Enhancing cancer prognostics with group penalty models: a comparative study on radiomics feature selection in lung adenocarcinomas and meningiomas

Introduction

Radiomics, the extraction of intricate features from diagnostic images such as computed tomography (CT) and magnetic resonance imaging (MRI), is revolutionizing cancer management by revealing patterns invisible to the naked eye (1). This technique enables quantitative analysis of tissue complexity, facilitating precise cancer prognosis prediction and personalized treatment planning.

Radiomics features inherently exhibit natural group structures based on their extraction methods and characteristics (2). These natural groupings reflect fundamental biological and technical relationships among features. For instance, features extracted from the same segmentation region share spatial context and biological relevance, while features computed using the same mathematical framework capture related aspects of tissue heterogeneity (3). This inherent group structure suggests that considering features collectively within their natural groups, rather than treating them as independent variables, may provide more robust and biologically interpretable results in cancer prognostics.

The least absolute shrinkage and selection operator (Lasso) (2) has been the standard method for feature selection in high-dimensional radiomics data, effectively reducing complexity into radiomics scores (rad-scores) (3-5). Despite the widespread use of Lasso, its effectiveness in analyzing complex, high-dimensional data remains uncertain, with critiques pointing to potential limitations in such scenarios (6). More critically, Lasso is challenged by the presence of highly correlated features. When predictors are highly correlated, Lasso tends to select only one feature from a group of correlated features and ignores the others (7), failing to leverage the collective information within feature groups. This limitation is particularly problematic in radiomics, where features naturally cluster into meaningful groups based on their mathematical properties or biological origins. Consequently, this arbitrary selection among correlated features leads to instability in feature selection—small changes in the data can result in different features being selected from the same correlated group, undermining the reliability and reproducibility of the model (8).

Recent work by Orton et al. (9) demonstrated that hierarchical correlation-based feature reduction could improve radiomics model interpretability in clear-cell renal cell carcinoma by considering the natural hierarchy of feature groups based on interpretability (shape features being more interpretable than first-order features, which in turn are more interpretable than texture features). However, their approach was limited to a single cancer type with predetermined, static group hierarchies and relied solely on group Lasso methodology.

Our study addresses these limitations by systematically comparing traditional Lasso with multiple group penalty approaches including group Lasso (10,11), group minimax concave penalty (group MCP) (11), group smoothly clipped absolute deviation (group SCAD) (11), and graper (12)—a Bayesian method that adaptively learns group-specific penalties from the data. Features within the same group often exhibit positive correlations, highlighting the potential of group-based analyses to enhance statistical power and improve reproducibility (13,14). To critically assess the effectiveness of these methods, we performed cross-validation based evaluation using existing radiomics datasets and validated the selected models through independent test sets. We hypothesized that group penalty models would provide more stable feature selection while maintaining predictive accuracy and improving interpretability of rad-scores. This comparative analysis was conducted through two case studies: the prediction of spread through air spaces (STAS) in early-stage lung adenocarcinomas (15) and the pre-surgical grading of meningiomas (16). We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1951/rc).

Methods

Patients

Lung adenocarcinomas study

Lung adenocarcinoma is the most prevalent form of lung cancer and the leading cause of cancer-related deaths. STAS, introduced as a novel prognostic factor in the 2015 World Health Organization (WHO) classification, is correlated with increased recurrence risk and diminished survival rates. Detecting STAS is crucial for surgical planning; however, challenges exist in its identification using current preoperative and intraoperative methods. Radiomics offers a potential solution for these challenges through quantitative imaging analysis (15). This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

The cohort used in this research was initially collected in a prior study (15). Data from 590 lesions with stage I lung adenocarcinoma and presurgical chest CT images were used to develop and validate the radiomics model. The training set comprised 432 lesions (281 STAS-negative and 151 STAS-positive) from Severance Hospital (January 2016 to December 2017), Korea University Anam Hospital, and Yongin Severance Hospital, while the test set included 158 lesions (121 STAS-negative and 37 STAS-positive) from Hospital A (January to December 2018). The patient demographics are summarized in Table S1. A flowchart illustrating the study population is depicted in Figure 1A. Radiomics features were extracted from three distinct regions: the nodule, its 3-mm dilated periphery, and a combined region, yielding 279 features per region using Pyradiomics (RRID:SCR_026019) (17). These features included histogram, texture [gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), neighboring gray tone difference matrix (NGTDM)], and shape descriptors (Table S2).

Figure 1 Selection flowchart for (A) STAS lung adenocarcinomas and (B) meningioma grading. Hospital A: Severance Hospital; Hospital B: Korea University Anam Hospital; Hospital C: Yongin Severance Hospital. STAS, spread through air spaces.

Meningiomas study

Preoperative grading of meningiomas, the most prevalent primary intracranial tumors in adults, is critical for planning surgery and predicting prognosis. The WHO classification classifies these tumors into WHO grade 1, 2, and 3. While most meningiomas are WHO grade 1 (“low-grade”), more than one-fifth are WHO grade 2 or 3 (“high-grade”), demonstrating aggressive behavior and recurrence, leading to lower 5-year survival rates. Utilizing noninvasive radiomics, analyzing features from T1-weighted contrast (T1C)-enhanced images and diffusion tensor imaging (DTI), aids in this differentiation, especially for incidentally discovered tumors, which may not require histological verification and can be monitored without immediate intervention (16).

The cohort used in this research consists of data previously collected in prior study (16). Data from 194 patients with confirmed pathology and preoperative MRI scans were used to develop and validate the radiomics model. The training set comprised 136 patients (108 with low-grade and 28 with high-grade tumors) from June 2010 to October 2016, while the test set included 58 patients (46 with low-grade and 12 with high-grade tumors) from November 2016 to March 2018. The patient demographics are summarized in Table S1. A flowchart depicting the study population is shown in Figure 1B. After preprocessing with N4 bias correction and intensity normalization, regions of interest were delineated on T1C-enhanced images. A total of 98 radiomics features were extracted from T1C, apparent diffusion coefficient (ADC), and fractional anisotropy (FA) maps using 3D Slicer (RRID:SCR_005619), including histogram-based, co-occurrence matrix-based, and run-length matrix-based parameters, plus morphological features (Table S3).

Least absolute shrinkage and selection operator (Lasso)

Lasso is a widely used statistical method for feature selection in high-dimensional data analysis (2). In radiomics studies, where hundreds of features are extracted from medical images, Lasso helps identify the most relevant features while avoiding overfitting—a common problem when the number of features exceeds the number of patients.

The key advantage of Lasso is its ability to automatically select features by shrinking less important feature coefficients to exactly zero, effectively removing them from the model. This results in a simpler, more interpretable model that focuses only on the most predictive radiomics features. The degree of feature selection is controlled by a regularization parameter (λ), which is typically determined through cross-validation to optimize model performance. The detailed mathematical formulation is provided in the Appendix 1.

However, Lasso has important limitations, particularly when dealing with correlated features—a common scenario in radiomics data.

Limitations of Lasso with correlated features

When predictors are highly correlated, Lasso tends to select only one feature from a group of correlated features and ignore the others, which can be problematic when multiple correlated predictors are all relevant (7). This tendency introduces instability in feature selection, where small changes in the data can lead to large changes in the set of selected predictors. For instance, when data is split into different training and validation sets, Lasso may select different sets of features across the splits due to the inherent correlation structure among predictors (8). This instability undermines the reliability and reproducibility of the model, making it less robust for inference and prediction. Thus, the instability in feature selection by Lasso necessitates exploring alternative methods that can handle correlations more effectively and provide more stable and interpretable models, such as group penalty models which consider groups of correlated features.

Group penalty models

Group penalty models extend individual feature selection to groups of related features, which is particularly useful when features are naturally correlated or grouped. These models prioritize selecting groups of features over individual features based on predefined groupings from the study design or data characteristics (Figure 2). Unlike traditional Lasso, which can suffer from instability in feature selection, group penalty models handle correlations more effectively by considering the inherent group structures within the data (13,14).

Figure 2 Contrast between Lasso’s selection of individual features and group penalty models, which select clusters of correlated features in radiomics data analysis. graper makes feature selections with group information as a prior but does not make group-level selections. ADC, apparent diffusion coefficient; FA, fractional anisotropy; graper, adaptive penalization regression with external covariates using variational Bayes; Lasso, least absolute shrinkage and selection operator.

The group penalty models evaluated in this study include group Lasso (10,11), group MCP (11), group SCAD (11), and graper (12). group Lasso applies a group-wise penalty that ensures either all features in a group are selected together or none at all. This all-or-nothing selection property makes it particularly suitable for radiomics applications where features within the same category are expected to be jointly relevant.

Group MCP and group SCAD employ non-convex penalties that address the bias issue inherent in group Lasso. While group Lasso applies constant shrinkage to all coefficients, potentially over-penalizing large coefficients, group MCP and group SCAD reduce this bias by applying minimal or no penalty to truly important feature groups. This property allows these methods to perform asymptotically as well as if the true model were known in advance.

Unlike the other group penalty methods that apply fixed penalties, graper takes a fundamentally different Bayesian approach. This method adaptively learns group-specific penalties from the data rather than applying the same penalty to all groups. Such flexibility allows automatic identification of important feature groups without extensive manual tuning. The method uses variational Bayes approximation to maintain computational efficiency while providing explicit importance scores for each feature group. Additionally, graper was implemented with Random Over-Sampling Examples (ROSE) (17) to address data imbalances in our cancer classification tasks.

The Appendix 1 provides detailed algorithmic descriptions and mathematical formulations for the group Lasso, group MCP, and group SCAD penalties, as well as the hierarchical Bayesian model and variational inference procedure employed in graper.

Feature grouping strategy for lung adenocarcinomas study

In the STAS prediction radiomics study, our methodological approach for prior group selection was anchored to three strategic feature classifications. (I) Segmentation regions: we grouped the features into three primary categories based on their distinct segmentation regions: nodule, peripheral, and a “nodule + peripheral” region. This classification establishes the foundation for our analysis by clearly differentiating the features according to their extraction areas. (II) Radiomics feature types: to further enhance the precision of our group penalty models, we performed a careful categorization of each radiomic feature into one of six distinct types based on their radiometric attributes: histogram, GLCM, GLRLM, GLSZM, NGTDM, and shape descriptors. This detailed categorization allowed for a more nuanced analysis of radiomic features. (III) Combination: this granular classification harmonized with the initial segmentation categories, leading to the creation of 18 unique groups. This sophisticated structure combines the segmentation regions (nodule, peripheral, and “nodule + peripheral”) with radiomics feature types (histogram, GLCM, GLRLM, GLSZM, NGTDM, and shape). A summary of these groups is presented in Figure 3 and Table S4, where the numbers in each cell indicate the count of features belonging to that specific group.

Figure 3 Group structure of radiomics feature groups for the lung adenocarcinomas study. Depicting the organization of features into three main levels: (I) segmentation regions; (II) radiomics feature types; (III) combination: combinations of the segmentation regions and radiomics feature types. GLCM, gray level co-occurrence matrix; GLRLM, gray level run length matrix; GLSZM, gray level size zone matrix; NGTDM, neighboring gray tone difference matrix.

Feature grouping strategy for meningiomas study

In the study of pre-surgical grading of meningioma, our methodological approach to prior group selection was anchored in three strategic feature classifications. (I) Imaging techniques: initially, we organized the features according to medical imaging techniques, identifying three main categories: T1C, ADC, and FA. Additionally, we set the morphology as a separate group, laying the groundwork for our analysis. (II) Radiomics feature types: to further refine our group penalty models, we classified each radiomic feature into one of four established types according to their mathematical and computational definitions: histogram, co-occurrence matrix (CM), run-length matrix (RLM), and morphology. This detailed categorization allowed for a more nuanced analysis of radiomic features. (III) Combination: this resulted in a comprehensive arrangement of ten distinct groups: three groups for each imaging techniques (T1C, ADC, and FA) combined with the three radiomics feature types (histogram, CM, and RLM) and one additional group for morphology. Figure 4 and Table S5 present a summary of these groups, where the numbers in each cell indicate the count of features belonging to that specific group.

Figure 4 Group structure of radiomics feature groups for the meningiomas study. Depicting the organization of features into three main levels: (I) imaging techniques; (II) radiomics feature types; (III) combination: combinations of the imaging techniques and radiomics feature types, including morphology feature types. CM, co-occurrence matrix; RLM, run-length matrix; T1C, T1-weighted contrast; DTI, diffusion tensor imaging; FA, fractional anisotropy.

Statistical analysis

In our study, we evaluated model performance using nested 10-fold cross-validation (18). This method divides the entire dataset into 10 equal parts, each of which sequentially serves as the validation fold while the remaining nine parts collectively form the training fold. This process is repeated 10 times, with each subset serving as the validation fold once. The process of the cross-validation is depicted in Figure S1. Within each training fold, we further performed a 10-fold cross-validation to determine the optimal value of the regularization parameter, λ, based on the “deviance” criterion. Specifically, we identified the λ_min, the value of λ that minimizes the deviance. For the group penalty models, we fixed the gamma values based on recommendations (11): γ=3 for group MCP and γ=4 for group SCAD, following the default suggestions by Fan and Li (19) and Zhang (20). For the graper method, model performance was evaluated using standard 10-fold cross-validation, reflecting its modeling strategy without the need for hyperparameter tuning. Features with posterior inclusion probability exceeding 0.5 were selected for the final model, following the median probability model criterion (21). ROSE-based oversampling was applied within each training fold to balance class distributions prior to model fitting, as shown in Figure S1.

The rad-score for each case was derived through a linear combination of selected features weighted by their coefficients, as determined by Lasso and group penalty models using the training folds. To assess the predictive performance of the rad-scores, we used the area under the receiver operating characteristic (ROC) curve (AUC) (22) as the primary metric. A high AUC value indicates the model’s superior ability to accurately predict the presence or absence of a disease, including its progression stages. To provide a comprehensive assessment of the model’s performance, we calculated and presented the mean and standard deviation (SD) of the AUC values obtained from the results of each validation fold, thereby evaluating the model’s consistency and reliability across different data subsets.

We employed the Jaccard Index (23) as a key metric for evaluating the stability of feature selection in our models. The Jaccard Index was calculated by dividing the number of common features selected in both folds (intersection) by the total number of unique features selected across both folds (union). For the 10-fold cross-validation, this computation was performed for every possible pair of folds, yielding 45 unique comparisons. The average Jaccard Index across these comparisons provided a quantitative measure of the consistency of the model in feature selection. This process is illustrated in Figure S2, which explains the Jaccard Index.

To assess the parsimony of our models, we examined the number and variability of selected features using the mean and SD across a 10-fold cross-validation. A more parsimonious model that selects fewer features is favored owing to its simplicity and interpretability.

To confirm the applicability in clinical practice, we selected the group penalty model that achieved the highest AUC for further evaluation. To compare these models, DeLong’s test for two correlated ROC curves (24) was performed.

To evaluate model calibration, we calculated Brier scores (mean squared error between predicted probabilities and actual outcomes) and generated calibration curves comparing predicted probabilities against observed frequencies (25).

Statistical analysis was performed using the statistical software R version 4.3.1 (R Foundation for Statistical Computing, RRID:SCR_001905) (26). The Lasso analysis was conducted using the “glmnet” package (version 4.1.8) (27). The group Lasso, group MCP, and group SCAD analyses were conducted using the “grpreg” package (version 3.5.0) (11) while the graper analysis was implemented with the “graper” package (version 1.20.0) (12). The ROSE method was implemented using the “ROSE” package (version 0.0.4) (17).

Results

Cross-validation performance in lung adenocarcinomas study

In the STAS prediction study, the group SCAD model employing prior group selection (III) Combination achieved the highest AUC of 0.804 (SD =0.056). It exhibited a Jaccard Index of 0.613 (SD =0.293) and the Brier score of 0.167 (SD =0.035), while selecting an average of 29.6 features (SD =17.431). The Lasso model achieved an AUC of 0.776 (SD =0.059), a Jaccard Index of 0.503 (SD =0.084), and selected an average of 16.3 features (SD =2.669). The results are summarized in Table 1. To provide a detailed visualization of the groups of features selected by the group penalty methods in each fold, heatmaps have been organized in Figures S3-S5.

Cross-validation performance in meningiomas study

In the meningioma grading study, the group Lasso model employing the prior group selection (I) Imaging techniques achieved the highest AUC of 0.816 (SD =0.143). It also exhibited a Jaccard Index of 0.7 (SD =0.39) and the Brier score of 0.139 (SD =0.043), while selecting an average of 14 features (SD =12.649). The Lasso model achieved an AUC of 0.743 (SD =0.223), a Jaccard Index of 0.41 (SD =0.183), and selected an average of 4.9 features (SD =1.595). These findings are summarized in Table 2. To provide a detailed visualization of the groups of features selected by the group penalty methods in each fold, heatmaps have been organized in Figures S6-S8.

Application

Lung adenocarcinomas study

Based on the cross-validation results, we applied the best-performing group SCAD [(III) Combination] model and Lasso to an independent test set for validation. The Lasso model achieved an AUC of 0.877 [95% confidence interval (CI): 0.815–0.939], while the group SCAD [(III) Combination] model achieved an AUC of 0.874 (95% CI: 0.814–0.934). The test results showed no significant difference between the AUCs of the two models (P=0.896), with a 95% CI for the difference in AUCs ranging from −0.034 to 0.039. This suggests that both models have comparable performance in terms of AUC.

Unlike Lasso, which selects features individually across multiple groups—potentially leading to fragmented interpretations—group SCAD focuses on the “nodule shape” and “nodule + peripheral histogram” feature groups. These results are summarized in Table 3.

Meningiomas study

Based on the cross-validation results, we applied the best-performing group Lasso [(I) Imaging techniques] model and Lasso to an independent test set for validation. The Lasso model achieved an AUC of 0.835 (95% CI: 0.706–0.965), while the group Lasso [(I) Imaging techniques] model achieved an AUC of 0.877 (95% CI: 0.779–0.974). The results showed no significant difference between the AUCs of the two models (P=0.391). The 95% CI for the difference in AUCs ranged from −0.137 to 0.054, indicating comparable performance between the two models.

Unlike Lasso, which selects features individually across multiple groups, group Lasso focuses on the morphology feature group. These results are summarized in Table 4.

Discussion

Our findings highlight the superiority of strict group penalty (group Lasso, group MCP and group SCAD) model over Lasso and graper in terms of both AUC and Jaccard Index, with only a few exceptions. This improvement highlights the benefits of integrating the group structures of features, which appear to enhance the models’ discriminatory capability and consistency in feature selection. Moreover, group penalty models achieved similar or lower Brier scores compared to Lasso, suggesting that the integration of group structures does not compromise the predictive performance of the models. We observed a correlation between the size of the feature group and the Jaccard Index, particularly evident when comparing (I) the smaller groupings with (III) the extensive groupings. This suggests that models incorporating broader grouping of features tend to exhibit more stable feature selection across different cross-validation folds. Despite Lasso models selecting fewer features on average, their stability in feature selection is less consistent. Additionally, although graper is flexible in incorporating group information as a prior, it showed lower Jaccard Index values than Lasso, indicating less stable feature selection, and exhibited the highest Brier scores among all methods, suggesting poor calibration. Consistent feature selection in radiomics holds multifaceted importance, primarily because it enhances the reliability, interpretability, and generalizability of predictive models, which are crucial for clinical applicability. Therefore, this consistency is foundational for advancing personalized medicine by translating complex radiomics data into reliable, understandable, and widely applicable diagnostic and prognostic tools (23,28).

Further analysis through train-test split comparisons in a clinical setting yielded encouraging results. The group penalty models demonstrated comparable predictive performance to Lasso, with no statistically significant differences in their AUC values for both lung adenocarcinomas and meningiomas studies. This finding is particularly promising as it shows that group-based feature selection can maintain high predictive accuracy while offering additional benefits. The group-based feature selection significantly contributed to clinically relevant insights. For instance, in the lung adenocarcinoma study, the group SCAD model focused on the “nodule shape” and “nodule + peripheral histogram” feature groups, highlighting their importance for STAS prediction. This contrasts sharply with the Lasso model, which selected features individually across multiple groups. The Lasso approach often results in the selection of a scattered set of features from various categories such as histogram, GLCM, GLSZM, and shape descriptors, leading to fragmented interpretations. This fragmentation can hinder the ability to draw comprehensive clinical insights, as the model does not consider the contextual synergy among the selected features. The group SCAD model’s emphasis on specific, clinically meaningful feature groups enables a more coherent interpretation. By focusing on the “nodule shape” and “nodule + peripheral histogram” groups, the group SCAD model integrates relevant radiomic features that collectively contribute to the biological understanding of STAS in lung adenocarcinomas. This approach not only improves interpretability but also aligns better with clinical reasoning, which often considers related features in conjunction rather than in isolation. In the meningioma study, the group Lasso model’s selection of the morphology group suggested that these features alone might be sufficient for accurate tumor grading, potentially simplifying preoperative evaluations. This is particularly significant as it implies that the detailed imaging techniques such as T1C and ADC, along with their respective complex feature types like histogram and co-occurrence matrix, might not be necessary for effective tumor grading. By focusing solely on morphology, which includes features such as volume, surface area, and sphericity, the group Lasso model simplifies the evaluation process, reducing the need for multiple imaging modalities and extensive feature extraction processes. This streamlined approach not only saves time and resources but also provides a more integrated understanding of tumor properties, enhancing the clinical utility and interpretability of the model.

However, our study has some limitations. The artificial grouping of features within our methodology may not fully capture the inherent structural complexities of the data, suggesting the need for further exploration of methodologies representing natural feature groupings more accurately.

While Lasso tends to under-select correlated features by choosing only one representative from a correlated group, group penalty models exhibit the opposite tendency—they may over-select features by including all features within a selected group. For instance, in our lung adenocarcinoma study, selecting the segmentation-based grouping could result in all 279 features from a single region being included in the model. This “all-or-nothing” property inherent to strict group penalty methods (group Lasso, group MCP, and group SCAD) may lead to potential overfitting, reduced generalizability, and diminished interpretability when large feature groups are selected in their entirety. Future studies should consider hybrid approaches, such as sparse group Lasso, which applies both group-level and within-group sparsity penalties to select relevant groups while simultaneously identifying the most informative individual features within those groups.

Furthermore, it is important to clarify the interpretation of the stability metrics reported in this study. The Jaccard Index in our group penalty models primarily reflects the consistency of group-level selection across cross-validation folds, rather than the precise selection of specific individual features based on their information value. When a group penalty model achieves high stability (e.g., Jaccard Index approaching 1.0), this indicates that the same feature groups are consistently selected, but all features within those groups are included together regardless of their individual predictive contributions. Therefore, “perfect stability” in group penalty models should be interpreted as consistent group selection rather than optimal individual feature identification. This distinction is crucial for clinical applications where understanding the contribution of specific features may be important for biological interpretation.

In addition, model calibration analysis revealed important considerations for clinical implementation (Figures S9-S11). The graper model consistently showed poor calibration compared to other methods, with systematic overestimation of predicted probabilities particularly evident in the calibration curves where graper consistently fell below the diagonal line. This is likely attributable to the ROSE oversampling procedure applied exclusively to this Bayesian method. Although ROSE effectively addressed class imbalance to improve discrimination, it may have distorted the probability estimates by altering the underlying class distribution (29). Furthermore, calibration curves exhibited substantial fluctuations in high-probability regions (predicted probability >0.7) due to limited sample sizes in these bins, a common challenge in medical imaging studies with moderate sample sizes (30).

Although we validated our models using independent test sets, these sets were obtained from the same institutions but from different time periods (temporal validation). This approach, while commonly used in medical imaging research due to practical constraints of multi-institutional data sharing, may not fully capture the variability encountered across different clinical sites, imaging protocols, and patient populations. Future multi-institutional external validation studies are warranted to confirm the generalizability of our findings.

Conclusions

In conclusion, our study strongly advocates for the use of group penalty models in radiomics to enhance the accuracy and reliability of predictive models in medical imaging. These models significantly improve analytical stability and provide a deeper understanding of complex biological and clinical interactions that traditional methods often overlook. By leveraging the inherent group structures within features, group penalty models enhance the interpretability of radiomic analyses. This improved interpretability is crucial in clinical settings, where it can profoundly influence diagnostic, prognostic, and therapeutic decisions. Integrating group penalty models into radiomics not only facilitates more robust and consistent feature selection but also aligns the analysis more closely with clinical reasoning, thereby advancing the field of personalized medicine.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1951/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1951/dss

Funding: This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (No. NRF-2021R1I1A1A01059893); and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2022R1A2C1091488).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1951/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Lambin P, Leijenaar RTH, Deist TM, Peerlings J, de Jong EEC, van Timmeren J, Sanduleanu S, Larue RTHM, Even AJG, Jochems A, van Wijk Y, Woodruff H, van Soest J, Lustberg T, Roelofs E, van Elmpt W, Dekker A, Mottaghy FM, Wildberger JE, Walsh S. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017;14:749-62. [Crossref] [PubMed]
2. Robert T. Regression shrinkage and selection via the lasso. J R Stat Soc Series B Stat Methodol 1996;58:267-88.
3. Huang Y, Liu Z, He L, Chen X, Pan D, Ma Z, Liang C, Tian J, Liang C. Radiomics Signature: A Potential Biomarker for the Prediction of Disease-Free Survival in Early-Stage (I or II) Non-Small Cell Lung Cancer. Radiology 2016;281:947-57. [Crossref] [PubMed]
4. Zheng BH, Liu LZ, Zhang ZZ, Shi JY, Dong LQ, Tian LY, Ding ZB, Ji Y, Rao SX, Zhou J, Fan J, Wang XY, Gao Q. Radiomics score: a potential prognostic imaging feature for postoperative survival of solitary HCC patients. BMC Cancer 2018;18:1148. [Crossref] [PubMed]
5. Xie T, Wang X, Li M, Tong T, Yu X, Zhou Z. Pancreatic ductal adenocarcinoma: a radiomics nomogram outperforms clinical model and TNM staging for survival estimation after curative resection. Eur Radiol 2020;30:2513-24. [Crossref] [PubMed]
6. Rahnenführer J, De Bin R, Benner A, Ambrogi F, Lusa L, Boulesteix AL, Migliavacca E, Binder H, Michiels S, Sauerbrei W, McShane L. for topic group “High-dimensional data” (TG9) of the STRATOS initiative. Statistical analysis of high-dimensional biomedical data: a gentle introduction to analytical goals, common approaches and challenges. BMC Med 2023;21:182. [Crossref] [PubMed]
7. Hastie T, Tibshirani R, Wainwright M. Statistical learning with sparsity: The Lasso and Generalizations. 1st edition. New York: Chapman and Hall/CRC; 2015.
8. Roberts S, Nowak G. Stabilizing the lasso against cross-validation variability. Computational Statistics & Data Analysis 2014;70:198-211.
9. Orton MR, Hann E, Doran SJ, Shepherd STC, Ap Dafydd D, Spencer CE, López JI, Albarrán-Artahona V, Comito F, Warren H, Shur J, Messiou C, Larkin J, Turajlic S, Koh DM. Interpretability of radiomics models is improved when using feature group selection strategies for predicting molecular and clinical targets in clear-cell renal cell carcinoma: insights from the TRACERx Renal study. Cancer Imaging 2023;23:76. [Crossref] [PubMed]
10. Yuan M, Lin Y. Model selection and estimation in regression with grouped variables. J R Stat Soc Ser B Stat Methodol 2006;68:49-67.
11. Breheny P, Huang J. Group descent algorithms for nonconvex penalized linear and logistic regression models with grouped predictors. Stat Comput 2015;25:173-87. [Crossref] [PubMed]
12. Velten B, Huber W. Adaptive penalization in high-dimensional regression and classification with external covariates using variational Bayes. Biostatistics 2021;22:348-64. [Crossref] [PubMed]
13. Huang J, Zhang T. The benefit of group sparsity. Ann Stat 2010;38:1978-2004.
14. Huang J, Breheny P, Ma S. A Selective Review of Group Selection in High-Dimensional Models. 2012;27:10.1214/12-STS392.
15. Suh YJ, Han K, Kwon Y, Kim H, Lee S, Hwang SH, Kim MH, Shin HJ, Lee CY, Shim HS. Computed Tomography Radiomics for Preoperative Prediction of Spread Through Air Spaces in the Early Stage of Surgically Resected Lung Adenocarcinomas. Yonsei Med J 2024;65:163-73. [Crossref] [PubMed]
16. Park YW, Oh J, You SC, Han K, Ahn SS, Choi YS, Chang JH, Kim SH, Lee SK. Radiomics and machine learning may accurately predict the grade and histological subtype in meningiomas using conventional and diffusion tensor imaging. Eur Radiol 2019;29:4068-76. [Crossref] [PubMed]
17. Lunardon N, Menardi G, Torelli N. ROSE: a package for binary imbalanced learning. R J 2014;6:79-89.
18. Borstelmann SMMachine Learning Principles for Radiology Investigators. Acad Radiol 2020;27:13-25. [Crossref] [PubMed]
19. Fan J, Li R. Variable selection via nonconcave penalized likelihood and its oracle properties. J Am Stat Assoc 2001;96:1348-60.
20. Zhang CH. Nearly unbiased variable selection under minimax concave penalty. Ann Stat 2010;38:894-942.
21. Barbieri MM, Berger JO. Optimal predictive model selection. Ann Stat 2004;32:870-97.
22. Obuchowski NA. Receiver operating characteristic curves and their use in radiology. Radiology 2003;229:3-8. [Crossref] [PubMed]
23. Khaire UM, Dhanalakshmi R. Stability of feature selection algorithm: A review. J King Saud Univ Comput Inf Sci 2022;34:1060-73.
24. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 1988;44:837-45.
25. Steyerberg EW, Vickers AJ, Cook NR, Gerds T, Gonen M, Obuchowski N, Pencina MJ, Kattan MW. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology 2010;21:128-38. [Crossref] [PubMed]
26. R Core Team. R: A language and environment for statistical computing. Vienna: R Foundation for Statistical Computing; 2013.
27. Friedman J, Hastie T, Tibshirani R. Regularization Paths for Generalized Linear Models via Coordinate Descent. J Stat Softw 2010;33:1-22.
28. Park JE, Park SY, Kim HJ, Kim HS. Reproducibility and Generalizability in Radiomics Modeling: Possible Strategies in Radiologic and Statistical Perspectives. Korean J Radiol 2019;20:1124-37. [Crossref] [PubMed]
29. van den Goorbergh R, van Smeden M, Timmerman D, Van Calster B. The harm of class imbalance corrections for risk prediction models: illustration and simulation using logistic regression. J Am Med Inform Assoc 2022;29:1525-34. [Crossref] [PubMed]
30. Riley RD, Ensor J, Snell KIE, Archer L, Whittle R, Dhiman P, Alderman J, Liu X, Kirton L, Manson-Whitton J, van Smeden M, Moons KG, Nirantharakumar K, Cazier JB, Denniston AK, Van Calster B, Collins GS. Importance of sample size on the quality and utility of AI-based prediction models for healthcare. Lancet Digit Health 2025;7:100857. [Crossref] [PubMed]