A Discordance-Aware Multimodal Framework with Multi-Agent Clinical Reasoning

A team of researchers has published a paper proposing a novel, discordance-aware multimodal AI framework designed to tackle a major challenge in diagnosing and managing knee osteoarthritis (OA). The core problem is the frequent mismatch, or discordance, between the structural damage visible in X-rays and MRIs and the level of pain reported by the patient. This makes clinical interpretation difficult and limits the effectiveness of current decision-support tools. The new system addresses this by first training separate predictive 'expert' models on different data types: a CatBoost model for tabular clinical data (demographics, biomarkers), and ResNet18-based neural networks to extract features from MRI and X-ray images. These predictions are fused using a stacking ensemble.

The framework's most significant advancement is its two-stage reasoning process. First, it calculates a 'pain-structure discordance score' by comparing a patient's reported pain to the level of pain expected based on their structural damage alone. This score is then fed into a multi-agent reasoning layer. This layer acts like a team of AI specialists, interpreting the complex predictive signals and discordance to assign the patient to a specific, clinically meaningful OA phenotype (e.g., 'structural damage dominant' or 'pain dominant'). Finally, the system generates phenotype-specific management recommendations, moving from raw prediction to actionable clinical guidance. The model was developed and validated using baseline data from the extensive FNIH Osteoarthritis Biomarkers Consortium.