ProtoMedX: Towards Explainable Multi-Modal Prototype Learning for Bone Health Classification

Explainability is built into the model design, not added afterward: ProtoMedX uses a prototype-based architecture that is "explainable by design," meaning clinicians can directly understand why the model makes specific decisions about bone health. This contrasts with most AI systems that rely on post-hoc explanations added after the fact, which can be unreliable or incomplete.

Combining medical images with patient records improves accuracy: The multi-modal approach that integrates both DEXA scans and patient history achieved 89.8% accuracy compared to 87.58% using images alone, demonstrating that responsible AI systems should leverage multiple data sources rather than relying on a single input type.

Real-world performance matters for medical AI: The study used data from 4,160 actual NHS patients rather than synthetic or limited datasets, which is essential for validating that AI systems will work reliably in clinical practice where stakes are high and patient populations are diverse.

Regulatory compliance and clinical trust go hand-in-hand: The paper explicitly references the EU AI Act, highlighting that explainable AI isn't just ethically preferable—it's becoming a legal requirement. For medical applications, transparency helps clinicians maintain appropriate oversight and catch potential errors.

Visual explanations enable clinician verification: ProtoMedX provides explanations that clinicians can "visually understand," which is crucial for building trust and allowing medical professionals to validate whether the model's reasoning aligns with established clinical knowledge about bone health.