Construction of a multimodal artificial intelligence model for differentiating benign and malignant anterior...

Multimodal data fusion approaches not only capture the multidimensional characteristics of diseases but also compensate for the limitations of single-modality data in clinical diagnostics. These methods are increasingly utilized in clinical practice to significantly improve diagnostic accuracy. Artificial intelligence (AI), particularly in the domain of machine learning, has demonstrated a strong ability to extract meaningful features from large-scale datasets and is now widely applied in medical research. Differentiating benign from malignant anterior mediastinal tumors remains challenging in clinical settings. This study aimed to develop a multimodal AI model to accurately distinguish benign from malignant anterior mediastinal tumors.

We retrospectively analyzed patients with anterior mediastinal tumors treated at a single hospital between 2021 and 2025. Clinical data, laboratory test results, imaging studies, and final pathological diagnoses were collected. Radiomics features were extracted from chest computed tomography (CT) images and analyzed. Machine learning techniques were used to select key variables. Four predictive models were constructed: a model based solely on clinical variables, a model based solely on imaging features, a multimodal model integrating variables via logistic regression, and a multimodal model based on a random forest algorithm. The performance of each model was evaluated through receiver operating characteristic curve analysis and comprehensive performance metrics to identify the optimal model.

A total of 104 patients met the inclusion criteria, including 58 cases of benign and 46 cases of malignant anterior mediastinal tumors. Five clinical features and seven imaging features were identified as predictive variables. The logistic regression-based multimodal model achieved the best performance [area under the curve (AUC) =0.94], followed by the clinical-only model (AUC =0.899), the random forest-based fusion model (AUC =0.881), and the imaging-only model (AUC =0.844).

The logistic regression-based multimodal model demonstrated superior performance in differentiating benign and malignant anterior mediastinal tumors compared to models based on clinical data alone, imaging data alone, or the random forest approach. This method holds promise for aiding preoperative diagnosis in clinical settings. Notably, features such as tumor cross-sectional diameter, short-axis length on CT, history of myasthenia gravis, lactate dehydrogenase, and carcinoembryonic antigen showed significant clinical value in differentiating tumor types. These findings suggest that clinicians should pay closer attention to these indicators during differential diagnosis of anterior mediastinal tumors.