Baseline Radiomics as a Prognostic Tool for Clinical Benefit from Immune Checkpoint Inhibition in Inoperable NSCLC Without Activating Mutations

Academic Editor: Avraham Eisbruch Received: 9 April 2025 Revised: 16 May 2025 Accepted: 26 May 2025 Published: 27 May 2025 Citation: Moiseenko, F.; Radulovic, M.; Tsvetkova, N.; Chernobrivceva, V.; Gabina, A.; Oganesian, A.; Makarkina, M.; Elsakova, E.; Krasavina, M.; Barsova, D.; et al. Baseline Radiomics as a Prognostic Tool for Clinical Benefit from Immune Checkpoint Inhibition in Inoperable NSCLC Without Activating Mutations. Cancers 2025,17, 1790. https://doi.org/10.3390/ cancers17111790 Copyright: © 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/ licenses/by/4.0/). Article Baseline Radiomics as a Prognostic Tool for Clinical Benefit from Immune Checkpoint Inhibition in Inoperable NSCLC Without Activating Mutations Fedor Moiseenko 1,2 , Marko Radulovic 3,* , Nadezhda Tsvetkova 1, Vera Chernobrivceva 1, Albina Gabina 1,2 , Any Oganesian 1,4, Maria Makarkina 1, Ekaterina Elsakova 1, Maria Krasavina 1, Daria Barsova 1, Elizaveta Artemeva 1 , Valeria Khenshtein 1 , Natalia Levchenko 1 , Viacheslav Chubenko 1 , Vitaliy Egorenkov 1 , Nikita Volkov 1, Alexei Bogdanov 1and Vladimir Moiseyenko 1 1N.P Napalkov Saint Petersburg Clinical Research and Practical Centre for Specialized Types of Medical Care (Oncological), Leningradskaya Str. 68A, 197758 Saint Petersburg, Russia; [email protected] (F.M.); [email protected] (N.T.); [email protected] (V.C.); [email protected] (A.G.); [email protected] (A.O.); [email protected]u (M.M.); [email protected] (E.E.); [email protected] (M.K.); [email protected] (D.B.); [email protected] (E.A.); [email protected] (V.K.); [email protected] (N.L.); [email protected] (V.C.); v[email protected]u (V.E.); [email protected] (N.V.); [email protected] (A.B.); [email protected] (V.M.) 2N.N. Petrov National Medical Research Center of Oncology, Ministry of Public Health of the Russian Federation, Leningradskaya Str. 68, 197758 Saint Petersburg, Russia 3 Department of Experimental Oncology, Institute of Oncology and Radiology of Serbia, 11000 Belgrade, Serbia 4Department of Oncology, Pediatric Oncology and Radiation Therapy, St.-Petersburg State Pediatric Medical University, St. Lithuanian 2, 194100 Saint Petersburg, Russia *Correspondence: [email protected] Simple Summary: This study introduces a powerful machine learning-based radiomics approach to help improve predictions of immunotherapy outcomes in patients with nonsmall cell lung cancer (NSCLC). We believed that the full potential of CT scan-based tumor analysis had not been achieved, partly due to limited use of model integrations (ensembles) in previous research. To address this, we tested 1680 combinations of data processing and machine learning methods, selecting the best-performing ones to create an integrated (ensemble) model. Using clinical and imaging data, our final model achieved an AUC of 0.86 for predicting 24-month patient survival, which, to our knowledge, exceeds previously published results for this diagnosis and disease outcomes. This approach reduces the weaknesses of relying on a single model and offers a more reliable and accurate tool for predicting immunotherapy outcomes. Abstract: Background/Objectives: Checkpoint inhibitors (ICIs) are key therapies for NSCLC, but current selection criteria, such as excluding mutation carriers and assessing PD-L1, lack sensitivity. As a result, many patients receive costly treatments with limited benefit. Therefore, this study aimed to predict which NSCLC patients would achieve durable survival ( ≥ 24 months) with immunotherapy. Methods: A comprehensive ensemble radiomics approach was applied to pretreatment CT scans to prognosticate overall survival (OS) and predict progression-free survival (PFS) in a cohort of 220 consecutive patients with inoperable NSCLC treated with first-line ICIs (pembrolizumab or atezolizumab, nivolumab or prolgolimab) as monotherapy or in combination. The radiomics pipeline evaluated four normalization methods (none, min-max, Z-score, mean), four feature selection techniques (ANOVA, RFE, Kruskal–Wallis, Relief), and ten classifiers (e.g., SVM, random forest). Using two to eight radiomics features, 1680 models were built in the Feature Explorer (FAE) Python package. Results: Three feature sets were evaluated: clinicopathological (CP) only, radiomics only, and a combined set, using 6and 12-month Cancers 2025,17, 1790 https://doi.org/10.3390/cancers17111790 Cancers 2025,17, 1790 2 of 16 PFS and 24-month OS endpoints. The top 15 models were ensembled by averaging their probability scores. The best performance was achieved at 24-month OS with the combined CP and radiomics ensemble (AUC = 0.863, accuracy = 85%), followed by radiomics-only (AUC = 0.796, accuracy = 82% ) and CP-only (AUC = 0.671, accuracy = 76%). Predictive performance was lower for 6-month (AUC = 0.719) and 12-month PFS (AUC = 0.739) endpoints. Conclusions: Our radiomics pipeline improved selection of NSCLC patients for immunotherapy and could spare non-responders unnecessary toxicity while enhancing cost-effectiveness. Keywords: NSCLC; prediction; prognosis; radiomics; machine learning; ensemble; immunotherapy; checkpoint inhibitors 1. Introduction Non-small cell lung cancer (NSCLC) is a common, deadly, and highly heterogeneous tumor type. This heterogeneity is evident in its clinical course, histological characteristics, genetic and expression profiles, and in responses to treatment. Different treatment modalities can lead to diverse outcomes; for example, some therapies provide immediate tumor shrinkage and symptom improvement, while others deliver a survival advantage over the long term [ 1 ]. The combination of these benefits forms the basis for treatment decisions in clinical practice. Historically, chemotherapy was the main NSCLC treatment, often resulting in significant tumor shrinkage, though typically for only a short duration. Targeted therapies, on the other hand, have shown the potential for longer-lasting responses but are limited to a small group of patients with specific molecular aberrations. More recently, immunotherapy, particularly checkpoint inhibitors, has emerged as a transformative treatment, offering remarkably long survival benefits for a subset of patients [2]. The selection of patients who are likely to benefit from immunotherapy is based on several factors, including the carcinoma’s origin (with smoking history playing a role), the molecular genetic profile (such as mutations associated with sensitivity or resistance and tumor mutational burden), and signs of immune system activation. Among these factors, PD-L1 expression is currently the primary criterion for estimating both the likelihood of an overall response and a long-lasting response. PD-L1 also serves as a marker of the mechanism by which tumor cells protect themselves from activated immune cells, highlighting the complex interactions among the tumor, its microenvironment, and the host organism. However, PD-L1 status has notable limitations; for instance, approximately 20% of patients with high PD-L1 expression still experience early progression, while a proportion of PD-L1-negative patients do respond to treatment. These challenges underscore the need for more refined patient-selection strategies, with the goal of improving outcomes and personalizing treatments in the heterogeneous landscape of NSCLC. The selection of NSCLC patients who are likely to benefit from immunotherapy is increasingly approached by use of CT radiomics, which reveals subtle intra-tumor heterogeneity that cannot be identified by visual inspection of imaging scans. Radiomics quantitatively characterizes tumor imaging data by extracting features that capture tumor size, shape, voxel intensity distribution, spatial relationships, and texture patterns within the tumor VOI [ 3 ]. These features are then entered into machine learning models to classify patients according to their clinical outcomes and therapy response [4]. Current state-of-the-art approaches for classifying therapy response and disease outcomes in NSCLC still leave considerable room for improvement [ 5 ]. Radiomics still achieves Cancers 2025,17, 1790 3 of 16 only moderate predictive power on held-out test sets, with reported AUCs rarely exceeding 0.75. Consequently, radiomics only has potential to become a supportive biomarker rather than a stand-alone decision tool. Progress is limited by the absence of optimized, standardized, and widely adopted pipelines for feature selection and classification only sporadic use of ensemble techniques and limited reproducibility due to rare sharing of code, hyperparameters, and imaging protocols. Until sequential internal validation and external validation become routine, radiomics will remain confined to proof-of-concept research instead of becoming a reliable clinical tool [ 6 – 13 ]. To address these gaps, we aimed to improve the consistency and predictive accuracy of radiomics-based classification by employing an ensemble approach that integrates the best-performing individual models. Such an ensemble radiomics framework mitigates the limitations of single-model variability and improves overall robustness and generalizability, providing a foundation for more reproducible and accurate outcome prediction in future studies. The novelty of this study is based on systematic evaluation of 1680 standardized pipelines combining normalization, preprocessing, feature selection, and classification steps. It also introduces an ensemble framework that integrates top-performing models, to reduce reliance on any single model’s chance success and enhance predictive stability. Importantly, ensembling was performed by soft voting, averaging continuous probability scores of individual models to avoid the bias of hard voting in selecting probability score thresholds. Several previous ensemble studies in NSCLC used XGBoost classifier alone [ 14 , 15 ], whereas we used a similar AdaBoost classifier alongside nine others. One study ensembled five classifiers using hard voting using a fixed 0.5 probability score threshold for each individual model [ 16 ]. The most comprehensive prior approach combined RF, SVM, and LASSO to build 54 predictive models [ 8 ]. Notably, many studies use the term “ensemble” to describe the integration of multiple feature types rather than distinct radiomic workflows [ 8 , 15 , 17 ]. Our current study thus presents a substantial methodological advance by addressing workflow optimization and standardization through an exhaustive pipeline, supported by openly shared code for reproducibility and wider adoption. Motivated by the clinical need for more reliable prognostic tools in immunotherapytreated NSCLC patients, we aimed to implement a comprehensive workflow to systematically explore 1680 combinations of normalization, preprocessing, feature selection, and classification strategies on pretreatment CT scans. The top-performing models were then integrated into a unified ensemble radiomics signature, aiming to enhance generalizability, reduce variance due to chance performance of individual models, and support more robust prognostic assessments. 2. Materials and Methods 2.1. Patient Group We used a retrospective database of patients treated for inoperable NSCLC at the N.P Napalkov Saint Petersburg Clinical Research and Practical Centre for Specialized Types of Medical Care (Oncological) in 2021. This study was approved by the institutional Ethics Committee (Approval No. 3, dated 14 March 2023). Patients received checkpoint inhibitors (pembrolizumab and bioanalogues, atezolizumab, nivolumab, or prolgolimab) as first-line palliative therapy, either as monotherapy or in combination. Only patients with available pre-immunotherapy images were included. Outcomes were defined as 6or 12-month progression-free survival (PFS) and 24-month overall survival (OS). Clinical characteristics are presented in Table 1. All the patients were treated according to national guidelines and MDT decisions, receiving a checkpoint inhibitor either alone or with chemotherapy as part of routine care. Treatment response was evaluated with CT every 6–8 weeks, adverse events were monitored per local practice, and overall survival Cancers 2025,17, 1790 4 of 16 was determined via patient phone contact or the national health database. Progression of disease was defined according to iRECIST criteria. Progression of disease was considered as a time from day one of first cycle to date of first registered progression, provided that this progression was subsequently confirmed by a second investigation. Table 1. Clinical characteristics of patients included in the study a. All Included (n = 220) OS ≥24 Months (n = 52) OS < 24 Months (n = 168) Sex Male 172 (78.5%) 38 (73.1%) 135 (80.4%) Female 48 (21.5%) 14 (26.9%) 33 (19.6%) Mean age (min-max) 63.3 (35–87) 64.6 (38–82) 62.9 (35–87) Smoking Yes 102 (46.4%) 31 (59.6%) 71 (42.3%) No 118 (53.6%) 21 (40.4%) 97 (57.7%) NLR ≥3109 (49.5%) 25 (48.1%) 84 (50%) Histology Adenocarcinoma 109 (49.5%) 29 (55.8%) 80 (47.6%) Squamous cell cancer 99 (45.0%) 21 (40.4%) 78 (46.4%) Large cell 12 (5.5%) 2 (3.8%) 10 (6%) Stage IIIA 33 (15.0%) 9 (17.3%) 24 (14.3%) IIIB 28 (12.7%) 6 (11.5%) 22 (13.1%) IV 159 (72.3%) 37 (71.2%) 122 (72.6%) Brain metastases Present 16 (7.3%) 6 (11.5%) 10 (6.0%) Liver metastases Present 17 (7.7%) 2 (3.8%) 14 (8.4%) a Tumor mutational burden (TMB) and circulating tumor DNA (ctDNA) biomarkers were not included due to incomplete availability across the cohort. Abbreviations: NLR, Neutrophil-to-Lymphocyte Ratio. The prospective sample size calculation was based on a pilot study of the first 120 chronologically included patients, which indicated that a minimum of 96 patients, including 12 positive cases, would be required. The calculations were based on an alpha value of 0.05, a beta value of 0.20, a positive-to-negative case ratio of 0.14, and an expected effect size corresponding to an AUC of 0.75. The final study included 220 patients, with at least 147, 75, and 37 survivors (positive cases) observed in the 6-, 12-, and 24-month groups, respectively. All actual parameters exceeded the initial sample size estimates, with positive-to-negative ratios of 0.67, 0.34, and 0.17, respectively, and an achieved highest prognostic AUC of 0.86. 2.2. Image Acquisition CT chest images from 220 patients were acquired in uncompressed DICOM format (slice thickness ≤2.5 mm) using a Siemens Somatom Definition 128 CT scanner (Siemens, Munich, Germany) at the N.P. Napalkov Saint Petersburg Clinical Research and Practical Centre for Specialized Types of Medical Care (Oncological). The dataset excluded low-dose protocols, while the use of contrast was not a limiting factor. During database creation, the radiologist was responsible for data selection, interpretation, and primary reporting, image labeling, and data deidentification. CT scanning and preliminary analysis were performed using a RadiAnt DICOM Viewer v2025.1. Data processing involved preparing the CT datasets, image pre-processing, detecting pathological lesions, and segmentation. Contouring of pathological lung lesions was performed in 3D Slicer v.5.8.1, an open-source medical image analysis platform, using multiplanar reconstruction (MPR) of chest CT Cancers 2025,17, 1790 5 of 16 images. This process was guided by initial radiology reports to ensure accuracy and consistency. Tumor VOIs were semi-automatically segmented slice-by-slice in the axial plane using the “Grow from Seeds” tool (FastGrowCut) in 3D Slicer’s Segment Editor by a radiation oncologist (F.M.) with 15 years of oncologic imaging experience. Threshold-based selection, paint, erase, and island removal tools were used to refine segmentations. The final tumor VOIs were reviewed for consistency and exported in MRB format for further analysis. Figure 1illustrates the final segmentation. Figure 1. Flowchart of CT-based prognostic model construction for prediction of lung cancer PFS or prognostication of OS. CT scans were acquired after diagnosis, and tumor VOIs were segmented. Prior to radiomics analysis, images underwent normalization, resampling to 1 × 1 × 1 mm, and interpolation. A total of 2157 radiomics features were extracted using PyRadiomics. The CP and radiomics features were combined by adding them to a CSV file along with the outcome column. The cohort was further divided into development and test sets in a 70:30 ratio. Feature values were normalized using z-score, mean normalization (to − 0.5, 0.5), or min-max normalization (to 0, 1). Subsequently, data balancing, preprocessing, feature selection, grid search for optimal hyperparameters, and classification were performed. Feature selection and classification were conducted by 10-fold cross-validation in the development set. The final prognostic model was constructed by combining the 15 best-performing models into ensembles using soft voting. 2.3. Feature Extraction For the radiomics analysis, we used the open-source Python package Pyradiomics v3.1.0, which is compliant with the Imaging Biomarker Standardization Initiative (IBSI) standards. [ 18 ]. Using the parameter file provided below, the software was configured to generate all available image transformations and feature types, resulting in a total of 2157 features computed per CT scan. Seven image transformations were applied wavelet, square, square root, logarithm, gradient, exponential, and Laplacian of Gaussian (LoG). To reduce inter-scan variability, the CT scans were first z-score normalized (mean intensity = 0, standard deviation = 1) and then resampled to an isotropic voxel size of 1×1×1 mm using ‘sitkBSpline.’ Radiomic features were extracted only from tumor VOIs, yielding 107 standard shape, intensity (first-order), and texture (second-order) fea- Cancers 2025,17, 1790 6 of 16 tures from the original images, while higher-order features were computed from the transformed images. The bin width was individually determined for each filter type in a pilot analysis of 75 CTs to keep the number of gray-level bins per scan between 30 and 130, which is considered optimal for textural reproducibility without causing over-smoothing or excessive noise sensitivity [ 19 ]. This was necessary because the seven image filter transformations produced images with distinct pixel intensity ranges, requiring separate bin-width settings for each filter to maintain a consistent bin count across all individual scans. For detailed descriptions of the extracted radiomic features, please refer to https://pyradiomics.readthedocs.io/en/latest/features.html (accessed on 25 May 2025). Below is the Params.yaml file: setting: normalize: true normalizeScale: 600 resampledPixelSpacing: [1, 1, 1] interpolator: 'sitkBSpline' voxelArrayShift: 1000 binWidth: 30.0 label: 2 imageType: Original: LoG: sigma: [1.0, 2.0, 3.0, 4.0, 5.0] binWidth: 15.0 Wavelet: binWidth: 8.0 Square: binWidth: 15 SquareRoot: binWidth: 25 Logarithm: binWidth: 50 Exponential: binWidth: 6 Gradient: binWidth: 14 featureClass: glcm: firstorder: shape2D: shape: glrlm: glszm: gldm: ngtdm: Cancers 2025,17, 1790 7 of 16 2.4. Model Selection The data were partitioned based on the chronological order of patient inclusion into training (n = 110), validation (n = 44), and test cohorts (n = 66) at the ratio of 5:2:3, for model training, validation, and independent evaluation, respectively. The endpoints were defined as binary outcomes: 6-month PFS, 12-month PFS, and 24-month OS. The supervised machine learning modeling was performed using the FeAture Explorer FAEv0.6.0.7z python package with NumPy, pandas, and scikit-learn [ 20 ]. The source code is openly available on GitHub (https://github.com/salan668/FAE.git (accessed on 25 May 2025)). The machine learning pipeline began with CSV files containing a binary outcome column and either CP-only, radiomics-only, or CP and radiomics feature columns. All features underwent identical normalization, feature selection, and classification procedures. Data balancing was performed through upsampling. At every step, preprocessing, feature selection, and classification, only one method was applied per each step, rather than combining multiple methods simultaneously. For example, feature pre-selection involved discarding features with Pearson’s correlation coefficient above 0.97 or, alternatively, applying principal component analysis. Feature selection was then performed using one of the following methods: ANOVA, Kruskal–Wallis (KW), Recursive Feature Elimination (RFE), or Relief. The remaining highly relevant features were used as input for one of these classifiers: support vector machine (SVM), linear discriminant analysis (LDA), logistic regression (LR), AdaBoost, Gaussian process (GP), multilayer perceptron (MLP), random forest (RF), least absolute shrinkage and selection operator (LASSO), decision tree (DT), or naïve Bayes (NB). The script was set to select between two and eight features. All possible combinations of the above components resulted in 1680 radiomics pipelines, trained on the development set and evaluated on the test dataset. Parameters such as slope, intercept, weight coefficients, and support vectors are learned from the data during training, while hyperparameters are not derived from the data. Instead, hyperparameters are tuned via grid-search based on the model’s performance on validation sets during cross-validation. 2.5. Ensemble Modeling The probability scores delivered by each model were standardized to a mean of 0 and a standard deviation of 1 using z-score normalization. The top 15 prognostic models in the test set were then integrated, averaging their continuous normalized scores to generate an overall ensemble probability. These ensembles were subsequently also evaluated on the reserved test set. We tested varying numbers of models in the ensemble during preliminary experiments and found that including more than approximately 15 models no longer improved prognostic performance. “Soft” voting, which aggregates continuous probability scores, was used instead of hard voting binarized class labels to avoid biases introduced by early categorization. 2.6. Statistical Analysis For each combination of features and endpoints, binary classification performance was evaluated on the reserved test set using the FAE Python package by calculating a comprehensive set of performance metrics, ROC analysis, F1 score, and the Youden index, to assess overall discrimination ability, while accuracy and balanced accuracy measure general correctness, with the latter accounting for class imbalance. True positives and true negatives were used as basic classification counts, while Matthews correlation coefficient (MCC) is a robust summary measure across all confusion matrix elements, especially in imbalanced datasets. Together, these metrics offer a wide and complementary evaluation of model performance. Cancers 2025,17, 1790 8 of 16 2.7. Validation Model performance was evaluated in two stages. First, ten-fold cross-validation was performed on the development set, using an internal validation subset within each fold. Next, the best-performing models in the test set were combined, and ensembles tested again on the test set, composed sequentially of the most recent 30% of patients. To prevent information leakage, the FAE radiomics pipeline was finalized prior to evaluation on the hold-out test subset. 3. Results 3.1. Patient Characteristics We included 220 patients with inoperable NSCLC who received checkpoint inhibitors in a real-world setting (Table 1). The majority of patients were male (78.5%) and had stage IV disease (72.3%). Approximately half were smokers (46.4%) and had non-squamous histology (49.5%). The vast majority of patients received either an atezolizumab-based combination (48.3%) or pembrolizumab (36.5%; Table 2). A response was achieved in 15.6% of patients. The median progression-free survival was 8.2 months [6.8–9.5], similar to that observed in registrational trials, and the median overall survival was 22.0 months ( 19.6–24.4, Table 2) [21]. Table 2. Treatment characteristics and results. All Included (n = 220) OS ≥24 Months (n = 52) OS < 24 Months (n = 168) Drug Atezolizumab 102 (48.3%) 19 (37.3%) 83 (51.9%) Pembrolizumab 77 (36.5%) 20 (39.2%) 57 (35.6%) Nivolumab 17 (8.1%) 3 (5.9%) 14 (8.8%) Prolgolimab 15 (7.1%) 9 (17.6%) 6 (3.8%) Best response CR 3 (1.4%) 0 (0%) 3 (1.8%) PR 26 (11.8%) 12 (23.1%) 14 (8.3%) SD 47 (21.4%) 9 (17.3%) 38 (22.6%) PD 110 (50.0%) 27 (51.9%) 83 (49.4%) Not assessed 34 (15.4%) 4 (7.7%) 30 (17.9%) RR 29 (15.6%) 12 (25%) 17 (12.3%) DCR 76 (40.8%) 21 (43.7%) 55 (39.9%) Median PFS (months) [95% CI] 8.2 [6.8–9.5] 15.4 [11.9–18.9] 6.97 [6.2–7.8] Median OS (months) [95% CI] 22.0 [19.6–24.4] 33.2 [30.7–35.7] 14.5 [12.3–16.7] Abbreviations: CR, Complete Response; PR, Partial Response; SD Stable Disease; PD, Progressive Disease; RR, Response Rate; DCR; Disease Control Rate. 3.2. Experimental Design Figure 1outlines the study workflow, which included CT scans from 220 patients. Briefly, patient and imaging data were curated, and 1680 CT-based radiomics models were trained, validated, and tested. These models were generated by combining three normalization methods, two dimensionality reduction techniques, four feature selection methods, and ten classifiers, with the number of selected features restricted to between two and eight (Figure 1). The split into development and test sets was based on the chronological order of patient inclusion, as detailed in the Methods section. This approach simulated a real-world clinical scenario in which retrospective data are used to predict outcomes for future patients from the same institution. Cancers 2025,17, 1790 9 of 16 3.3. Performance of Individual Models The best-performing individual models in the test set were then combined into an ensemble and evaluated in the reserved test set for stratification of PFS and OS following immune checkpoint inhibitor therapy. Models were developed for 6-, 12-, and 24-month endpoints, and the evaluation metrics included AUC, accuracy, balanced accuracy, true positives, true negatives, MCC, F1 score and the Youden index. Among the top 15 individual models for the 24-month endpoint that included CP and radiomics features, the most frequent normalization methods were mean (40%) and min-max (33%). The most frequent preprocessing methods were PCC (53%) and PCA (47%). The most frequent feature selectors were Relief (60%) and Kruskal–Wallis (20%), while the most frequent classifiers were AdaBoost (60%) and Auto-Encoder (12%). The risk of insufficient diversification due to AdaBoost dominance is unlikely, because the three upstream optimization layers introduced additional variability in the selection of features entering AdaBoost. Figure 2displays the classification evaluation of the individual models that best prognosticated 24-month overall survival in the test set. Applying a 24-month survival endpoint allowed us to define a patient subgroup whose PFS exceeded double that of the remaining cohort, highlighting those who derived maximal benefit from immunotherapy. As expected, individual models exhibited much higher prognostic performance on the training set compared to the validation and test sets (Figure 2). This performance gap reflects the degree of overfitting and serves as an indicator of each model’s generalizability. Figure 2. The classification performance of the best-performing individual models in the test set for the 24-month endpoint. The models were based on CP features alone, radiomics features alone, or a combination of CP and radiomics features. Panels (a–c) show the prognostic evaluation of the best-performing model in the test set for each feature combination. We addressed potential temporal confounding by comparing prognostic performance in the chronologically selected test set with that in randomly selected validation subsets. This comparison provided a direct way to assess time-related bias. Among the 15 top-performing models for the 24-month endpoint that combined CP and radiomics features, the average AUC in the randomly selected validation subsets was 0.74 (SD 0.12, 95% CI 0.67 to 0.82), while, in the chronologically selected test set, the average AUC was 0.64 (SD 0.03, 95% CI 0.62 to 0.66). These AUC values refer to averages of individual models, not to the ensemble models. The AUC values obtained in the validation and test sets were significantly different, with t-statistic of 2.81 and p= 0.015 by an independent two- Cancers 2025,17, 1790 16 of 16 35. Velazquez, E.R.; Parmar, C.; Jermoumi, M.; Mak, R.H.; van Baardwijk, A.; Fennessy, F.M.; Lewis, J.H.; De Ruysscher, D.; Kikinis, R.; Lambin, P.; et al. Volumetric CT-based segmentation of NSCLC using 3D-Slicer. Sci. Rep. 2013,3, 3529. [CrossRef] 36. Cama, I.; Guzmán, A.; Campi, C.; Piana, M.; Lekadir, K.; Garbarino, S.; Díaz, O. Segmentation variability and radiomics stability for predicting Triple-Negative Breast Cancer subtype using Magnetic Resonance Imaging. arXiv 2025, arXiv:2504.01692. [CrossRef] Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.