may indicate immunotherapeutic responsiveness, as necrotic antigen release activates T cells.⁵³ 2. “lbp3Dk_glszm_SmallAreaLowGrayLevelEmphasis”: It quantifies 3D small low-density regions. Micro-necrotic zones release damage-associated molecular patterns (DAMPs, eg, HMGB1, ATP), activating antigen-presenting cells to promote T-cell recruitment. Studies suggest micro-necrosis also induces chemokines (CXCL9/10), directly attracting CD8+ T cells.^54–56 3. “exponential_ngtdm_Complexity” : High values reflect local texture disorder, suggesting strong spatial heterogeneity. Elevated complexity is positively correlated with tumor mutational burden and potential sensitivity to ICI.⁵⁷

Radiomics modeling successfully predicted true response to sequential PD-1 inhibitors after CCRT in 82.35% (14/17) of patients in the prospective validation set and accurately identified true response in 75.76% (25/33) of patients in the test set. Clinical Decision Curve Analysis revealed that the radiomic model provided a significantly higher net benefit compared to TNM staging across the clinically relevant risk threshold range (0.1-0.4). This superiority indicates that the radiomic approach reduces overtreatment (ie, false positives) while maintaining effective interventions for high-risk patients (ie, true positives). Prognostic analysis following patient stratification into high- and low-risk groups based on the cutoffs calculated by the LR radiomics model demonstrated that patients were effectively stratified for PFS and OS (P < 0.05). This suggests that a better prognosis can be achieved for patients with unresectable stage III NSCLC who are sensitive to treatment with sequential PD-1 inhibitors following CCRT. It also indicates that the radiomics model can efficiently and accurately predict the efficacy of sequential PD-1 inhibitors after CCRT, as well as stratify risk for PFS and OS, which may help achieve individualized treatment.

This study has several limitations. First, to minimize bias, all ROIs were manually outlined by a radiation oncologist; however, this approach limits the ability to assess the reproducibility of radiological feature extraction. Future studies will involve multiple annotators to evaluate inter-observer variability. Second, existing studies suggest that machine learning algorithms require a minimum sample size of 80 to achieve a root mean square error < 0.01 and at least 30 samples to ensure statistical validity. Empirical data further indicate that the recommended number of features for binary classification models should be approximately one-tenth of the sample size.⁴⁰ Training Set (N = 38): Enhanced via 5-fold cross-validation (effective sample size = 190 iterations), meeting the ‘10 events per variable’ threshold (38/6 ≈ 6.3 events per feature). Validation (N = 17) & Test Sets (N = 33): Combined sample (N = 50) exceeds the minimum requirement (N = 30) for preliminary validation per FDA AI/ML guidance.⁵⁸ Although the feature count and sample size in this multicenter study met these general requirements, the focus on patients with unresectable stage III NSCLC inherently limited the cohort size. As an increasing number of stage IV NSCLC patients are now being treated with sequential PD-1 inhibitors following CCRT, additional data should be included in future studies to validate the generalizability and accuracy of the model. We plan to expand collaboration with multiple centers to establish a CT protocol-sharing platform and standardize scanning parameters. Third, this study is currently limited to radiomics data, and multimodal data such as tissue sections and genomics will be included in the future to further explore the potential mechanism of this therapy and to reveal the associations between the radiomics features and the biological properties of tumors. Fourth, existing hospital PACS systems lack native support for radiomics analysis, leading to two key bottlenecks: Data Format Compatibility: DICOM images must be converted to NIfTI or PNG formats prior to model input, prolonging preprocessing time. Real-Time Inference Delay: Manual export of images to external platforms for analysis fails to meet clinical real-time requirements. While fully resolving these challenges remains difficult, we are collaborating with PACS vendors to develop lightweight plugins for automated image retrieval and preprocessing.

Conclusions

In this study, a machine learning-based radiomics model was successfully developed and validated to non-invasively and accurately predict the efficacy of sequential PD-1 inhibitors in patients with unresectable stage III NSCLC undergoing CCRT with risk stratification to guide individualized treatment.

Hospital Pronouns

The Second Affiliated Hospital of Guizhou Medical University stands for Hospital A; Guiyang Pulmonary Hospital stands for Hospital B; Qiandongnan Prefecture People's Hospital stands for Hospital C; and Qiannan Prefecture Hospital of Traditional Chinese Medicine stands for Hospital D.

Supplemental Material

sj-docx-1-tct-10.1177_15330338251351109 - Supplemental material for Pretreatment CT-Based Machine Learning Radiomics Model Predicts Response in Inoperable Stage III NSCLC Treated with Concurrent Radiochemotherapy Plus PD-1 Inhibitors

Supplemental material, sj-docx-1-tct-10.1177_15330338251351109 for Pretreatment CT-Based Machine Learning Radiomics Model Predicts Response in Inoperable Stage III NSCLC Treated with Concurrent Radiochemotherapy Plus PD-1 Inhibitors by Ya Li, Bachelor, Min Zhang, Bachelor, Yong Hu, MM, Bo Du, Bachelor, Youlong Mo, Bachelor, Tianchu He, MM, Mingdan Zhao, Bachelor, Benlan Li, Bachelor, Ji Xia, Bachelor, Zhongjun Huang, Bachelor, Fangyang Lu, MD, Zhen Huang, Bachelor, Bing Lu, MD, and Jie Peng in Technology in Cancer Research & Treatment

Footnotes

Abbreviations

Acknowledgments

We thank all members for their invaluable contributions.

ORCID iD

Jie Peng

Ethics Statement

The present study was conducted in strict accordance with the principles of the Declaration of Helsinki. The study was approved by the Ethics Committee of the Second Affiliated Hospital of Guizhou Medical University (Approval No. 2020-LS-03), the Ethics Committee of Guiyang Pulmonary Hospital (Approval No. 2022-KY-24), the Ethics Committee of Qiandongnan Prefecture People's Hospital (Approval No. 2023-LS-21), and the Ethics Committee of Qiannan Prefecture Hospital of Traditional Chinese Medicine (Approval No. 2023-LS-12).

Author Contributions

Conception and design: YL, BL, and JP. Administrative support: BL, JP, and FY L. Provision of study materials and enrollment of patients: YL, MZ, YH, BD, YL M, TC H, and MD Z. Collection and assembly of data: YL. Data analysis and interpretation: YL, MZ, and JP. Manuscript writing: all authors. All authors contributed to the article and approved the submitted version.

Funding

This work was supported by the Qian Dong Nan Science and Technology Program (No. qdnkhJz [2023] 14), Scientific Research Project of Guizhou Provincial Health and Wellness Commission (No. gzwkj 2024-099), and Spark Program (No. XHJH-0048).

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

Raw data supporting the conclusions of this paper are available from the authors.

Supplemental Material

Supplemental material for this article is available online.

References

Siegel

Kimberly

Nikita Sandeep

Ahmedin

. Cancer statistics, 2023. CA Cancer J Clin. 2023;73(1):17-48. DOI:https://doi.org/10.3322/caac.21763

Barta

Powell

Wisnivesky

. Global epidemiology of lung cancer. Ann Glob Health. 2019;85(1):8. DOI:https://doi.org/10.5334/aogh.2419

Puri

Saltos

Perez

, et al. Locally advanced, unresectable non-small cell lung cancer. Curr Oncol Rep. 2020;22(4):31. DOI:https://doi.org/10.1007/s11912-020-0882-3

Postmus

Kerr

Oudkerk

, et al. Early and locally advanced non-small-cell lung cancer (NSCLC): ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Ann Oncol. 2017;28(0):iv1-iv21. DOI:https://doi.org/10.1093/annonc/mdx222

Curran

Jr Paulus

Langer

, et al. Sequential vs concurrent chemoradiation for stage III non–small cell lung cancer: Randomized phase III trial RTOG 9410. J Natl Cancer Inst. 2011;103(19):1452-1460. DOI:https://doi.org/10.1093/jnci/djr325

Aupérin

Le Péchoux

Rolland

, et al. Meta-analysis of concomitant versus sequential radiochemotherapy in locally advanced non–small-cell lung cancer. J Clin Oncol. 2010;28(13):2181-2190. DOI:https://doi.org/10.1200/JCO.2009.26.2543

Goldstraw

Chansky

Crowley

, et al. The IASLC lung cancer staging project: Proposals for revision of the TNM stage groupings in the forthcoming (eighth) edition of the TNM classification for lung cancer. J Thorac Oncol. 2016;11(1):39-51. DOI:https://doi.org/10.1016/j.jtho.2015.09.009

Antonia

Villegas

Daniel

, et al. Durvalumab after chemoradiotherapy in stage III non–small-cell lung cancer. N Engl J Med. 2017;377(20):1919-1929. DOI:https://doi.org/10.1056/NEJMoa1709937

Antonia

Villegas

Daniel

, et al. Overall survival with durvalumab after chemoradiotherapy in stage III NSCLC. N Engl J Med. 2018;379(24):2342-2350. DOI:https://doi.org/10.1056/NEJMoa1809697

10.

Spigel

Faivre-Finn

Gray

, et al. Five-year survival outcomes from the PACIFIC trial: Durvalumab after chemoradiotherapy in stage III non–small-cell lung cancer. J Clin Oncol. 2022;40(12):1301-1311. DOI:https://doi.org/10.1200/jco.21.01308

11.

Zhou

Chen

Jiang

, et al. Sugemalimab versus placebo after concurrent or sequential chemoradiotherapy in patients with locally advanced, unresectable, stage III non-small-cell lung cancer in China (GEMSTONE-301): Interim results of a randomised, double-blind, multicentre, phase 3 trial. Lancet Oncol. 2022;23(2):209-219. DOI:https://doi.org/10.1016/s1470-2045(21)00630-6

12.

Ettinger

Wood

Aisner

, et al. Non-small cell lung cancer, Version 2.2021 featured updates to the NCCN guidelines. J Natl Compr Cancer Network. 2021;19(3):254-266. DOI:https://doi.org/10.6004/jnccn.2021.0013

13.

Peng

Zou

Zhang

, et al. A novel sub-regional radiomics model to predict immunotherapy response in non-small cell lung carcinoma. J Transl Med. 2024;22(1):87. DOI:https://doi.org/10.1186/s12967-024-04904-6

14.

Peng

Huang

, et al. Development and validation of a pyradiomics signature to predict initial treatment response and prognosis during transarterial chemoembolization in hepatocellular carcinoma. Front Oncol. 2022;12(0):853254. DOI:https://doi.org/10.3389/fonc.2022.853254

15.

Hosny

Schabath

, et al. Artificial intelligence in cancer imaging: Clinical challenges and applications. CA Cancer J Clin. 2019;69(2):127-157. DOI:https://doi.org/10.3322/caac.21552

16.

Shankar

Schöder

Sharon

, et al. Harnessing imaging tools to guide immunotherapy trials: Summary from the National Cancer Institute Cancer Imaging Steering Committee workshop. Lancet Oncol. 2023;24(3):e133-e143. DOI:https://doi.org/10.1016/S1470-2045(22)00742-2

17.

, et al. Non-invasive multimodal CT deep learning biomarker to predict pathological complete response of non-small cell lung cancer following neoadjuvant immunochemotherapy: A multicenter study. J Immunother Cancer. 2024;12(9):e009348-e009348. DOI:https://doi.org/10.1136/jitc-2024-009348

18.

Saad

Hong

Aminu

, et al. Predicting benefit from immune checkpoint inhibitors in patients with non-small-cell lung cancer by CT-based ensemble deep learning: A retrospective study. Lancet Digit Health. 2023;5(7):e404-e420. DOI:https://doi.org/10.1016/S2589-7500(23)00082-1

19.

Tian

, et al. Assessing PD-L1 expression in non-small cell lung cancer and predicting responses to immune checkpoint inhibitors using deep learning on computed tomography images. Theranostics. 2021;11(5):2098-2107. DOI:https://doi.org/10.7150/thno.48027

20.

Kwint

Stam

Proust-Lima

, et al. The prognostic value of volumetric changes of the primary tumor measured on Cone Beam-CT during radiotherapy for concurrent chemoradiation in NSCLC patients. Radiother Oncol. 2020;146(0):44-51. DOI:https://doi.org/10.1016/j.radonc.2020.02.002

21.

Hosny

Zeleznik

, et al. Deep learning predicts lung cancer treatment response from serial medical imaging. Clin Cancer Res. 2019;25(11):3266-3275. DOI:https://doi.org/10.1158/1078-0432.CCR-18-2495

22.

Kim

Hwang

Kim

, et al. Predicting treatment outcomes using 18F-FDG PET biomarkers in patients with non-small-cell lung cancer receiving chemoimmunotherapy. Ther Adv Med Oncol. 2022;14(0):17588359211068732. DOI:https://doi.org/10.1177/17588359211068732

23.

Lababede

Meziane

. The eighth edition of TNM staging of lung cancer: Reference chart and diagrams. Oncologist. 2018;23(7):844-848. DOI:https://doi.org/10.1634/theoncologist.2017-0659

24.

Ettinger

Wood

Aisner

, et al. NCCN Guidelines® Insights: Non–small cell lung cancer, version 2.2023: Featured updates to the NCCN guidelines. J Natl Compr Cancer Network. 2023;21(4):340-350. DOI:https://doi.org/10.6004/jnccn.2023.0020

25.

Gao

, et al. National guidelines for diagnosis and treatment of lung cancer 2022 in China (English version) National Health Commission of the People's Republic of China. Chin J Cancer Res. 2022;34(3):176-206. DOI:https://doi.org/10.21147/j.issn.1000-9604.2022.03.03

26.

Zhao

Zhang

, et al. P2. 04-29 preliminary results with tislelizumab in Chinese patients with non-small cell lung cancer (NSCLC). J Thorac Oncol. 2018;13(10):S741-S742. DOI:https://doi.org/10.1016/j.jtho.2018.08.1253

27.

Chen

Yang

, et al. Radiotherapy improves the outcomes of immunotherapy with Sintilimab in non-small-cell lung cancer: A real-world analysis. Front Immunol. 2022;13(0):991431. DOI:https://doi.org/10.3389/fimmu.2022.991431

28.

Seymour

Bogaerts

Perrone

, et al. iRECIST: Guidelines for response criteria for use in trials testing immunotherapeutics. Lancet Oncol. 2017;18(3):e143-e152. DOI:https://doi.org/10.1016/S1470-2045(17)30074-8

29.

Grégoire

Mackie

. State of the art on dose prescription, reporting and recording in Intensity-Modulated Radiation Therapy (ICRU report No. 83). Cancer/Radiothérapie. 2011;15(6-7):555-559. DOI:https://doi.org/10.1016/j.canrad.2011.04.003

30.

Gary

Johannes

Douglas

Karel

GMM

. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ. 2015;350(jan07 4):g7594-g7594. DOI:https://doi.org/10.1136/bmj.g7594

31.

Mosmann

Borba

Macedo

, et al. Solitary pulmonary nodule and 18 F-FDG PET/CT. Part 1: Epidemiology, morphological evaluation and cancer probability. Radiol Bras. 2016;49(1):35-42. DOI:https://doi.org/10.1590/0100-3984.2014.0012

32.

Bradley

Komaki

, et al. Long-term results of NRG oncology RTOG 0617: Standard-versus high-dose chemoradiotherapy with or without cetuximab for unresectable stage III non–small-cell lung cancer. J Clin Oncol. 2020;38(7):706-714. DOI:https://doi.org/10.1200/JCO.19.01162

33.

Zhu

Zhang

Zheng

, et al. Combination strategies to maximize the benefits of cancer immunotherapy. J Hematol Oncol. 2021;14(1):156. DOI:https://doi.org/10.1186/s13045-021-01164-5

34.

Wang

Qiao

Jiang

, et al. The landscape of immune checkpoint inhibitor plus chemotherapy versus immunotherapy for advanced non-small-cell lung cancer: A systematic review and meta-analysis. J Cell Physiol. 2020;235(5):4913-4927. DOI:https://doi.org/10.1002/jcp.29371

35.

Casares

Pequignot

Tesniere

, et al. Caspase-dependent immunogenicity of doxorubicin-induced tumor cell death. J Exp Med. 2005;202(12):1691-1701. DOI:https://doi.org/10.1084/jem.20050915

36.

Wen

Everitt

Bedő

, et al. Spleen volume variation in patients with locally advanced non-small cell lung cancer receiving platinum-based chemo-radiotherapy. PloS One. 2015;10(11):e0142608. DOI:https://doi.org/10.1371/journal.pone.0142608

37.

Zhu

Wang

Tang

Cao

. Radiotherapy induced immunogenic cell death by remodeling tumor immune microenvironment. Front Immunol. 2022;13(0):1074477. DOI:https://doi.org/10.3389/fimmu.2022.1074477

38.

Tian

Goldstein

Wang

, et al. Mutual regulation of tumour vessel normalization and immunostimulatory reprogramming. Nature. 2017;544(7649):250-254. DOI:https://doi.org/10.1038/nature21724

39.

Zheng

Fang

Liu

, et al. Increased vessel perfusion predicts the efficacy of immune checkpoint blockade. J Clin Invest. 2018;128(5):2104-2115. DOI:https://doi.org/10.1172/JCI96582

40.

Gillies

Kinahan

Hricak

. Radiomics: Images are more than pictures, they are data. Radiology. 2016;278(2):563-577. DOI:https://doi.org/10.1148/radiol.2015151169

41.

Thawani

Mustafa

. The future of radiomics in lung cancer. Lancet Digit Health. 2020;2(3):e103. DOI:https://doi.org/10.1016/S2589-7500(20)30022-4

42.

Leijenaar

Zerka

Vaidyanathan

, et al. 5MO Prospective validation of a radiomics signature for chemoradiotherapy lung cancer patients. Ann Oncol. 2020;31(0):S246. DOI:https://doi.org/10.1016/j.annonc.2020.08.158

43.

Theelen

WSME

Chen

Verma

, et al. Pembrolizumab with or without radiotherapy for metastatic non-small-cell lung cancer: A pooled analysis of two randomised trials. Lancet Respir Med. 2021;9(5):467-475. DOI:https://doi.org/10.1016/S2213-2600(20)30391-X

44.

Heidt

Bohn

Stollmayer

, et al. Delta-radiomics features of ADC maps as early predictors of treatment response in lung cancer. Insights Imaging. 2024;15(1):218. DOI:https://doi.org/10.1186/s13244-024-01787-5

45.

Provenzano

Favali

Mazzeo

, et al. 178P Explainable radiomics, machine and deep learning models to predict immune-checkpoint inhibitor treatment efficacy in advanced non-small cell lung cancer patients. Ann Oncol. 2023;34(0):S252. DOI:https://doi.org/10.1016/j.annonc.2023.09.2902

46.

Xie

Chen

. Comparison of the efficacy and safety of domestically produced tislelizumab, camrelizumab, and imported pembrolizumab in the treatment of advanced NSCLC: A real-world retrospective study. Clin Transl Oncol. 2025;27(1):204-212. DOI:https://doi.org/10.1007/s12094-024-03565-7

47.

Daei Sorkhabi

ZareDini

Fazlollahi

, et al. The safety and efficacy of tislelizumab, alone or in combination with chemotherapy, for the treatment of non-small cell lung cancer: A systematic review of clinical trials. BMC Pulm Med. 2023;23(1):495. DOI:https://doi.org/10.1186/s12890-023-02755-3

48.

Zhang

Lin

Tan

, et al. Sintilimab for the treatment of non-small cell lung cancer. Biomark Res. 2022;10(1):23. DOI:https://doi.org/10.1186/s40364-022-00363-7

49.

Zhou

Zhang

Guo

, et al. Five-year follow-up of neoadjuvant PD-1 inhibitor (sintilimab) in non-small cell lung cancer. J Immunother Cancer. 2024;12(8):e009355. DOI:https://doi.org/10.1136/jitc-2024-009355

50.

Tang

Cong

Zheng

, et al. Concurrent sintilimab with sequential chemoradiotherapy for unresectable, stage III non-small cell lung cancer: A retrospective study. Front Oncol. 2023;13(0):1129989. DOI:https://doi.org/10.3389/fonc.2023.1129989

51.

Yang

Bai

, et al. Efficacy and safety of pembrolizumab versus sintilimab treatment in patients with advanced squamous lung cancer: A real-world study in China. Front Oncol. 2023;13(0):1147903. DOI:https://doi.org/10.3389/fonc.2023.1147903

52.

Litvin

Burkin

Kropinov

Paramzin

. Radiomics and digital image texture analysis in oncology. Sovremennye Tehnologii v Medicine. 2021;13(2):97-104. DOI:https://doi.org/10.17691/stm2021.13.2.11

53.

Lujiao

Lulin

Hongxia

, et al. Prediction of CD3 T cells and CD8 T cells expression levels in non-small cell lung cancer based on radiomic features of CT images. Front Oncol. 2023;13(0):1-1. DOI:https://doi.org/10.3389/fonc.2023.1104316

54.

Yasuhiro

Taku

Tomoya

. Motility dynamics of T cells in tumor-draining lymph nodes: A rational indicator of antitumor response and immune checkpoint blockade. Cancers (Basel). 2021;13(8):4616-4616. DOI:https://doi.org/10.3390/cancers13184616

55.

Karan

Venu

Teresa

. Key chemokines direct migration of immune cells in solid tumors. Cancer Gene Ther. 2022;29(1):10-21. DOI:https://doi.org/10.1038/s41417-021-00303-x

56.

Katelyn

Courtney

Rosemarie

, et al. Neoadjuvant selicrelumab, an agonist CD40 antibody, induces changes in the tumor microenvironment in patients with resectable pancreatic cancer. Clin Cancer Res. 2021;27(16):4574-4586. DOI:https://doi.org/10.1158/1078-0432.ccr-21-1047

57.

Jan

Daniel

Michael

, et al. Tumour mutational burden: Clinical utility, challenges and emerging improvements. Nat Rev Clin Oncol. 2024;21(10):725-742. DOI:https://doi.org/10.1038/s41571-024-00932-9

58.

Gautier

Anthony

Thomas

, et al. Verification and validation of computational models used in biopharmaceutical manufacturing: Potential application of the ASME verification and validation 40 standard and FDA proposed AI/ML model life cycle management framework. J Pharm Sci. 2021;110(4):1540-1544. DOI:https://doi.org/10.1016/j.xphs.2021.01.016

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.01 MB