Ridge-Cox beats Lasso-Cox and deep learning in landmark radiomics survival benchmark

Radiomics promises to turn ordinary CT scans into quantitative crystal balls. By extracting hundreds of shape, texture, and wavelet features from tumour images, researchers hope to predict which cancer patients will survive longer without ever touching a biopsy needle. But a nagging question has haunted the field: with so many survival prediction methods and feature engineering strategies available, which combination actually works? A new benchmark published in Heliyon by Wei Liu and Wei Wang delivers one of the most comprehensive answers yet, and its conclusions are turning conventional wisdom on its head.

The stakes are enormous. Radiomics-based survival models have shown prognostic value in glioblastoma, lung cancer, ovarian cancer, and other malignancies, yet quantitative imaging features suffer from the curse of dimensionality and from uncertainties introduced by different scanners, imaging protocols, organ motion, and manual tumour delineation by different radiation oncologists. Compared with genomic features, radiomic features exhibit substantially lower stability across datasets. Feature selection, which filters out non-informative and volatile features, or feature extraction, which constructs new informative features by integrating original ones, therefore becomes a necessary strategy for radiomics risk modelling. The problem is that the optimal pairing of survival models with feature engineering methods has remained unclear.

Previous benchmarking efforts had important blind spots. Parmar and colleagues, who first compared machine learning methods for radiomic features, dichotomized continuous survival outcomes into two categories rather than directly modelling time-to-event data while accounting for censoring. Such dichotomization discards valuable time-to-event information, introduces arbitrary thresholds, and reduces statistical power. Leger and colleagues later evaluated 11 survival prediction methods with 12 feature selection methods, but relied on a single dataset pair and ignored feature extraction entirely. Deep learning survival models, which relax the proportional hazards and linear assumptions, had been applied to radiomics with wildly varying architectures, from one to five hidden layers and three to 100 nodes per layer, reflecting a lack of consensus about appropriate network complexity.

To close these gaps, Liu and Wang assembled three groups of publicly available radiomic datasets representing different cancer types, feature sets, and levels of feature stability. Group I contained 224 non-small cell lung cancer patients from the Moffitt Cancer Center with 636 features, validated on 88 patients from the MAASTRO Clinic. Group II contained 213 epithelial ovarian cancer patients from Hammersmith Hospital with 657 features, validated on 70 patients from The Cancer Genome Atlas. Group III contained 422 NSCLC patients from MAASTRO with 1561 features extracted using PyRadiomics, validated on 143 patients from a Stanford radiogenomics cohort. Crucially, Group III also included a test-retest dataset of 28 patients scanned 15 minutes apart and an inter-observer dataset of repeated delineations by five radiation oncologists, allowing the researchers to filter for 780 features with stability correlations above 0.9.

The benchmark then pitted 16 survival prediction methods against 12 feature selection and extraction methods, evaluated with three complementary metrics: the concordance index, which measures the proportion of patient pairs whose predictions and outcomes agree; the integrated area under the curve, which averages time-dependent discrimination over the follow-up period; and the integrated Brier score, which quantifies the accuracy of predicted survival probabilities using inverse probability of censoring weighting. Each discovery dataset was randomly split into training and test sets, models were fitted exclusively on training data to avoid leakage, and the trained models were applied unchanged to both test and external validation datasets. This procedure was repeated 100 times, yielding performance estimates and stability measures via the relative standard deviation, with overall rankings derived through rank aggregation across all datasets and metrics.

The headline finding is a shock to the field’s default settings. Lasso-Cox, the most widely used radiomics survival approach because it integrates feature selection with Cox regression, ranked only seventh overall. Ridge-Cox, which retains all features while controlling overfitting through L2 regularization rather than enforcing sparsity, ranked first in both prediction accuracy and stability according to the C-index and IAUC. The authors argue that this common preference for Lasso-Cox may overlook an effective strategy: unlike genomics, where measuring features is costly, radiomic features can be extracted automatically within seconds at negligible cost, so retaining informative features through regularization may provide superior predictive performance when accurate prognosis is the primary objective. Boosting methods glmboost and CoxBoost ranked second and third by C-index and IAUC, while survivalSVM consistently achieved favourable performance and stability regardless of metric.

Among feature engineering methods, supervised principal component analysis claimed the top overall rank, proving particularly advantageous for datasets with relatively unstable radiomic features. SPCA integrates multiple weak signals into informative meta-features and was associated with improved predictive performance and model stability. Hierarchical clustering performed best on the NSCLC2 dataset, which contained a large number of highly discriminative and stable features, by clustering correlated features to generate less redundant representations. Among pure feature selection methods, the simple Cox p-value method achieved the best balance between predictive accuracy and stability, followed by random survival forest variable importance. Notably, applying no feature selection at all also ranked highly, because several survival methods already perform embedded feature selection, making an additional step potentially wasteful.

The deep learning results challenge another widespread expectation. The five neural network methods, DeepSurv, DeepHit, CoxTime, LogHazard, and PCHazard, generally achieved lower C-index and IAUC values with larger variability than traditional models, indicating weaker risk ranking on radiomic data. Yet when evaluated by the integrated Brier score, the deep learning methods outperformed conventional survival models, with CoxTime ranking first on two datasets. The discrepancy reflects different optimization objectives: the C-index and IAUC primarily evaluate patient risk ranking, whereas the Brier score emphasizes accuracy of predicted survival probabilities, and models like DeepHit and PCHazard explicitly estimate survival probabilities at each time point without the proportional hazards constraint. A two-way analysis of variance across 180 optimized models revealed that the number of hidden layers had the greatest influence on performance, with accuracy declining as depth increased but improving as nodes per layer increased. The sweet spot: one to two hidden layers with 32 to 128 nodes per layer.

Feature stability emerged as the hidden variable governing everything else. The median feature RSD decreased from 63.91 in NSCLC1 to 61.48 in OV and further to 21.91 in NSCLC2, and median C-index values rose in parallel from 0.520 to 0.567 to 0.577. Datasets with high censoring ratios and small sample sizes produced less stable feature-survival associations, which in turn destabilized prediction models. Training failure rates told the same story: the abess method failed in roughly 90 percent of runs on the unstable NSCLC1 training set, while the spike-and-slab lasso Cox model, sslasso, trained successfully almost every time even on unstable data, ranking fourth overall and offering a promising balance of performance, interpretability, and robustness. Applying SPCA before model construction substantially reduced failure rates across the board.

For a field racing toward clinical deployment, the practical guidance is clear. When the goal is accurate risk ranking or patient stratification, traditional methods such as Ridge-Cox, glmboost, and CoxBoost are preferable. When accurate estimation of survival probabilities over time matters most, deep learning methods such as CoxTime, DeepHit, and PCHazard offer advantages. Before choosing any model, the authors recommend evaluating the discriminative ability and stability of radiomic features in a new dataset, since methods that performed well on datasets with similar feature characteristics represent suitable candidates. Rather than relying on a single best-performing method, multiple top-ranked combinations should be evaluated and externally validated. The principles may extend beyond radiomics to other high-dimensional omics survival analyses, and with all data and code publicly available on GitHub, the benchmark invites the community to stress-test its conclusions at scale.

Subject of Research: Radiomics survival prediction benchmark reveals Ridge-Cox advantage and deep learning limits

Article Title: A comparative study of survival prediction and feature selection/extraction methods for quantitative radiomics risk modelling

Article References: Liu, W., & Wang, W. (2026). A comparative study of survival prediction and feature selection/extraction methods for quantitative radiomics risk modelling. Heliyon, 12(16), Article e45602. https://doi.org/10.1016/j.heliyon.2026.e45602

Image Credits: AI Generated

DOI: 10.1016/j.heliyon.2026.e45602

Keywords: radiomics, survival prediction, machine learning, deep learning, cancer prognosis, feature selection, benchmarking, medical imaging, comparative, survival, prediction, feature