Motion2Meaning: A Clinician-Centered Framework for Contestable LLM in Parkinson's Disease Gait Interpretation

Clinicians need meaningful ways to challenge AI decisions in healthcare: The paper demonstrates that current AI clinical dashboards lack transparency and give doctors no recourse when they disagree with AI predictions. Motion2Meaning addresses this by creating a "Contestable AI" system where clinicians can actively question and override AI recommendations about Parkinson's Disease severity.

A new safeguard called XMED can detect when AI models are unreliable: The system identifies problematic predictions by measuring "explanation discrepancies"—finding a five-fold difference between incorrect predictions (7.45% discrepancy) and correct ones (1.56%). This gives clinicians an automated warning system for when to be especially skeptical of the AI's output.

There's a critical trade-off between AI accuracy and usability: The paper reveals that making the AI more factually grounded (more accurate) can make it harder for clinicians to read and use quickly. This suggests that responsible AI design requires balancing multiple competing goals rather than optimizing for a single metric.

Wearable sensor data combined with AI can objectively measure Parkinson's Disease progression: The system achieved 89.0% accuracy in predicting disease severity stages using ground reaction force data from wearable sensors, suggesting this approach could provide objective, continuous monitoring outside traditional clinical settings.

Human-centered design is essential for responsible AI in medicine: The framework was specifically built around clinician workflows and needs rather than purely technical optimization, emphasizing that trustworthy AI requires understanding how doctors actually work.