Prostate Cancer Classification Using Multimodal Feature Fusion and Explainable AI

Combining different types of medical data improves cancer detection: The system achieved 98% accuracy by analyzing both written clinical notes (using BERT) and numerical lab results (using Random Forest) together, outperforming either data type alone—particularly for intermediate cancer stages where it reached 90% recall compared to 72.5% using text alone.

Explainability is built into the diagnostic tool, not added afterward: The researchers used SHAP analysis to provide transparent rankings of which features (clinical factors) most influenced each diagnosis decision, enabling doctors to understand why the AI made a specific recommendation rather than treating it as a "black box."

Simpler, interpretable approaches can match complex AI systems: Rather than using cutting-edge deep learning architectures, the researchers demonstrated that a straightforward combination of established techniques (BERT and Random Forest) delivers clinically meaningful results while remaining computationally efficient enough for hospitals to actually implement.

The system is designed for practical hospital deployment: By balancing high performance (89% F1 score) with interpretability and computational efficiency, this approach addresses real-world constraints that prevent many AI tools from being adopted in clinical settings.

Different data types reveal complementary diagnostic information: The ablation studies proved that textual clinical notes provide diagnostic value beyond what numerical lab data alone can offer, suggesting that comprehensive AI systems should integrate multiple information sources.