An intelligent prediction method for osteosarcoma metastasis based on a multi-modal fusion model

Osteosarcoma (OS) is one of the most common malignant bone tumors in adolescents, and distant metastasis is a major factor affecting patient prognosis. To improve the accuracy of OS metastasis risk prediction, this study proposes an intelligent prediction method based on a multi-modal fusion model. This method fully integrates multi-omics data, including g61ene expression, DNA methylation, and gene mutation, and employs a multi-modal variational autoencoder (VAE) for feature extraction and dimensionality reduction. This effectively removes data noise and learns latent representations. The extracted features are concatenated through a modality fusion module and input into a transformer model for deep feature interaction and learning. The multi-head attention mechanism of the transformer captures the complex relationships between different modal features and highlights critical features, thereby enhancing the predictive power of the model. Finally, the fused features are classified through a fully connected neural network (FCNN), achieving precise prediction of OS metastasis risk. Experimental results on the therapeutically applicable research to generate effective treatments (TARGET) dataset demonstrate that the proposed model outperforms traditional single-modal prediction methods in terms of AUC, F1-score, and other metrics. The combination of multi-modal VAE and transformer effectively addresses issues such as high dimensionality, noise, and limited sample sizes in omics data, significantly improving the robustness and generalizability of the model. Furthermore, the attention mechanism provides an intuitive visualization of molecular features critical to metastasis, offering valuable insights for clinical diagnosis and personalized treatment of OS.