Sarcomas are malignant tumors of mesenchymal origin, covering soft tissue and bone. While they represent only about 1% of all adult cancers, they account for 8-20% of all cancers diagnosed in children under 15, making them a disproportionately significant problem in pediatric oncology. The biological heterogeneity of sarcoma, covering more than 100 histological subtypes ranging from osteosarcoma and Ewing sarcoma to synovial sarcoma and rhabdomyosarcoma, makes unified prognostic modeling particularly difficult.
The recurrence problem: Despite multimodal treatment combining surgery, chemotherapy, and radiotherapy, recurrence is common and carries a poor prognosis. Data from Soole et al. in synovial sarcoma showed a median time to first relapse of 24 months, with 73% of recurrences localized and 24.3% metastatic. Even with aggressive salvage approaches, including resection in 75.7% of cases, second-line chemotherapy in 73.0%, and radiotherapy in 48.6%, response rates were low (e.g., 36.4% for ifosfamide-based regimens). Five-year event-free survival was only 32.8%, and overall survival 42.1%. Identifying which patients will relapse early, before recurrence is clinically apparent, is therefore a priority.
Why classical statistics fall short: Traditional survival analyses and multivariate regression models require a minimum number of events relative to the number of candidate variables, a constraint that cannot be met for rare pediatric tumors at single institutions. Prior studies applying conventional statistical methods have produced inconsistent results, with some finding that standard prognostic factors (histological subtype, tumor size, surgical margins) are not statistically significant when sample sizes are small. Machine learning approaches specifically designed for high-dimensional, low-sample-size settings offer a principled alternative.
The authors of this 2025 paper, published in Children (MDPI), set out to apply a LASSO-regularized logistic regression framework to a micro-cohort of 23 pediatric sarcoma patients treated at a single tertiary-care center in Germany. They aimed to identify a parsimonious, biologically plausible risk signature for early recurrence, knowing that formal statistical power was limited and that findings would be hypothesis-generating rather than practice-changing. The study is notable for its transparent handling of these constraints and its use of modern interpretability tools.
The analytic dataset comprised 23 consecutive pediatric patients (all under 18 years) with histologically confirmed sarcoma treated at a single tertiary-care center, approved under ethics protocol 18-6469 of Ruhr University Bochum and conducted in compliance with the Declaration of Helsinki. The mean age at primary resection was 13.0 plus or minus 4.7 years (range 2-17 years, interquartile range 11.5-16.3 years). The median observational period was 76 months (approximately 6.3 years), generating roughly 158 patient-years of aggregate follow-up. Crucially, no patient died of disease- or treatment-related causes during this interval, meaning overall survival remained at 100% and first tumor recurrence was adopted as the primary endpoint.
Tumor spectrum: The cohort reflected the biological diversity characteristic of pediatric sarcoma. Synovial sarcoma was the leading subtype in 7 patients (30%), followed by dermatofibrosarcoma protuberans (DFSP) in 4 patients (17%), and epithelioid sarcoma in 3 patients (13%). Aggressive fibromatosis, alveolar rhabdomyosarcoma, and low-grade fibromyxoid sarcoma each appeared in 2 patients (9% each). Single-patient occurrences (4% each) included chondrosarcoma, dedifferentiated liposarcoma, fibrosarcoma, and an osteoblastoma-like osteosarcoma variant. Primary tumor widths ranged widely from 0.2 cm to 8.7 cm, with a median of 1.5 cm.
Recurrence rate and events-per-variable challenge: Recurrences were documented in 5 of 23 patients (22%), yielding an events-per-variable (EPV) ratio of approximately 1:9 across the 46 baseline covariates extracted per patient. These 46 variables spanned clinical demographics, histological characteristics, tumor metrics, comorbidities (including cardiovascular history, hypertension, and metabolic conditions), and treatment-related factors. An EPV of 1:9 is far below the traditional rule of thumb requiring at least 10 events per variable for stable logistic regression coefficients, making standard maximum-likelihood estimation highly prone to overfitting and separation.
The 46-variable feature set was preprocessed with median imputation for numeric variables and mode imputation for categorical variables. Numeric features were z-standardized (mean 0, standard deviation 1) and categorical variables underwent one-hot encoding with drop_first = True to avoid multicollinearity. A unified ColumnTransformer from scikit-learn 1.4 was applied strictly inside the cross-validation loop to prevent any information leakage from the preprocessing steps into held-out test folds.
The authors chose L1-penalized (LASSO) logistic regression as their primary modeling strategy, a well-justified decision for this rare-event setting. LASSO introduces a penalty term proportional to the absolute value of the model coefficients, shrinking small coefficients to exactly zero and thereby selecting a sparse subset of features. This automatic variable selection is valuable when the number of candidate predictors approaches or exceeds the number of events, preventing the unbounded coefficient inflation that afflicts ordinary logistic regression in such settings. The LASSO solver was set to "saga," a stochastic average gradient variant suitable for L1 regularization, with class weights balanced to account for the 5:18 event ratio.
Regularization path and hyperparameter selection: The regularization strength was governed by the inverse penalty parameter C, with C = 1.0 (corresponding to lambda = 1.0) selected as optimal via leave-one-out cross-validation (LOOCV) across a log-spaced grid from 10^-2 to 10^2 (100 values). LOOCV was chosen explicitly because it is the only unbiased cross-validation scheme available when n = 23: with fewer observations than a typical 5-fold CV would partition, leaving out one patient at a time ensures each prediction is made on a fully held-out case with the model trained on all 22 remaining observations.
Dimensionality reduction via PCA: To visualize the latent structure of the 46-variable feature space and identify atypical cases, principal component analysis (PCA) was applied to the full preprocessed design matrix with n_components = 2 using svd_solver = 'full'. The first two principal components explained 50% of total variance. Each patient's distance from the multivariate centroid in PC space was computed as a Euclidean distance, and the three patients with the largest distances were flagged as potential outliers for clinical chart review.
SHAP interpretability: SHAP (SHapley Additive exPlanations) values were computed using the TreeExplainer applied to the linear LASSO model, providing additive, patient-level explanations of how each variable contributed to the predicted recurrence probability. Features were ranked by mean absolute SHAP value across the cohort, offering an intuitive visualization of variable importance that goes beyond the raw LASSO coefficients. The entire analysis was implemented in Python 3.11 using pandas 2.2.1, scikit-learn 1.4.2, numpy 1.26.4, matplotlib 3.8.4, and shap 0.45.0 in a reproducible Jupyter environment.
Despite the severe statistical constraints imposed by the micro-cohort design, the LASSO model isolated four non-zero predictors of recurrence from the initial 46 candidates. This parsimony is a feature of L1 regularization, not a limitation: by shrinking irrelevant coefficients to zero, LASSO filters out variables whose apparent associations are driven by noise rather than signal, yielding a clinically interpretable, four-variable signature.
Histological grade and tumor width: These were the two dominant predictors. Each incremental increase in histological grade was associated with an odds ratio (OR) of approximately 2.18 (beta coefficient approximately +0.78), meaning each upward step in tumor grade roughly doubled the odds of recurrence. Similarly, each additional centimeter of primary tumor width carried an OR of approximately 2.04 (beta approximately +0.71). These two variables dominated the SHAP importance ranking and align directly with established oncological principles: higher-grade tumors exhibit increased proliferative capacity and metastatic potential, while larger tumors are more likely to harbor microscopic spread and present greater surgical challenges for achieving R0 resection. Both are also central inputs to the Sarculator, a widely used prognostic nomogram for soft tissue sarcoma.
Arterial hypertension: The most novel finding was the inclusion of arterial hypertension as the third non-zero predictor, carrying an OR of 1.70. Hypertension was uncommon in this pediatric cohort but consistently co-occurred with recurrence across modeling strategies. The authors propose several mechanistic hypotheses: elevated VEGF expression (a known driver of sarcoma angiogenesis) also causes systemic hypertension through capillary rarefaction and increased vascular tone. HIF-1 alpha, frequently upregulated in sarcomas under hypoxic conditions, is independently linked to both VEGF expression and systemic hypertension. Whether hypertension in this context reflects a true biological driver, a surrogate of tumor aggressiveness, or an iatrogenic effect of corticosteroids or pain management remains unresolved.
Extremity localization: The fourth predictor was tumor localization to the extremities, associated with an OR of 1.90. This echoes prior reports in the pediatric sarcoma literature documenting local recurrence rates as high as 25% in limb sarcomas versus 13.6% for trunk wall and head/neck tumors. The authors interpret extremity localization as a surrogate marker of surgical complexity rather than a biological driver, reflecting the anatomical constraints that make achieving clear resection margins more difficult in limb-compartment tumors. The SHAP analysis confirmed that these four variables dominated the prediction landscape, with all other 42 candidate variables shrunk to zero coefficients.
The LASSO model achieved a balanced accuracy of 0.61 plus or minus 0.08 and an AUROC of 0.47 plus or minus 0.12 under LOOCV. These figures require careful contextualization. Balanced accuracy of 0.61 means the model correctly classified recurrence and non-recurrence cases 61% of the time after accounting for the 5:18 class imbalance, exceeding the 50% baseline of random guessing. However, the AUROC of approximately 0.47 indicates near-chance discriminative ability: when a randomly chosen child who recurred is paired with one who did not, the model assigns the higher predicted probability to the correct patient only slightly more than half the time.
Understanding the AUROC/balanced accuracy discrepancy: The divergence between these two metrics is not unusual in micro-datasets and is instructive. Balanced accuracy is a threshold-dependent metric, meaning it measures classification performance at a specific probability cutoff, while AUROC is threshold-free and measures the model's overall ability to rank patients by risk. In a dataset with only 4-5 positive events, a single misclassified case can shift the entire ROC curve substantially. The wide standard deviations (plus or minus 0.08 and plus or minus 0.12, respectively) directly reflect this sensitivity to individual LOOCV folds. The authors correctly characterize the model as a "hypothesis generator" rather than a clinically deployable risk score.
Calibration: Model calibration was assessed using the Hosmer-Lemeshow test across four equal-sized risk quartiles. The result was a chi-square statistic of 1.9 (df = 2, p = 0.39), indicating no evidence of lack of fit. A calibration plot comparing observed and expected recurrence rates by risk quartile showed close agreement, suggesting the model's predicted probabilities are not systematically overconfident or underconfident. Good calibration at modest discrimination is a meaningful finding: the model is honest about uncertainty even if it cannot reliably separate individual patients.
Sensitivity analysis via Firth logistic regression: A bias-reduced (Firth-penalized) logistic regression was applied as a sensitivity check to confirm that the four identified predictors were not artifacts of small-sample complete separation, a well-known problem in logistic regression where a predictor perfectly separates outcomes in small datasets. The Firth estimates showed less than 5% deviation in odds ratios and identical directionality for all four predictors (grade: LASSO OR 2.18 vs. Firth OR 2.25; width: 2.04 vs. 2.10; hypertension: 1.70 vs. 1.72; extremity: 1.90 vs. 1.88). This concordance provides meaningful evidence that the signal is not driven by statistical instability.
Beyond the predictive model, the authors applied PCA to the full 46-variable preprocessed design matrix to visualize the latent structure of the patient cohort. The first two principal components collectively explained 50% of total variance, providing a two-dimensional map of how patients cluster based on their baseline clinical and histological profiles. This dimensionality reduction technique is particularly valuable in rare disease settings where sample sizes are too small for reliable cluster analysis or unsupervised classification.
Outlier identification: Three patients emerged as conspicuous outliers, occupying the periphery of the PC space with Euclidean distances substantially larger than the rest of the cohort. The authors flagged these three patients a priori for clinical chart review before examining their outcomes. Two of these outliers encapsulated the full "high-risk triad," namely high histological grade, large primary width, and documented arterial hypertension, and both experienced early recurrence. This post-hoc validation of the outlier-identification approach lends biological plausibility to the PCA framework as a complementary screening tool.
The third outlier: The third outlier presented an instructive counterexample: this patient harbored the smallest recorded tumor volume in the cohort and remained disease-free throughout the entire observation period. The authors raise two possibilities: data mis-entry (which cannot be ruled out in a retrospective single-center study) or exceptional therapeutic response. The existence of such an outlier serves as a cautionary note about over-reliance on PCA distance as a proxy for clinical risk without careful chart review.
In the context of rare pediatric sarcoma, PCA outlier analysis may represent a practical tool for generating hypotheses about atypical patient phenotypes that warrant intensified surveillance or treatment modification. The approach requires no additional biological specimens or molecular assays, relying only on routinely collected clinical data. Its value is not in definitive risk stratification but in flagging cases whose clinical profile deviates enough from the cohort norm to merit closer attention from the multidisciplinary team.
Sample size and statistical power: The most fundamental limitation is the n = 23 sample with only 5 recurrence events. At this scale, statistical power is inadequate for formal hypothesis testing, and confidence intervals around all reported metrics (balanced accuracy, AUROC, individual odds ratios) are necessarily wide. The LASSO regularization addresses overfitting but cannot compensate for genuine information deficiency: with so few events, the model cannot learn complex, non-linear decision boundaries or reliably detect interactions between predictors. Every identified association must be regarded as a signal hypothesis, not a validated finding.
Retrospective single-center design: The retrospective cohort design introduces selection bias, as only patients who were treated at this institution and met the inclusion criteria are represented. Treatment protocols, chemotherapy regimens, radiation dosing, and surgical approaches may differ significantly between institutions, between eras, and between treating surgeons. Pathological grading, in particular, is subject to interobserver variability even within a single institution. The single-center setting also limits the diversity of patient demographics, comorbidity profiles, and sarcoma subtypes represented.
Hypertension measurement limitations: The novel finding of arterial hypertension as a predictor requires cautious interpretation beyond its biological plausibility. Blood pressure elevation in pediatric patients can be situational, driven by white-coat effect, procedural anxiety, acute pain, or steroid administration during treatment. The study did not differentiate between established chronic hypertension and isolated blood pressure elevations recorded in the chart. VEGF, HIF-1 alpha, and other candidate molecular mediators of the hypertension-sarcoma link were not measured, leaving the mechanistic hypothesis entirely speculative.
Generalizability: Even if a larger prospective study confirmed the four-variable signature, its clinical utility would depend on whether it adds prognostic information above and beyond what experienced sarcoma clinicians already incorporate into their management decisions. Histological grade and tumor size are already standard inputs to tools like the Sarculator. Whether hypertension monitoring adds independent, actionable information requires prospective testing against validated risk tools in a well-characterized multicenter cohort, ideally with molecular profiling to dissect the biological mechanism.
Multicenter prospective validation: The immediate next step is external validation of the four-variable signature in independent pediatric sarcoma cohorts at multiple institutions. A prospective study embedding the four predictors (grade, width, hypertension status, extremity localization) into a structured data collection protocol would allow prospective testing against established tools and clinical endpoints. Multicenter participation is essential not only for statistical power but also to assess whether the signature generalizes across treatment protocols, patient demographics, and pathological grading practices. The authors explicitly call for systematic blood pressure surveillance to be incorporated into forthcoming pediatric sarcoma trials.
Molecular profiling integration: To test the hypertension hypothesis mechanistically, future studies should systematically collect VEGF expression levels, HIF-1 alpha immunohistochemistry or mRNA quantification, and microvessel density measurements alongside blood pressure data. Incorporating multi-omics data (RNA sequencing, copy number variation, epigenomic profiling) into LASSO or other regularized regression frameworks could yield a richer risk signature that captures both clinical and molecular heterogeneity. Pediatric sarcoma lends itself to this approach because institutional biobanking for rare tumors is already common at major sarcoma centers.
Expanded machine learning architectures: With larger cohorts, gradient-boosted tree ensembles (XGBoost, LightGBM), random forests, and survival analysis extensions (such as Cox-LASSO or random survival forests) would allow capture of non-linear predictor relationships and time-to-recurrence modeling rather than binary classification at a fixed endpoint. Deep learning approaches on histopathology whole-slide images have shown promise for grade and subtype classification in adult sarcoma and could be extended to pediatric cohorts as annotated digital pathology datasets expand through consortia such as the Children's Oncology Group (COG).
Federated learning for rare pediatric tumors: The fundamental bottleneck in pediatric sarcoma AI research is the rarity of individual subtypes at any single institution. Federated learning frameworks, where institutions train local model updates on their own data and share only gradients rather than raw patient records, represent a privacy-preserving route to assembling the multi-hundred-patient datasets needed for robust predictive models across sarcoma subtypes. International consortia building on existing pediatric oncology cooperative group infrastructure are well positioned to operationalize this approach.