Non-small cell lung cancer accounts for 85% of all lung cancer cases and is the leading cause of cancer death worldwide. While surgical resection offers the best chance of cure for early-stage disease, recurrence rates are devastating: 50 to 90 percent of patients relapse within just 2 years after surgery, and up to 95 percent within 5 years.
Traditional TNM staging is the standard framework for guiding treatment decisions, but patients with identical disease stages often have dramatically different survival outcomes, confirming that staging alone cannot capture the full complexity of NSCLC biology.
Previous efforts to improve recurrence prediction using radiomics have focused almost exclusively on features from the tumor itself (the gross tumor volume, GTV) or the immediately surrounding tissue. However, cancer is a systemic disease, and the metabolic and anatomic signatures of distant organs captured in whole-body PET/CT scans contain information that has been largely ignored.
This study introduces MetaPredictomics, a stacked ensemble machine learning framework that integrates clinicopathologic data, GTV radiomics, and radiomic features extracted from 19 presumed healthy organs (organomics) from presurgical PET/CT scans to predict recurrence-free survival in NSCLC patients.
The NSCLC RadioGenomics dataset from The Cancer Imaging Archive provided 145 patients with complete imaging and clinicopathologic data. All patients underwent 18F-FDG PET/CT before surgery between 2008 and 2012, with surgical tissue analysis providing histopathologic and gene mutation data. Of the 145 patients, 41 experienced recurrence during follow-up.
A deep learning nnU-Net model automatically segmented 19 organs from CT images including whole lung, affected lung lobe, whole heart, cardiac substructures, aorta, esophagus, liver, pancreas, spleen, adrenal glands, rectum, rib cage, vertebra, sacrum, and hip. Visual quality checks were performed on all segmentations.
Using PyRadiomics, 107 features were extracted from each region of interest from both CT and PET images, covering first-order statistical features, 3D and 2D morphology, gray-level co-occurrence matrix (GLCM) texture, and other texture matrix families. This produced 214 features per organ, yielding organomics feature sets for each of the 19 organs plus the GTV.
The clinicopathologic feature set incorporated patient characteristics such as age, sex, and smoking status, histologic grade, lymphovascular and pleural invasion status, KRAS and EGFR mutation status, pathologic TNM staging, conventional PET biomarkers including SUVmax and total lesion glycolysis, and treatment history including chemotherapy and radiation.
The MetaPredictomics framework uses a two-level stacked ensemble approach. In the first level, a separate time-to-event prediction model (glmboost, a gradient boosted generalized linear model) is trained independently for each feature set: GTV radiomics, each of the 19 organomics sets, and the clinicopathologic set.
Feature selection within each first-level model used Spearman correlation to remove redundant features and univariate Cox proportional hazards analysis with 100 bootstrap iterations to rank features by their average concordance index (C-index). The top 10 features from each set were retained for model training.
The risk scores output by each first-level model serve as input features for a second-level meta-model, also a glmboost. All combinations of first-level models with average C-index of at least 0.6 were explored via grid search to identify the optimal ensemble combinations, evaluated by 3-fold cross-validation with preservation of recurrence event ratios across folds.
Model performance was compared using the Mann-Whitney U test on C-index distributions across 1,000 bootstrap iterations per fold, with Benjamini-Hochberg correction for multiple comparisons. A nomogram was also constructed from the best meta-model to provide individualized 1, 3, and 5-year recurrence probability estimates.
The clinicopathologic model achieved the highest standalone performance with an average C-index of 0.67 across three external folds. This confirms that conventional biomarkers such as pathologic TNM staging, PET metrics including SUVmax, SUVpeak, and total lesion glycolysis, and treatment history remain the most informative single source for recurrence prediction.
The GTV radiomics model and the whole-lung model both matched with a C-index of 0.65, closely behind the clinicopathologic model. The aorta organomics model also reached 0.65, followed by esophagus and adrenal glands at 0.63, demonstrating that information from distant organs carries meaningful prognostic signal.
In total, 14 of the 21 first-level models achieved an average C-index of 0.6 or higher and were eligible for inclusion in the stacked ensemble. The whole-lung model outperformed the affected-lung-lobe model (0.65 versus 0.63), likely because some tumors extend beyond a single lobe, making the whole-lung region a more comprehensive region of interest.
Kaplan-Meier analysis confirmed that GTV, whole-lung, sacrum, and aorta models significantly stratified patients into low and high-risk groups by log-rank test, validating the survival discrimination ability of these individual models despite their moderate C-index values.
All top 100 meta-models significantly outperformed every first-level model. The top 100 meta-models ranged from C-index 0.703 to 0.731, compared to the best first-level C-index of 0.67, demonstrating robust and consistent improvement from multi-source integration.
The best-performing meta-model combined clinicopathologic, GTV, whole-lung, esophagus, and pancreas models, achieving a C-index of 0.731. The second-best model (clinical plus GTV plus whole-lung plus sacrum) achieved a C-index of 0.816 on fold 1 alone, the highest single-fold performance observed in the study.
Examining which first-level models appeared most often among the top 100 meta-models revealed a clear hierarchy: the clinicopathologic model was present in 98 of 100, whole lung in 71, esophagus in 69, pancreas in 62, and GTV in 61, identifying these five as the most consistently valuable input sources for ensemble prediction.
The best meta-model produced the lowest log-rank P-value of 0.00041 in Kaplan-Meier analysis when stratifying patients into high and low-risk groups, significantly better than any individual first-level model, and calibration plots showed good agreement between predicted and observed recurrence at 1 and 2 years.
Pathologic T and N stage, SUVmax, SUVpeak, total lesion glycolysis, and radiation therapy history were selected in all three cross-validation folds of the clinicopathologic model, confirming their established importance and the robustness of these features as recurrence predictors.
In the GTV radiomics model, two CT features were consistently selected across all folds: the 90th percentile intensity, which captures dense tumor components such as high-cellularity areas and fibrotic tissue associated with aggressive biology, and the maximal correlation coefficient from the gray-level co-occurrence matrix, which reflects tumor textural heterogeneity linked to treatment failure.
In whole-lung organomics, the GLSZM gray-level non-uniformity on FDG-PET was selected in all folds, capturing global metabolic heterogeneity in the lungs that may reflect occult tumor spread, inflammation, or systemic tumor-related metabolic effects beyond the visible lesion.
Liver organomics features capturing metabolic texture variability were significantly associated with recurrence risk, consistent with the interpretation that deviations from the liver's normally uniform FDG uptake may signal systemic metabolic stress, inflammatory responses, or early occult metastatic involvement, all of which worsen prognosis.
The organomics concept is based on the hypothesis that the anatomic and metabolic state of distant organs captured in whole-body PET/CT contains information about systemic disease progression that is invisible in the primary tumor image alone. Organs such as the aorta and esophagus may show radiologic changes from direct tumor extension or lymphatic spread through the mediastinum, the primary route of NSCLC regional advancement.
The adrenal glands appeared as strong contributors, consistent with their status as a common metastatic site in NSCLC due to their rich vascular supply and drainage from the thoracic lymphatics. Increased FDG uptake heterogeneity or structural changes in the adrenal glands may reflect early or subclinical metastatic seeding before it is clinically detectable.
The pancreas, which appeared in 62 of 100 top meta-models, is not a common site of NSCLC metastasis, suggesting its metabolic signature may reflect systemic inflammatory or metabolic states associated with cancer-related cachexia or physiological stress that correlates with poorer prognosis rather than direct tumor involvement.
This work builds on a previous study from the same group that showed organomics improved overall survival prediction in NSCLC beyond GTV-only models. The present study extends this to recurrence-free survival and demonstrates that the value of distant-organ information is amplified when combined through ensemble learning rather than used in isolation.
A prognostic nomogram was constructed from the best meta-model to enable individualized prediction of recurrence probability at 1, 3, and 5 years post-surgery. Calibration analysis confirmed good agreement between predicted and observed outcomes at 1 and 2 years, though accuracy declined at 5 years, a known limitation of long-term prognostic models in heterogeneous cohorts.
The framework's modular design allows flexibility for clinical implementation: patients lacking certain data types can still receive predictions from the available subset of models, with the ensemble adjusting to include only the applicable first-level inputs. This is particularly valuable in real-world settings where not all imaging or biomarker data may be available for every patient.
The automated organ segmentation tool built on nnU-Net was essential for making organomics clinically feasible, as manual delineation of 19 organs per patient would impose an impractical time and labor burden on clinical radiologists. The open-source automated tool removes this barrier, enabling scalable deployment of the organomics pipeline.
Future clinical integration would benefit from prospective validation in larger and more diverse NSCLC cohorts, inclusion of brain imaging for patients whose PET/CT scan covered the skull, and incorporation of genomic or proteomic data as additional first-level inputs to further refine the ensemble model's predictive power.
MetaPredictomics demonstrates that maximizing the use of already-acquired PET/CT imaging data by extracting features from all visible organs, not just the tumor, substantially improves NSCLC recurrence prediction. No additional scans or invasive procedures are needed, making this an immediately translatable enhancement to existing clinical workflows.
The stacked ensemble approach was critical to realizing this improvement: individually, most organomics models achieved only modest C-index values, but their combination through meta-modeling consistently and significantly outperformed all standalone models, illustrating the power of integrating complementary information sources.
The finding that radiomics should complement rather than replace conventional clinical biomarkers is important for clinical adoption: pathologic staging, PET metrics, and treatment history remain indispensable, and the framework is most effective when these are combined with imaging-derived organ signals.
Scaling this approach to larger multicenter datasets with recurrence site documentation would allow future studies to determine whether specific organ models predict recurrence in those same organs, potentially enabling organ-targeted surveillance and adjuvant treatment strategies for high-risk NSCLC patients.