A Diagnostic Nomogram Integrating Preoperative Clinical Features and laboratory indicators for Predicting Pathological Subtypes of Subcutaneous Hemangioma: A Diagnostic Study (Preprint)

<sec> <title>BACKGROUND</title> Subcutaneous vascular tumors and vascular malformations frequently exhibit considerable overlaps in their clinical presentations and imaging characteristics, often necessitating histopathological examination for definitive differentiation. Accurate distinction between these two entities is crucial for determining appropriate clinical management strategies and for prognostic assessment of subcutaneous vascular lesions. </sec> <sec> <title>OBJECTIVE</title> This study aims to develop a clinical prediction model—presented as a nomogram—based on machine learning approaches to preoperatively predict the pathological subtype of subcutaneous vascular lesions. The model is intended to serve as an accurate and practical decision-support tool for clinicians, facilitating reliable differentiation among pathological subtypes and thereby optimizing diagnostic and therapeutic strategies to improve patient outcomes. </sec> <sec> <title>METHODS</title> We retrospectively analyzed 257 patients who underwent surgical resection of subcutaneous vascular lesions between June 2019 and August 2024, including 30 cases of vascular malformations and 227 cases of vascular tumors. Preoperative clinical characteristics and laboratory biochemical parameters were extracted from inpatient medical records. Missing values in continuous variables were imputed using mean substitution. The dataset was then randomly partitioned into training and validation sets at a 7:3 ratio. To mitigate class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied exclusively to the training set to prevent data leakage. Key predictive features were selected from the training set using Least Absolute Shrinkage and Selection Operator (LASSO) regression. Four machine learning models—logistic regression, random forest, support vector machine (SVM), and extreme gradient boosting (XGBoost)—were developed and compared based on their performance in the training set. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC), and calibration was evaluated via calibration plots. The validation set was used for final performance assessment. Given its favorable interpretability and robust predictive performance, the LASSO-derived logistic regression model was selected as the foundation for a nomogram to preoperatively estimate the probability of a lesion being classified as either a vascular tumor or a vascular malformation. Additionally, SHAP (Shapley Additive Explanations) analysis was performed on tree-based models (e.g., XGBoost) to quantify the contribution of individual features to the prediction outcomes. </sec> <sec> <title>RESULTS</title> A total of 257 patients with subcutaneous vascular anomalies were included in this study, comprising 227 cases of vascular tumors and 30 cases of vascular malformations. Six predictors were selected by LASSO regression for the pathological classification model: Ethnicity, Absolute Eosinophil Count, Indirect Bilirubin, Bicarbonate, Sodium, and Lactate Dehydrogenase Isoenzyme 1 (LDH-1).Among the four machine learning models evaluated on the validation set, the area under the receiver operating characteristic curve (AUC) ranged from 0.678 to 0.939. The random forest model achieved the highest performance, with an AUC of 0.939 (95% CI: 0.897–0.971), precision of 0.944, recall of 0.750, and an F1-score of 0.836. </sec> <sec> <title>CONCLUSIONS</title> This study developed and validated a high-accuracy, interpretable machine learning model for the non-invasive preoperative differentiation of subcutaneous vascular tumors from vascular malformations using only routinely available preoperative laboratory tests. Among four algorithms systematically evaluated, the random forest model demonstrated superior discriminative performance (AUC = 0.939) and calibration, and was selected as the optimal predictor. Relying on just six widely accessible hematological and biochemical variables—and enhanced by SHAP interpretability analysis—the model not only achieved high predictive accuracy but also highlighted key biomarkers such as LDH-1 and Absolute Eosinophil Count as discriminative features. This approach offers a low-cost, easily deployable decision-support tool, particularly valuable in resource-limited settings, with the potential to streamline preoperative classification and reduce reliance on advanced imaging or invasive biopsy. Future multicenter prospective studies are warranted to validate the model’s generalizability and facilitate its integration into clinical decision-support systems. </sec>