Rare diseases: FusionRare combines four categories of AI and identifies the correct diagnosis as the top result in 79 per cent of cases

Symbolic reasoning, deep learning, similar case search and LLMs: by combining these approaches, FusionRare outperforms GPT 5.5 and DeepRare on 2,528 individual clinical cases, saving doctors 21 per cent of their time.

Some 300 million people worldwide live with a rare disease, and many wait years before receiving a diagnosis. A team of researchers has presented FusionRare on arXiv, a system that does not rely on a single AI technique but combines four: symbolic reasoning based on medical knowledge, pattern recognition via deep learning, the identification of patients with similar profiles, and heuristic inference from a large language model. Their opinions are statistically aggregated and then reviewed by an agent that draws on a knowledge base.

Results measured without data leakage

The authors took care to ensure that the test cases had not been seen during training – a common bias with LLMs. Out of 2,528 clinical cases isolated in this way, the correct diagnosis ranked first in 51 per cent of cases based on symptoms alone, and in 79 per cent when genetic data was included — outperforming the agent-based DeepRare system, GPT 5.5, and the specialised tools PhenoBrain and LIRICAL.

A randomised crossover trial involving clinicians shows that the tool places the correct diagnosis higher up in doctors’ rankings for 45 per cent of decisions, boosts their confidence in 62 per cent of cases and reduces the time needed to reach a conclusion by 21 per cent.