Skip to main content
NCBI home page
Journal for Immunotherapy of Cancer logoLink to Journal for Immunotherapy of Cancer
. 2025 Oct 5;13(10):e013074. doi: 10.1136/jitc-2025-013074

Causal AI-based clinical and radiomic analysis for optimizing patient selection in combined immunotherapy and SABR in early-stage NSCLC: a secondary analysis of the phase II I-SABR trial

Maliazurina B Saad 1,0, Eman Showkatian 1,0, Vivek Verma 2, Qasem Al-Tashi 1, Muhammad Aminu 1, Xinyan Xu 1,2, Muhamed Qayati Mohamed 1,§, Morteza Salehjahromi 1, Sheeba J Sujit 1, Yuliya Kitsel 3, Steven H Lin 2, Zhongxing Liao 2, Saumil Gandhi 2, David Qian 2, David Jaffray 1,4, Caroline Chung 4,5, Natalie I Vokes 3,6, Jianjun Zhang 3,6, J Jack Lee 7,0, John V Heymach 3,0, Jia Wu 1,3,4,*,0, Joe Y Chang 2,✉,0
PMCID: PMC12506444  PMID: 41052882

Abstract

Background

The recent phase II randomized stereotactic ablative radiotherapy with and without immunotherapy (I-SABR) trial has shown improved event-free survival (EFS) when adding immunotherapy to stereotactic ablative radiotherapy (SABR) for early-stage inoperable non-small cell lung cancer (NSCLC). However, optimizing patient selection thereof is critical, because not every patient benefits from immunotherapy. Leveraging the powerful use of artificial intelligence, this secondary analysis of the I-SABR trial developed a modeling system (named “I-SABR-SELECT”) based on clinical and radiomic factors to address which patients should receive additional immunotherapy.

Methods

The discovery/validation cohorts were from the I-SABR trial, with external validation from the single-arm STARS trial. Individual treatment effect scores, estimating the benefit of adding immunotherapy, were derived from radiomic and clinical predictors using counterfactual reasoning. Dimensionality reduction was applied to mitigate overfitting and enhance model robustness. We also evaluated the average treatment effect between subgroups of patients who were treated following versus against the model’s recommendation.

Results

The model recommended that 49% (69/141) patients enrolled in the I-SABR trial switch treatments (65% (49/75) in the SABR arm and 30% (20/66) in the I-SABR arm). Patients treated by the model’s recommendation had higher EFS, with HRs of 0.06 (in the I-SABR arm, p<0.001) and 0.26 (in the SABR alone arm, p=0.0042) from the I-SABR trial population, and 0.38 (p=0.031) for the STARS trial. Following model stratification, among patients recommended for SABR+immunotherapy, the restricted mean survival time for EFS is prolonged by 1.43 years compared to those who received SABR alone. The absolute risk reduction of the added immunotherapy effect was over twofold greater than that observed in the I-SABR trial without selection.

Conclusions

Combining clinical and radiomic parameters, I-SABR-SELECT uses causal reasoning to individualize treatment selection for patients with early-stage inoperable NSCLC. If validated, it could serve as a foundation for a treatment-focused digital twin by integrating real-time adaptive decision-making.

Code: https://github.com/WuLabMDA/ISABR-SELECT

Keywords: Biomarker, Lung Cancer, Radiotherapy/radioimmunotherapy, Survivorship, Statistics

WHAT IS ALREADY KNOWN ON THIS TOPIC

  • Immune checkpoint inhibitors (ICIs) have transformed cancer care, yet only a subset of patients with metastatic disease benefit from ICI monotherapy. In our phase II randomized I-SABR trial (NCT03110978, Lancet, 2023), adding ICI to stereotactic ablative radiotherapy (SABR) for early-stage non-small cell lung cancer (NSCLC) reduced the risk of recurrence, progression, or death by 62%. However, 53% of patients treated with SABR alone remained event-free at 4 years, indicating that some may achieve durable disease control without immunotherapy. This underscores the need to tailor ICI use to maximize benefit and avoid toxicity.

  • Two recent phase III trials (KEYNOTE-867; SWOG/NRG S1914) comparing SABR alone versus SABR plus immunotherapy in early-stage NSCLC failed to meet statistical significance for recurrence-free survival, highlighting the urgency for precision selection strategies such as I-SABR-SELECT.

  • Existing biomarkers (eg, programmed cell death ligand-1, tumor mutational burden) have limited predictive value, are constrained by tissue availability and intratumoral heterogeneity, and were developed primarily in metastatic disease, leaving their utility in early-stage NSCLC uncertain.

WHAT THIS STUDY ADDS

  • We developed I-SABR-SELECT, the first predictive model leveraging artificial intelligence to integrate clinical and radiomic features for identifying patients with early-stage NSCLC most likely to benefit from adding immunotherapy to SABR.

  • In the original randomized I-SABR trial cohort, our model indicated that 49% of patients might have achieved better outcomes with an alternative treatment to the one they received.

  • Applying model-guided treatment selection more than doubled the absolute benefit of adding immunotherapy—from 23.3% to 47.0%—in the predicted benefit group, while also identifying a subgroup in whom SABR alone was the superior option.

  • Several radiomics features, including blood vessel surrounding tumor, tumor size, lung and tumor texture, as well as radiation dose and smoking status are found to be closely correlated with the benefit of adding immunotherapy.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

  • I-SABR-SELECT could serve as a tool to personalize patient with early-stage NSCLC management to optimize the treatment effects and clinical outcomes.

  • Our artificial intelligence framework may inform the design of future clinical trials combining immunotherapy with SABR to maximize the likelihood of success, and pave the way for adding immunotherapy in well-defined subgroups rather than unselected population.

Background

In efforts to continually improve outcomes for inoperable early-stage non-small cell lung cancer (NSCLC), important recent evidence has illustrated that the addition of immunotherapy to stereotactic ablative radiotherapy (SABR) may improve outcomes. In the randomized phase II I-SABR trial, combined-modality therapy (“I-SABR”)1 2 reduced the risk of recurrence, disease progression, or death by 62%.

However, it is known that SABR alone is adequate enough to cure a substantial proportion of early-stage NSCLC3 4; the I-SABR trial illustrated that 53% of patients who did not receive IO remained without evidence of recurrence at 4 years. Moreover, in the metastatic NSCLC setting, less than half of patients benefit from IO.5 Hence, it is crucial to optimize patient selection for IO, recognizing that not all patients with early-stage NSCLC would benefit. Carefully individualizing management in this population would recommend adding IO to those who may benefit and not recommend it for those who may not benefit to avoid overtreatment, unnecessary toxicities, and financial burden.

Because IO benefited patients with a variety of clinical parameters, certain biomarkers such as programmed cell death ligand-1 (PD-L1) expression,6,8 additional methodologies are imperative to implement for patient selection purposes. Radiomics is an emerging area that allows for relatively precise prediction of tumor biology and outcomes based on assessment of imaging features.9,14

However, the use of radiomics has primarily been investigated as a prognostic marker—providing information about a patient’s overall outcome regardless of treatment—rather than as a predictive tool for selecting specific therapies. Specifically, no known studies have used radiomics to quantify individualized treatment benefit and provide evidence-based guidance for treatment escalation in early-stage NSCLC. This limitation can be addressed with rapidly expanding futuristic approaches such as causal-based artificial intelligence (AI), but few data in this realm exist to date. Leveraging causal AI, this study used randomized data from the I-SABR trial to develop “I-SABR-SELECT,” a system based on clinical and radiomic factors, to identify patients with early-stage NSCLC who benefit from adding IO to SABR.

Methods

Study design and participants

This report addresses a prespecified secondary objective of the I-SABR trial to explore predictive biomarkers of treatment effect. The trial, approved by the institutional review board at The University of Texas MD Anderson Cancer Center, obtained informed consent from all patients. All procedures adhered to the 1964 Declaration of Helsinki. Clinical data were sourced from medical records, and baseline contrast CT images, taken within 3 months before treatment, were extracted from the IntelliSpace PACS system.

In this study, the term “predictive” specifically refers to the ability to predict treatment effect—a predictive biomarker modeling approach—rather than the conventional use of prediction for prognostic modeling (eg, predicting survival outcomes regardless of treatment intervention). Thus, our objective was to formulate a system that identifies patients who would benefit from combining IO with SABR, versus those for whom IO could be safely omitted. Patients from the I-SABR trial2 served as discovery and validation cohorts for the model, which was then externally validated using data from the STARS trial,4 a prospective single-arm SABR study for early-stage NSCLC. These two trials are described briefly below.

I-SABR trial cohort

The I-SABR trial was the first reported randomized phase II study that revealed a significant improvement in EFS at 4 years (the primary endpoint), which improved from 55% (SABR alone) to 77% when SABR was combined with an immune checkpoint inhibitor (I-SABR) in 141 patients with early-stage NSCLC2 (figure 1a).

Figure 1. Schematic representation of I-SABR randomized trial and development of the I-SABR-SELECT model. (a) Schematic illustration of the randomized phase II I-SABR clinical trial. (b) Steps in generating the overall “I-SABR-SELECT” framework: model construction->validation->interpretation. NSCLC, non-small cell lung cancer; SABR, stereotactic ablative radiotherapy; SHAP, SHapley Additive exPlanations.

Figure 1

Open in a new tab

Revised STARS trial cohort

To validate the results of the original randomized STARS trial (SABR vs surgery for operable stage I NSCLC4), we enrolled another 80 patients with stage I NSCLC who were treated with SABR, and compared the clinical outcomes with 80 propensity-matched patients who underwent video-assisted thoracoscopic surgery over a similar period.3 The prospective SABR arm from the revised STARS trial was used as an independent external validation cohort for this study.

I-SABR-SELECT: treatment decisions for individual patients

To better characterize variations in treatment outcomes between SABR alone or I-SABR, we integrated specific imaging characteristics with clinical risk factors as the basis for developing I-SABR-SELECT, a computational tool that encompasses several key steps (figure 1b). These clinical variables included age, sex, smoking history, tumor histology, Eastern Cooperative Oncology Group (ECOG) performance status, disease stage, tumor size, and radiation dose, which were analyzed in the previous STARS study.3 First, we harmonized CT images and extracted radiomic features to characterize different regions, including tumors, peritumoral areas, lung parenchyma, and blood vessels. Subsequently, we filtered the features to retain those that were highly robust and non-redundant. We then identified an optimal subset for model construction by considering their residual predictive value. To estimate the individualized treatment effect (ITE), we used counterfactual reasoning design, modeling the treatment difference between I-SABR and SABR. State-of-the-art machine learning technologies were used to mitigate model overfitting, including bootstrapping, repeated cross-validation, and ensemble learning. To assess the reliability and generalizability, the model was validated based on the pooled analysis of repeated cross-validation on the I-SABR cohort as well as on the STARS trial as an external validation cohort. Finally, we investigated the model’s explainability. Details of these steps are given in the subsequent paragraphs.

Step 1: processing and qualifying radiomic data

As a part of clinical protocol-mandated workup, patients received contrast-enhanced diagnostic CT imaging prior to any treatment with a slice thickness of ≤3 mm, 120 kVp, and used standard convolution kernel reconstruction. Contrast-enhanced CT scans were used instead of non-contrasted planning CTs to allow precise segmentation of regions of interest (ROIs), particularly blood vessels. To normalize CT numbers, we applied gray-level absolute discretization and voxel resampling to harmonize the CT data and reduce variability among different scans. Images were resampled to 1×1×1 mm3 voxels by using three-dimensional B-spline interpolation. Each scan was normalized based on lung window (W/L=1500 HU/−600 HU) and mediastinal window (W/L=350 HU/40 HU) settings.

We then carried out a series of fully automated pre-processing steps, including delineation of ROIs, extraction of features, and analysis of feature reproducibility and redundancy. Different ROIs were segmented by applying validated algorithms for lung tumors,15 16 blood vessels,17 and lungs,18 which were further refined by three independent expert thoracic radiologists using RayStation (RaySearch Laboratories, New York, USA).

Radiomic features (n=165; details in online supplemental tables S1−S3) were extracted with PyRadiomics (V.2.1.0, online supplemental methods S4)19 and in-house MATLAB (R2022b) codes in order to comprehensively profile individual patients by characterizing the tumor, peritumoral, and parenchymal regions (details in online supplemental figure S1). To characterize intratumoral heterogeneity, we profiled tumor shape, intensity, and texture. For peritumoral regions, we expanded the tumor segmentation mask by a radius of 5 or 10 mm to assess regional variationsc (online supplemental figure S9) and quantify tumor-mediated blood vessels. The regional variation was analyzed to quantify the phenotypic contrast between a tumor and its surrounding parenchyma.12 Tumor-mediated blood vessels were assessed to profile the vascular network that feeds a tumor, including the vessel density, volume, and counts (online supplemental methods S1). The background lung parenchyma was characterized in terms of its intensity and texture.

Next, we analyzed the quality of radiomic features to screen for stability, redundancy, and clinical relevance. A multistep feature qualification was implemented that included normalization, inter-reader agreement, correlation, and treatment modification potential. Every feature was z-scored and inter-reader agreement was estimated by using the intra-class correlation coefficient (ICC) and concordance correlation coefficient (CCC) among radiomic features extracted from two sets of ROI segmentations (online supplemental figure S2) including manual and automated contours by the “segment anything” model.20 Only radiomic features with ICC and CCC>0.85 were retained. Then, the Pearson correlation coefficient was calculated, and redundant features (p>0.90) were dropped. Finally, we assessed the predictive effect of individual features by modeling the interaction significance between individual features and the treatment indicator (online supplemental tables S1−S3).

Step 2: training and hyperparameters tuning

Dimensionality reduction via gray wolf optimization

The inherently high dimensionality of radiomic and clinical features requires reducing the features by identifying the most informative and synergistic subset. This step is crucial for efficiently modeling the intricate relationship between features and heterogeneous treatment outcomes, a task characterized by complexity and high non-linearity. We proposed a framework based on swarm intelligence to pinpoint potential effect modifiers. This framework adeptly addresses the “curse” of dimensionality of the concatenating radiomic and clinical features through applying feature selection, resampling, and bootstrapping to identify an optimal subset associated with treatment outcome. The repeated twofold cross-validation (n=15) scheme was conducted on the I-SABR trial data, with bootstrapping in the training set to ensure the robustness of the selected feature set. Thereafter, we nested a feature selection algorithm—specifically, the gray wolf optimizer21—within a wrapper framework to optimize the predictive capacity by removing strong prognostic effects through martingale residuals calculation.22 This strategy, agreed on by statistical experts through multidisciplinary discussion as the most robust solution to address sample size limitation. Further details are given in online supplemental methods S2.

Optimizing individualized treatment effects via counterfactual reasoning

Once the candidate features were established, the counterfactual reasoning approach used to model an individual’s constellation of candidate features and the expected ITE. A meta-learner, comprising two separately trained Cox proportional hazards (CPH), was used to model patient’s event-free survival (EFS: defined as local or regional recurrence, distant metastasis, secondary lung cancer and death, per-protocol2 in distinct treatment arms and estimate heterogeneous treatment effects. The ITE was then computed as a relative risk difference by contrasting the CPH models for I-SABR versus SABR (figure 1b). Mathematical formulations of ITE are given in online supplemental methods S3. For example, when a patient is treated with I-SABR, the ITE can be assessed by also modeling the risk probability under SABR-alone; subsequently, the treatment associated with lower risk is recommended.

Step 3: clinical validation

To assess the model’s predictive efficacy, we examined personalized recommendations within the I-SABR trial across two treatment arms (subgroup by treatment received), by pooling the prediction from the aforementioned repeated cross-validation. The model categorized patients into two treatment subgroups (I-SABR vs SABR), aligning with or diverging from the actual therapy they received, denoted as “follow I-SABR-SELECT” or “anti I-SABR-SELECT” (the latter being the reference group), respectively. Subsequently, we gaged clinical significance by conducting a comprehensive analysis by comparing the “follow” versus “anti I-SABR-SELECT” subgroups in terms of Kaplan-Meier analysis, log-rank tests, HRs, and restricted mean EFS time difference (RMSTD).23 The RMSTD facilitated comparisons of mean EFS times between the two stratified groups, addressing the fundamental question of how much longer of EFS, on average, those undergoing combination therapy (I-SABR) relative to those who received SABR monotherapy.

To quantify the net EFS benefit by the proposed recommendation system over the original trial, we reported effect-size metrics, including absolute risk reduction (ARR),24 the number needed to treat (NNT), and the net benefit.25 Briefly, ARR represents the proportion of patients potentially spared from unfavorable outcomes by adding IO to SABR; NNT is the average number of patients needing treatment with combination therapy to yield one patient who benefits. The net benefit represents the impact of treatment strategies by considering both event risks and treatment rates. We further investigated the model’s performance by stratifying patients based on the model’s recommendation, including subgroups of patients who were recommended to add IO or lack thereof.

Step 4: model interpretation and reasoning

Finally, we investigated the association of individual clinical/radiomic features with I-SABR-SELECT model recommendations (ie, I-SABR vs SABR). Moreover, to obtain a holistic view of feature importance in the context of the entire model, we quantified the influence by using the SHAP (SHapley Additive exPlanations) algorithm26 to account for interactions and dependencies between different features.

Statistical analysis

The primary outcome of the I-SABR trial in this secondary analysis was EFS, measured from randomization or treatment start date for I-SABR and STARS trial, respectively, to an event occurred (local, regional, or distant, secondary lung cancer, or death). Patients without events were censored at their last follow-up. Clinical data collection was locked for outcome analysis on January 18, 2023.2 Predictive value was assessed using Cox regression with interaction p values. Kaplan-Meier analysis and log-rank test were used to evaluate statistical significance of patient stratification. To adjust for multiple statistical tests, the Benjamini-Hochberg method was used to control the false discovery rate (FDR). Bootstrapping was performed to mitigate the potential bias of a small sample size. All statistical tests were two-sided, with a p value or FDR less than 0.05 considered to be statistically significant. All statistical analyses were done with R V.3.6.1.

Results

Patient characteristics in two clinical trials—I-SABR and STARS

This secondary analysis of the I-SABR trial2 (summarized in figure 1a) shows baseline patient characteristics in online supplemental table S6. The overall mean age at randomization was 70.9±7.4 years, and 54 patients (38.3%) were men. 14 patients (9.9%) were never-smokers, and 27 (19.1%) had a PD-L1 tumor proportion score of 1%. For purposes of independent external model validation, we also included patients from the STARS trial,3 a prospective single-arm study of SABR with a similar patient population (online supplemental tables S6).

I-SABR-SELECT optimizes treatment selection

The steps constituting the development of the I-SABR-SELECT system are presented in figure 1b. Given the 165 radiomic features to characterize the tumor, peritumoral vasculature, and parenchymal regions, 43 robust and non-redundant features passed the screening and were retained for subsequent analysis (online supplemental tables S1−S3 and online supplemental figure S3d). These radiomic features were then integrated with clinical risk factors to build the predictive model termed “I-SABR-SELECT”, with number of selected features between 10 and 22. We observed that when patients were treated according to the model’s recommendation, they were associated with better outcomes in the I-SABR arm (HR (anti subgroup as reference)=0.06, p<0.001, figure 2a) as well as the SABR arm (HR=0.26, p=0.0042, figure 2b). A similar trend was observed in the STARS trial external validation cohort (HR=0.38, p=0.031, figure 2c). Per the model recommendation, a 1.3 years increase in the EFS time observed for the patients who received IO than those who were not recommended to receive IO, as measured by the difference in RMSTD (figure 2a). Furthermore, the composite model (ie, I-SABR-SELECT) outperformed unimodal models of radiomics alone (online supplemental figure S4) or clinical features alone (online supplemental figure S5). Further, the stratification of the I-SABR-SELECT model remained predictive when stratified by its recommendation. Among patients recommended for I-SABR, patients received the combination therapy had longer EFS than those who received SABR alone (HR (SABR as reference)=0.06, p<0.001), with the corresponding RMSTD for EFS extended by 1.43 years (figure 2d). Among patients recommended for SABR alone, patients received SABR did better than patients received I-SABR (HR, I-SABR as reference=0.30, p=0.024, RMSTD=1.2 years, figure 2e).

Figure 2. Predictive performance of I-SABR SELECT; (a–c) show subgroup analyses by treatment arms and (d–e) show subgroup analyses by model recommendation. Restricted mean survival time difference (RMSTD) was calculated at the median follow-up time, which is 33 months. EFS, event-free survival; SABR, stereotactic ablative radiotherapy.

Figure 2

Open in a new tab

Clinical values from stratification by I-SABR-SELECT over I-SABR trial’s original randomization

The ITE for each patient was estimated as the relative risk in the two treatment arms modeled by two CPH models (figure 3a, online supplemental figure S3c). The I-SABR-SELECT model recommended that 69 (49%) of the 141 patients in the I-SABR trial population switch treatments (specifically, 49 (65%) of the 75 patients in the SABR arm and 20 (30%) of those in the I-SABR arm). The predictive modeling showed that 95 (67%) of the 141 patients were expected to benefit from adding IO (ITE>0, IO-benefit stratum), with the remainder not expecting a benefit (ITE<0, IO-spare stratum). In the IO-benefit stratum, the average IO effect was more than twofold greater than in the randomized trial. The ARR improved from 23.3% to 47.0% while the NNT reduced from four to two patients, and the net benefit improved from 14.0% to 37.0% after I-SABR-SELECT was applied to optimize treatment (figure 3b; details in online supplemental tables S4 and S5). In the IO-spare stratum (ie, without an expected benefit from IO), we observed negative effect measures, with ARR of –26%, and a net benefit of –35%. Interestingly, in the STARS trial cohort, 48 (60%) of the 80 patients (figure 3a,c) were predicted to benefit from the addition of IO.

Figure 3. Individualizing treatment for I-SABR and STARS trials. (a) Waterfall plots demonstrating that prediction of treatment escalation; if at least 70% of the cross-validated models predicted no benefit from adding immunotherapy for a patient, he/she will be classified into IO-spare group. (b) Effect measures calculated based on the IO-benefit stratum and the IO-spare stratum for patients in the I-SABR trial, as well as the percentage of patients in each cohort whose treatment matched the model recommendation or were recommended to switch their original treatment. (c) IO-benefit and IO-spare stratum for patients in the STARS trial. No effect measures could be calculated for this single-arm trial. *ARR, absolute risk reduction; IO, immunotherapy; NNT, number needed to treat; RRR, relative risk reduction; SABR, stereotactic ablative radiotherapy.

Figure 3

Open in a new tab

Furthermore, we observed a high significance of interaction effects based on the interaction plots (online supplemental figure S6), with the crossed lines in I-SABR and STARS indicating substantial treatment effects. Patients who followed the model’s recommendation had significant risk reduction, such as >25% risk reduction when the model suggested not adding IO.

Interpretation and reasoning behind I-SABR-SELECT

We further assessed the importance of each individual feature to the predictions of our I-SABR-SELECT model. Specifically, we computed SHAP values, which are based on cooperative game theory, to distribute the contribution of a prediction among its features. During the cross-validation, the individual models selected a subset of 10–22 features; the most frequently selected features were radiomic features quantifying the tumor microenvironment, tumor, and surrounding vessels (online supplemental figure S3b). These important radiomic features are displayed in figure 4a; ECOG performance status and tobacco exposure were among the most important treatment moderators among the clinical factors, meaning they had weights that could modify the treatment effect.

Figure 4. Visualization of the radiomics features incorporated into the I-SABR-SELECT model. (a) Imaging biomarkers selected by swarm intelligence. For vessel features, the hot-spot area was defined as a 10 mm radius from the tumor core. (b) SHapley Additive exPlanations summary plot illustrating the features ranked by their importance and direction in defining treatment effects. ECOG, Eastern Cooperative Oncology Group; SABR, stereotactic ablative radiotherapy.

Figure 4

Open in a new tab

The associated SHAP values of these clinical and radiomic features, ranked in order of their effect magnitudes, are shown in figure 4b. The complexity of the vessels surrounding the tumor as well as tumor morphologies was among the top features that influenced treatment selection, with a more complex surrounding angiogenesis network associated with no benefit from combined IO and SABR, while large tumors derived more benefit from the combined regimen. Radiomic features measuring the intensity and texture of lung parenchyma and tumor also contributed to the model’s predictions, such as higher contrast difference between tumor and its surrounding tissue having been associated with greater benefit from combined therapy. Regarding the clinical variables included, a poorer ECOG performance status as well as smokers correlated with augmented benefit from combination treatment. We also observed that, even though some features on their own were significantly associated with different recommended treatment groups by I-SABR-SELECT on univariate analysis (online supplemental figure S7a−d), they did not reach the same performance as I-SABR SELECT (online supplemental figure S8). This further proved that the complex interplay between these moderators jointly defined the treatment effect modification.

To showcase the clinical promise of our predictive modeling, we chose representative cases from both the model-predicted IO-benefit and IO-spare groups (figure 5). Within the IO-benefit group, some patients underwent SABR only, but the I-SABR-SELECT algorithm pinpointed these cases as having positive treatment effects, suggesting that these patients may have benefited from adding IO. Notably, these patients experienced events, such as distant and regional failures, within 4 months of treatment completion. Interestingly, we found no singular radiomic feature within the tumor, vessels, or lungs that was consistently associated with the predictions, indicating that the model’s highly non-linear nature was driven by the collective interactions among these features. Similarly, within the IO-spare group, we observed analogous patterns, with patients potentially able to forego IO, as evidenced by the early events that occurred between 7 and 20 months after the addition of IO.

Figure 5. Radiologic images and multidimensional features contributing to I-SABR-SELECT model recommendations. Cases labeled in blue are patients who received SABR during the randomized controlled trial, but I-SABR SELECT predicted a positive treatment effect from adding immunotherapy (IO). These patients developed early progression between 3 and 4 months after treatment, and I-SABR-SELECT interpreted them as having high-risk diseases that would benefit from the addition of immunotherapy. Cases labeled in green are patients who received I-SABR during the randomized controlled trial, but I-SABR-SELECT predicted a negative treatment effect from receiving IO. These patients had early events between 7 and 20 months after treatment and thus were considered not to benefit from immunotherapy and thus could be safely excluded from treatment escalation. ECOG, Eastern Cooperative Oncology Group; SABR, stereotactic ablative radiotherapy.

Figure 5

Open in a new tab

Discussion

This analysis aimed to refine patient selection for patients with early-stage NSCLC undergoing SABR who may benefit from added IO. This novel prespecified analysis of a randomized phase II trial used AI technology to combine clinical and radiomic data to address this question and thereby expand the notion of “precision medicine”. Results revealed that the interplay of performance status, SABR dose, tumor size, tobacco exposure, and multiple novel radiomic features was predictive. The model doubled the absolute EFS benefit compared with the initial I-SABR trial randomization.2 If validated, this model could guide future trials involving combined IO and SABR. It could also serve as a reference for the off-label use of I-SABR in real-world scenarios.

While radiomics is widely used for outcome prognostic purposes, its predictive role in treatment selection before treatment is delivered remains limited.27 Distinct from previous radiomic models focused on predicting certain clinical outcomes as a prognostic tool, in early-stage lung cancer,28,30 I-SABR-SELECT is the first predictive model for guiding treatment selection, not just a prognostic tool. Using AI (counterfactual reasoning), it contrasted outcomes between randomized arms to identify clinical-radiomic patterns linked to the benefits of adding IO.

The I-SABR-SELECT algorithm draws from the inaugural randomized trial demonstrating the safety and efficacy of combining IO with SABR for early-stage NSCLC.2 Unlike other reports illustrating only 30% of metastatic patients benefit from IO alone,8 this model identified that 67% (95 of 141) of early-stage patients in the I-SABR trial and 60% (48 of 80) in the STARS trial could benefit from adding IO. This may be due to the potential synergistic effects of combining IO and SABR, such as enhanced tumor antigen exposure and immune activation.1 2 Further validation is planned through ongoing phase III trials (KEYNOTE-867, NCT03924869; PACIFIC-4, NCT03833154, and SWOG/NRG S1914, NCT04214262). Recent news releases of KEYNOTE-867 showed that the combined IO/SABR arm did not cross the statistical significance boundary in recurrence-free survival as compared with SABR alone in early-stage NSCLC. SWOG/NRG S1914 was also closed prematurely after interim analysis. These data appear to conflict with the I-SABR results, thereby making the individualized I-SABR-SELECT approach even more necessary and urgent.

We used SHAP analysis to interpret the relationship between the model selection process and its features. Radiomic features of tumor microenvironments, lung, and vessels were associated with the benefit of combining SABR with IO. Alilou et al17 previously reported a correlation between radiomic vasculature features and immune responses to checkpoint inhibitors in metastatic NSCLC. Smokers likely benefit more from IO due to a higher mutational burden caused by smoking-related carcinogens and fewer EGFR/ALK mutations. Poor performance status (fragile/comorbid) patients may require additional immune stimulation to boost the immune response, explaining their greater benefit with added IO. Consistent with other reports,31 we observed that larger, irregularly shaped tumors tended to benefit from additional IO. Interestingly, higher SABR doses (70 Gy, with integrated boost to the gross tumor to 80 Gy in 10 fractions) were associated with a greater immunotherapeutic benefit. While case selection with 70 Gy in 10 fractions (typically used for more centrally located and/or larger tumors) could lead to higher recurrence (thereby supporting IO in that manner), the impact of SABR regimens (daily/total doses and its distribution in the chest) merits further exploration. Although biologically effective doses of 50 Gy in 4 fractions and 70 Gy in 10 fractions are similar (112.5–119 Gy), the immunological effective doses of these regimens remain an open question. It is worth emphasizing that the decision-making process of the I-SABR-SELECT model was highly non-linear, as demonstrated in representative cases (figure 5), in that it is not driven by individual features but rather by the collective interactions among them.

A notable strength of our research lies in the high-caliber study population from the phase II randomized I-SABR trial, which helped to minimize enrollment biases and confounding variables. To address limited validation cohorts, we used a repeated cross-validation to develop the first predictive model for guiding the choice of therapy. However, several limitations herein merit further discussion. First, because it was based on the I-SABR trial, the sample size was limited, and no similar external datasets exist as yet to validate these findings. Although the validation was conducted with the single-arm STARS trial (SABR only) was intriguing, further validation with comparable and larger I-SABR cohorts in the future is still needed. Second, our approach involved the application of hand-crafted radiomic features in conjunction with clinical risk factors for predictive modeling. As we demonstrated previously,14 leveraging deep learning on extensive samples has the potential to uncover valuable patterns beyond hand-crafted features. This avenue could be explored in the future when larger datasets of I-SABR-treated samples become available. Future improvements could include integrating baseline positron emission tomography/computed tomography (PET/CT) scans to assess tumor metabolism given their demonstrated clinical value in lung SABR32 and molecular profiling (eg, driver gene mutation or PD-L1 status) to enhance predictive value. Additionally, the model could evolve from the single-objective optimization to a multi-objective problem by factoring in side effects and EFS, thereby mirroring real clinical decision-making scenarios. With further development, I-SABR-SELECT could serve as the foundation for a treatment-focused digital twin by incorporating continuous patient monitoring, real-time adaptive modeling, and personalized simulations to dynamically refine treatment recommendations over time.

In summary, the I-SABR-SELECT model is the first known predictive tool for guiding treatment decisions in early-stage NSCLC, using AI to optimize patient selection for adding IO to SABR. The stratification led to a twofold increase in absolute EFS benefit compared with the original outcomes in the I-SABR trial. These promising results emphasize the need for further validation in larger trial cohorts to fine-tune the model and assess its clinical utility for guiding therapeutic choices. If further validated, this novel AI strategy could also have utility for locally advanced and stage IV NSCLC to improve the therapeutic ratio and establish individualized radiotherapy and IO.33

Supplementary material

online supplemental file 1
jitc-13-10-s001.docx (5.2MB, docx)
DOI: 10.1136/jitc-2025-013074

Acknowledgements

We thank Christine Wogan of The University of Texas MD Anderson Cancer Center Division of Radiation Oncology for editorial help. We thank Rachel C Maguire and Luyang Yao for their help with participant screening, enrollment, and data collection.

Footnotes

Funding: This work was supported by Bristol-Myers Squibb and MD Anderson Cancer Center Alliance (CA209-925); and the National Cancer Institute at the National Institutes of Health through the Cancer Center Core (Support) Grant (P30CA016672) supported by the Clinical Trials Support Resource; and through a Clinical and Translational Science Award (UL1 RR024148) to the MD Anderson Cancer Center, as well as MD Anderson Lung Moon Shot Program and QIAC Partnership in Research Grant. The principal investigator of the I-SABR trial (JYC) is a recipient of the Texas 4000 Distinguished Professorship at the MD Anderson Cancer Center and the Joan and Herb Kelleher Charitable Foundation. The funders had no role in the study design; in the collection, analysis, or interpretation of data; in the writing of the report; or in the decision to submit the paper for publication.

Provenance and peer review: Not commissioned; externally peer reviewed.

Patient consent for publication: Not applicable.

Data availability free text: The raw CT images of ISABR and STARS are not publicly shared to protect patient privacy, but are available for research use from the corresponding author. MTA is required to be approved by MD Anderson committees by providing the research plan and is restricted to non-commercial academic research purposes. Request can be submitted to corresponding authors and will receive an internal review response within 30 days.

Data availability statement

Data are available upon reasonable request.

References

  • 1.Bernstein MB, Krishnan S, Hodge JW, et al. Immunotherapy and stereotactic ablative radiotherapy (ISABR): a curative approach? Nat Rev Clin Oncol. 2016;13:516–24. doi: 10.1038/nrclinonc.2016.30. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Chang JY, Lin SH, Dong W, et al. Stereotactic ablative radiotherapy with or without immunotherapy for early-stage or isolated lung parenchymal recurrent node-negative non-small-cell lung cancer: an open-label, randomised, phase 2 trial. Lancet. 2023;402:871–81. doi: 10.1016/S0140-6736(23)01384-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Chang JY, Mehran RJ, Feng L, et al. Stereotactic ablative radiotherapy for operable stage I non-small-cell lung cancer (revised STARS): long-term results of a single-arm, prospective trial with prespecified comparison to surgery. Lancet Oncol. 2021;22:1448–57. doi: 10.1016/S1470-2045(21)00401-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Chang JY, Senan S, Paul MA, et al. Stereotactic ablative radiotherapy versus lobectomy for operable stage I non-small-cell lung cancer: a pooled analysis of two randomised trials. Lancet Oncol. 2015;16:630–7. doi: 10.1016/S1470-2045(15)70168-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Ribas A, Wolchok JD. Cancer immunotherapy using checkpoint blockade. Science. 2018;359:1350–5. doi: 10.1126/science.aar4060. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Samstein RM, Lee C-H, Shoushtari AN, et al. Tumor mutational load predicts survival after immunotherapy across multiple cancer types. Nat Genet. 2019;51:202–6. doi: 10.1038/s41588-018-0312-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Hellmann MD, Ciuleanu T-E, Pluzanski A, et al. Nivolumab plus Ipilimumab in Lung Cancer with a High Tumor Mutational Burden. N Engl J Med. 2018;378:2093–104. doi: 10.1056/NEJMoa1801946. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Fehrenbacher L, Spira A, Ballinger M, et al. Atezolizumab versus docetaxel for patients with previously treated non-small-cell lung cancer (POPLAR): a multicentre, open-label, phase 2 randomised controlled trial. Lancet. 2016;387:1837–46. doi: 10.1016/S0140-6736(16)00587-0. [DOI] [PubMed] [Google Scholar]
  • 9.Wu J, Mayer AT, Li R. Integrated imaging and molecular analysis to decipher tumor microenvironment in the era of immunotherapy. Semin Cancer Biol. 2022;84:310–28. doi: 10.1016/j.semcancer.2020.12.005. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology. 2016;278:563–77. doi: 10.1148/radiol.2015151169. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Lambin P, Leijenaar RTH, Deist TM, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol. 2017;14:749–62. doi: 10.1038/nrclinonc.2017.141. [DOI] [PubMed] [Google Scholar]
  • 12.Wu J, Li C, Gensheimer M, et al. Radiological tumor classification across imaging modality and histology. Nat Mach Intell. 2021;3:787–98. doi: 10.1038/s42256-021-00377-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Bera K, Braman N, Gupta A, et al. Predicting cancer outcomes with radiomics and artificial intelligence in radiology. Nat Rev Clin Oncol. 2022;19:132–46. doi: 10.1038/s41571-021-00560-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Saad MB, Hong L, Aminu M, 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:e404–20. doi: 10.1016/S2589-7500(23)00082-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Isensee F, Jaeger PF, Kohl SAA, et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18:203–11. doi: 10.1038/s41592-020-01008-z. [DOI] [PubMed] [Google Scholar]
  • 16.Primakov SP, Ibrahim A, van Timmeren JE, et al. Automated detection and segmentation of non-small cell lung cancer computed tomography images. Nat Commun. 2022;13:3423. doi: 10.1038/s41467-022-30841-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Alilou M, Khorrami M, Prasanna P, et al. A tumor vasculature-based imaging biomarker for predicting response and survival in patients with lung cancer treated with checkpoint inhibitors. Sci Adv. 2022;8:eabq4609. doi: 10.1126/sciadv.abq4609. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Hofmanninger J, Prayer F, Pan J, et al. Automatic lung segmentation in routine imaging is primarily a data diversity problem, not a methodology problem. Eur Radiol Exp. 2020;4:50. doi: 10.1186/s41747-020-00173-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res. 2017;77:e104–7. doi: 10.1158/0008-5472.CAN-17-0339. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Kirillov A, Mintun E, Ravi N, et al. Segment anything. arXiv. 2023:230402643. doi: 10.1109/ICCV51070.2023.00371. [DOI] [Google Scholar]
  • 21.Mirjalili S, Mirjalili SM, Lewis A. Grey Wolf Optimizer. Adv Eng Softw. 2014;69:46–61. doi: 10.1016/j.advengsoft.2013.12.007. [DOI] [Google Scholar]
  • 22.Danieli C, Bossard N, Roche L, et al. Performance of two formal tests based on martingales residuals to check the proportional hazard assumption and the functional form of the prognostic factors in flexible parametric excess hazard models. Biostatistics. 2017;18:505–20. doi: 10.1093/biostatistics/kxw056. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Weir IR, Marshall GD, Schneider JI, et al. Interpretation of time-to-event outcomes in randomized trials: an online randomized experiment. Ann Oncol. 2019;30:96–102. doi: 10.1093/annonc/mdy462. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Ranganathan P, Pramesh CS, Aggarwal R. Common pitfalls in statistical analysis: Absolute risk reduction, relative risk reduction, and number needed to treat. Perspect Clin Res. 2016;7:51–3.:51. doi: 10.4103/2229-3485.173773. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25.Dorresteijn JAN, Visseren FLJ, Ridker PM, et al. Estimating treatment effects for individual patients based on the results of randomised clinical trials. BMJ. 2011;343:d5888. :d5888. doi: 10.1136/bmj.d5888. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Lundberg SM, Lee S-I. A unified approach to interpreting model predictions. Adv Neural Inf Process Syst. 2017;30 [Google Scholar]
  • 27.Al-Tashi Q, Saad MB, Muneer A, et al. Machine Learning Models for the Identification of Prognostic and Predictive Cancer Biomarkers: A Systematic Review. Int J Mol Sci. 2023;24:7781. doi: 10.3390/ijms24097781. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Aerts HJWL, Velazquez ER, Leijenaar RTH, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun. 2014;5:4006. doi: 10.1038/ncomms5006. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Dissaux G, Visvikis D, Da-Ano R, et al. Pretreatment 18F-FDG PET/CT Radiomics Predict Local Recurrence in Patients Treated with Stereotactic Body Radiotherapy for Early-Stage Non-Small Cell Lung Cancer: A Multicentric Study. J Nucl Med. 2020;61:814–20. doi: 10.2967/jnumed.119.228106. [DOI] [PubMed] [Google Scholar]
  • 30.Vaidya P, Bera K, Gupta A, et al. CT derived radiomic score for predicting the added benefit of adjuvant chemotherapy following surgery in stage I, II resectable non-small cell lung cancer: a retrospective multicohort study for outcome prediction. Lancet Digit Health. 2020;2:e116–28. doi: 10.1016/S2589-7500(20)30002-9. [DOI] [PubMed] [Google Scholar]
  • 31.Thor M, Fitzgerald K, Apte A, et al. Exploring published and novel pre-treatment CT and PET radiomics to stratify risk of progression among early-stage non-small cell lung cancer patients treated with stereotactic radiation. Radiother Oncol. 2024;190:109983. doi: 10.1016/j.radonc.2023.109983. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Wu J, Aguilera T, Shultz D, et al. Early-Stage Non-Small Cell Lung Cancer: Quantitative Imaging Characteristics of (18)F Fluorodeoxyglucose PET/CT Allow Prediction of Distant Metastasis. Radiology. 2016;281:270–8. doi: 10.1148/radiol.2016151829. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Chang JY, Verma V, Weichselbaum RR. Reconciling the discrepancies in randomized data of combining immunotherapy and radiation therapy: Not all radiotherapy is created equal. Eur J Cancer. 2024;201:113972. doi: 10.1016/j.ejca.2024.113972. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

online supplemental file 1
jitc-13-10-s001.docx (5.2MB, docx)
DOI: 10.1136/jitc-2025-013074

Data Availability Statement

Data are available upon reasonable request.

Articles from Journal for Immunotherapy of Cancer are provided here courtesy of BMJ Publishing Group

RESOURCES