An approach with Visual and Tabular Mamba to multimodal medical data using Mixed Fusion

Researchers propose a Mamba-based architecture for multimodal medical data fusion that combines visual and tabular processing to improve cancer classification interpretability. Testing on skin and oral cancer datasets shows competitive performance with enhanced explainability through SHAP analysis, positioning state space models as viable alternatives to Transformers in medical AI applications.

This research demonstrates a meaningful advancement in medical AI by leveraging Mamba, a relatively new state space model architecture, for multimodal data integration in oncology. The mixed fusion approach represents a pragmatic solution to a persistent challenge in medical AI: balancing predictive accuracy with clinical interpretability. By processing imaging data and clinical metadata through separate specialized pathways before final fusion, the architecture mirrors how clinicians naturally synthesize information.

The broader context reflects a significant industry shift away from Transformer-dominated approaches. While Transformers revolutionized AI, they face computational constraints and interpretability limitations that concern medical practitioners. Mamba's state space modeling offers reduced computational overhead and cleaner decision pathways, making it attractive for resource-constrained healthcare settings. This work validates that alternative architectures can compete with established methods.

For healthcare AI developers and hospital systems evaluating classification tools, these findings suggest Mamba-based approaches merit serious consideration, particularly where sensitivity (minimizing false negatives) matters more than perfect accuracy—critical for cancer detection. The SHAP integration enables regulatory compliance and physician trust, addressing adoption barriers that have slowed medical AI deployment. The superior recall metrics on oral cancer data carry specific clinical weight, as missed cancer cases carry severe consequences.

The technical community should monitor whether Mamba architectures generalize across other multimodal medical tasks. Success here could accelerate adoption in resource-limited healthcare environments and inspire similar mixed fusion designs in other domains requiring explainable predictions with heterogeneous data types.

• →Mamba-based architectures offer competitive alternatives to Transformers for medical image classification with lower computational requirements
• →Mixed fusion design enables separate processing of visual and tabular data, improving interpretability through SHAP analysis
• →Superior recall performance on oral cancer data suggests particular utility in high-stakes screening where false negatives are costly
• →State space models demonstrate viability for multimodal healthcare AI, addressing interpretability barriers to clinical adoption
• →Findings indicate Mamba effectiveness varies by dataset, with advantages particularly notable on histopathological image tasks