# Machine learning identifies immune-perinatal predictors of infantile hemangioma

## TL;DR

This study uses machine learning to identify immune and perinatal factors that predict infantile hemangioma, a common vascular tumor in infants.

## Contribution

A novel machine learning model using immune-perinatal features to predict infantile hemangioma risk is developed and validated.

## Key findings

- Prematurity, multiple gestation, and elevated VEGF, CRP, and SAA levels are key risk factors for infantile hemangioma.
- The XGBoost model achieved high predictive accuracy with AUCs of 0.952 in training and 0.870 in external validation.
- SHAP analysis revealed SAA, VEGF, and low birth weight as the most influential predictors of IH.

## Abstract

Infantile hemangioma (IH), the most common vascular tumor of infancy, exhibits hallmark features of immune and inflammatory dysregulation. While most cases are self-limiting, a subset progresses with potentially severe complications. Despite its benign classification, IH offers a unique model to investigate immune-mediated mechanisms in early tumorigenesis. However, risk stratification models incorporating immune-inflammatory markers remain underdeveloped.

A total of 1,466 infants and young children were enrolled, including 81 with IH. Comprehensive perinatal, clinical, and laboratory data were collected. Candidate risk factors were identified using logistic regression. Four machine learning algorithms—XGBoost, Random Forest, Support Vector Machine, and k-Nearest Neighbors—were employed to construct predictive models. Model performance was assessed through internal and external validation. SHapley Additive exPlanations (SHAP) were applied to interpret feature contributions and immune-inflammatory signatures.

Key risk factors included prematurity, multiple gestation, low birth weight, and elevated levels of VEGF, CRP, and SAA—markers linked to inflammation and immune activation. The XGBoost model achieved superior performance, with an AUC of 0.952 (training), 0.935 (internal validation), and 0.870 (external validation). SHAP analysis highlighted SAA, VEGF, and low birth weight as the most influential predictors, reflecting a critical link between innate immune dysregulation and IH development.

This study presents a robust, interpretable machine learning model that leverages immune-perinatal features to predict IH risk. Our findings support the notion that IH may serve as a paradigm for inflammation-associated vascular tumorigenesis, with implications for early detection and personalized intervention strategies in immune-driven neoplasms.