Soft tissue sarcomas are a heterogeneous group of rare malignant tumors arising from mesenchymal tissues, including fat, muscle, nerves, fibrous tissue, and blood vessels. They account for roughly 1% of all adult malignancies. Despite this low prevalence, they are clinically demanding: surgical resection is the primary treatment modality and is often extensive, carrying significant risk for perioperative complications. Among the most consequential of these complications is wound infection, which can impair wound healing, prolong hospitalization, compromise oncological outcomes, and drive up healthcare costs.
The rarity of soft tissue sarcomas creates a fundamental research problem. No single institution accumulates enough patients to conduct statistically robust studies of wound-related complications. Prior investigations have been limited in scale, often retrospective single-center series that lack the power to isolate individual risk factors. This means clinicians have had to rely on extrapolated data from more common malignancies or general surgical literature to inform perioperative decisions for sarcoma patients.
The HIRA database: To overcome this barrier, the authors turned to the Health Insurance Review and Assessment (HIRA) Service Hub of the Republic of Korea, a nationwide administrative database covering approximately 50 million individuals. This database captures demographic data, diagnoses coded by the International Classification of Diseases (ICD-10), prescribed medications, and surgical procedure records for essentially the entire Korean population. It represents one of the most comprehensive national health datasets in the world for this kind of analysis.
The study also addresses a specific concern in big data analysis: when sample sizes are very large, traditional statistical significance tests become hypersensitive and will flag trivially small differences as statistically significant. The authors therefore deliberately chose a machine learning framework paired with explainability tools to understand which risk factors truly matter in a clinically meaningful sense, not just whether a p-value crosses an arbitrary threshold.
The study enrolled patients who underwent wide excision of soft tissue sarcomas between 2010 and 2021. Patient identification relied on ICD-10 code C49 (malignant neoplasm of other connective and soft tissue), paired with specific Anatomical Therapeutic Chemical (ATC) procedure codes for wide resection. Wound infections were defined as post-operative infections requiring a return to the operating room within the post-surgical window, a clinically stringent definition that excluded superficial infections managed conservatively. A total of 10,969 patients met inclusion criteria, making this by far the largest sarcoma wound infection cohort in the published literature.
The study was approved by the Korea University Anam Hospital Institutional Review Board (approval code 2022AN0378, August 2022), and a waiver of informed consent was granted given the retrospective, anonymized nature of the dataset. Each patient in the HIRA system carries a unique Korean Resident Registration Number, making deduplication automatic and complete. This structural feature of the database effectively eliminates one of the most common sources of error in administrative data studies.
Random Forest classifier: The machine learning method chosen was the random forest algorithm, an ensemble technique based on constructing multiple decision trees and aggregating their outputs through bootstrap aggregation (bagging). Each tree within the forest is built on a randomly sampled subset of both patients and features. Random forests are well-suited for administrative health data because they handle mixed data types (binary comorbidities, continuous age, ordinal socioeconomic status), are robust to overfitting relative to single decision trees, and natively provide feature importance estimates.
SHAP values as the explainability layer: After training the random forest, the authors applied Shapley Additive Explanations (SHAP) to interpret the model output. SHAP values are grounded in cooperative game theory: each feature receives a contribution value (positive or negative) based on how much it shifts the model's predicted probability of wound infection relative to the average prediction. This approach provides both global feature importance (which features matter most across all patients) and local interpretability (how each feature contributed to an individual patient's prediction).
Of the 10,969 patients who underwent wide sarcoma resection, 886 (8.08%) developed post-operative infections that required a return to the operating room. The overall perioperative blood transfusion rate was 20.67% (2,267 patients). Among infected patients specifically, the transfusion rate was 36.5%, compared to 19.3% in non-infected patients, a statistically significant difference (p less than 0.05). This raw comparison already suggests that transfusion is a meaningful risk factor, but the SHAP analysis quantifies this more rigorously.
The SHAP summary plot ranked features by their mean absolute SHAP value across all patients. Perioperative blood transfusion emerged as the single strongest predictor of wound infection. In the SHAP framework, a positive SHAP value indicates that the feature pushes the model prediction toward higher wound infection risk; a negative value pushes it toward lower risk. Patients who received transfusions (coded as 1) consistently displayed positive SHAP values, while non-transfused patients (coded as 0) showed negative SHAP values, a divergence that graphically captures transfusion's contribution to infection risk.
Other top-ranked features: Following transfusion, the highest mean SHAP values were observed for male sex, older age, and lower socioeconomic status (SES). Specific comorbidities, including liver disease (prevalence 32.7% overall, 39.6% in infected patients, p less than 0.05), diabetes mellitus (27.8% overall, 34.9% in infected patients, p less than 0.05), iron deficiency anemia (7.3% overall, 11.6% in infected patients, p less than 0.05), and peripheral vascular disease also contributed positively to infection risk predictions, though with lower SHAP importance than the sociodemographic variables.
Model performance: The area under the ROC curve (AUC) for the random forest classifier was 0.6422. The authors acknowledge openly that this represents modest classification performance. In practical terms, an AUC of 0.64 means the model correctly ranks a randomly chosen infected patient above a randomly chosen non-infected patient only about 64% of the time. For a study of this nature, the predictive accuracy of the model matters less than the interpretability of which factors drive predictions, and the SHAP framework serves that purpose regardless of overall AUC.
Beyond the global summary ranking, the authors generated SHAP dependence plots for each major predictor. These plots graph each patient's feature value on the x-axis against their individual SHAP value for that feature on the y-axis, with a color scale typically overlaid to show interaction with a second variable. Dependence plots reveal non-linear relationships and interactions that a simple regression coefficient or odds ratio would miss.
Transfusion dependence plot: Patients without a transfusion (x-axis value 0.0) clustered with negative SHAP values, indicating they were at reduced infection risk relative to the population average. Patients who received a transfusion (x-axis value 1.0) clustered with positive SHAP values. Notably, the interaction color overlay revealed that transfused patients tended to be older, suggesting that older patients both receive transfusions at higher rates and face compounding wound infection risk from both factors simultaneously.
Sex dependence plot: Males (coded as 0) displayed higher SHAP values than females (coded as 1), indicating that being male is associated with increased infection risk. This aligns with a broader body of literature documenting sex-based differences in immune response, sepsis epidemiology, and post-surgical infection rates. Multiple investigations have shown that males face higher rates of bacteremia and sepsis-related complications, attributed to differences in sex hormone profiles and a greater background prevalence of malnutrition and comorbidities in older male cohorts.
Age dependence plot: SHAP values for wound infection increased progressively with patient age. The interaction color overlay on the age dependence plot additionally showed that a disproportionately large fraction of very young patients (under age 20) received perioperative transfusions. This is consistent with the clinical reality that pediatric sarcoma surgeries often involve larger relative blood volumes. At every given age, transfused patients displayed higher SHAP values than non-transfused patients of the same age, confirming that transfusion contributes to infection risk independently of age itself.
Socioeconomic status plot: SHAP values for wound infection were highest in patients with the lowest SES codes. Lower SES is associated with greater comorbidity burden, higher tobacco use prevalence, dietary deficiencies, and diminished access to post-operative follow-up care, all of which impair wound healing. The SES dependence plot shows this as a monotonic increase in SHAP contribution as SES decreases, a gradient that is clinically consistent with known disparities literature.
The association between perioperative red blood cell transfusion and post-operative infection is not new to the surgical literature, but the biological mechanisms remain incompletely understood. In the context of soft tissue sarcomas, this is the first large-scale study to quantify that association explicitly. The authors discuss several candidate mechanisms through which allogeneic blood transfusion may predispose patients to wound infections.
Transfusion-related immunomodulation (TRIM): Allogeneic blood components carry donor leukocytes, foreign antigens, and bioactive lipids that accumulate during storage. These components can suppress recipient T-cell proliferation, natural killer cell activity, and macrophage function. The resulting immunosuppressed state is sometimes called transfusion-related immunomodulation and has been proposed as a mechanism for increased infection rates, impaired tumor surveillance, and potentially worse long-term oncological outcomes in cancer patients.
Iron loading and bacterial growth: Red blood cell transfusions also deliver supplemental iron, both intracellularly within donor erythrocytes and as free hemoglobin released during storage-related hemolysis. Iron is an essential growth factor for many bacteria, and elevated serum iron levels following transfusion may facilitate bacterial proliferation at surgical wound sites. This is consistent with the observation in the current study's demographic table that iron prescription (a proxy for iron deficiency anemia) was itself independently associated with higher wound infection rates: 11.6% in infected patients vs. 7.0% without infection (p less than 0.05).
The authors note that in the context of a global blood shortage that intensified during and after the COVID-19 pandemic, understanding when transfusions are truly necessary versus potentially avoidable has renewed clinical urgency. If transfusion independently raises wound infection risk in sarcoma patients, and if blood availability is constrained, this data supports more conservative transfusion thresholds and expanded use of autologous blood salvage or preoperative erythropoiesis-stimulating agents where clinically feasible.
The SHAP analysis revealed that after transfusion, the largest contributors to wound infection risk were sociodemographic rather than disease-specific. Male sex, older age, and lower socioeconomic status outranked individual comorbidities such as diabetes or liver disease in terms of mean SHAP contribution. The authors interpret this as a reflection of the systemic nature of wound healing impairment: these broad-category factors encompass a constellation of physiological and behavioral variables that no single comorbidity can capture.
Male sex: Multiple epidemiological studies have documented higher sepsis and surgical site infection rates in males. Proposed explanations include sex hormone differences, with testosterone associated with immunosuppression and estrogen with enhanced innate immunity, as well as higher background rates of malnutrition in older male populations and greater prevalence of pressure ulcers and peripheral vascular comorbidities. The SHAP finding that male sex is the second highest risk predictor for wound infection in soft tissue sarcoma is clinically actionable: male patients undergoing wide resection may benefit from more intensive perioperative wound surveillance protocols.
Age: The progressive increase in SHAP values with age is physiologically expected. Older patients carry greater burdens of underlying disease, including diabetes, liver disease, peripheral vascular disease, and nutritional deficiencies, all of which impair wound healing through distinct pathways. Reduced skin perfusion, impaired collagen synthesis, blunted neutrophil and macrophage responses, and immunosenescence each contribute. The age dependence plot's interaction term further demonstrated that older patients who also received transfusions faced the highest infection risk of any subgroup in the analysis.
Socioeconomic status: The literature consistently links low SES to higher wound complication rates across surgical specialties. Contributing factors include higher prevalence of tobacco use, obesity, dietary inadequacy, and difficulty accessing post-discharge wound care. In the Korean HIRA context, SES is captured through insurance premium categories, which serve as a validated proxy for income level. The SHAP finding that low SES patients carry systematically elevated wound infection risk echoes a global pattern of healthcare disparity and argues for targeting additional resources toward these patients in the perioperative period.
The authors are candid about several concrete limitations that qualify the study's conclusions. Most fundamentally, this is a retrospective cohort study based entirely on administrative billing codes rather than clinical records. While the HIRA database's retrospective accuracy for diagnosis coding has been validated at 70 to 90%, meaning up to 30% of diagnoses could be miscoded or missing, this level of coding error introduces systematic noise into all comorbidity variables.
Missing tumor characteristics: Perhaps the most significant limitation is that HIRA does not capture tumor-specific variables. The database contains no information on soft tissue sarcoma histologic subtype, tumor size, anatomical location, surgical margin status, resection complexity, use of adjuvant radiation therapy, or chemotherapy regimens. All of these factors are known to influence perioperative wound outcomes in sarcoma patients. Large tumors requiring complex flap reconstruction, irradiated tissue, or extensive anatomical dead space carry inherently higher wound complication rates. Because these variables are absent, the model's residual unexplained variance may substantially reflect tumor biology and surgical complexity rather than unmeasured patient factors.
Confounding by indication: Transfusion is not randomly administered. Patients who received perioperative transfusion are systematically different from those who did not: they had larger tumors, more extensive surgery, greater intraoperative blood loss, and more baseline anemia. These differences are the reason for transfusion, and they are themselves risk factors for wound infection independently of transfusion. Without controlling for these indications, the observed transfusion SHAP values likely overestimate the independent causal effect of transfusion on wound infection.
Model performance: The AUC of 0.6422 represents only modest discrimination. In a clinical decision-support context, this would be considered inadequate for individual-level risk prediction. The model's primary contribution is therefore feature attribution (understanding which variables matter) rather than actual risk scoring. Future work incorporating tumor-level variables from hospital records would be expected to substantially improve both AUC and the precision of SHAP attribution.
Despite its limitations, this study makes a genuine methodological and substantive contribution to sarcoma perioperative care research. It is the first retrospective cohort study of wound infection in soft tissue sarcomas conducted using a truly nationwide dataset, and at 10,969 patients it dwarfs any prior single-institution or multi-institution series on this topic. The combination of random forest modeling with SHAP explainability represents a step forward from purely descriptive administrative data analyses, providing a replicable framework for other rare cancer research groups facing similar data constraints.
The SHAP framework as a model for rare cancer big data: The authors emphasize that as administrative health databases grow in scale and coverage globally, SHAP-based explainability will become increasingly important for extracting actionable knowledge from large-scale datasets where traditional statistical methods become unreliable due to statistical power inflation. The SHAP approach sidesteps the p-value inflation problem inherent in large-N studies by focusing on magnitude of contribution rather than statistical significance.
Clinical implications for current practice: The study suggests three practice-relevant conclusions. First, perioperative transfusion should be subject to conservative thresholds in sarcoma surgery where feasible, given its association with wound infection risk. Second, male patients, elderly patients, and patients with low SES undergoing wide resection should be flagged for intensified wound surveillance and early intervention protocols. Third, the comorbidity profiles, particularly liver disease (32.7% prevalence), diabetes (27.8%), and peripheral vascular disease, should prompt proactive optimization before elective wide resection.
Future research priorities: The authors specifically identify hospital-based data integration as the next step. Linking HIRA administrative records to institutional tumor registries would add histologic subtype, tumor size, surgical approach, radiation history, and surgical margin data. These additions would address the most significant gap in the current analysis. The authors also suggest that future studies could examine post-operative MRI surveillance patterns and their relationship to wound complication detection timing, given that MRI is the primary imaging modality for soft tissue sarcoma follow-up and provides detailed soft tissue edema and seroma characterization.