AI framework improves rare disease diagnosis support in pediatric cases
KGRD outperformed its underlying AI model across a 420-case benchmark
Written by |
An artificial intelligence (AI) framework that combines patients’ symptoms and genetic findings with established knowledge about rare diseases may improve diagnostic support for children with rare genetic disorders, such as AADC deficiency, a study suggests.
Called KGRD, the framework is built on a large language model (LLM) — a type of AI used to generate human-like language — but rather than relying on the LLM alone to reach a diagnosis, it uses specialized components to evaluate different types of patient and medical information, combines the evidence, and checks how strongly it supports a proposed diagnosis. The process is designed to mimic selected parts of how a medical team weighs different sources of evidence when evaluating a patient.
KGRD outperforms its underlying AI model in 420 cases
When tested on 420 rare-disease cases, the best-performing version of KGRD generated a clinically useful candidate diagnosis in 81.9% of cases, compared with 73.6% when the same underlying LLM was used on its own. That amounted to 35 additional cases in which KGRD suggested a diagnosis judged close enough to the confirmed diagnosis.
“These results indicate that KGRD provides effective diagnostic support for paediatric rare diseases,” the researchers wrote.
The study, “KGRD: a knowledge-graph-augmented automated reasoning framework for diagnosis and counselling of paediatric rare genetic disorders,” was published in npj Digital Medicine.
Rare diseases collectively affect an estimated 3.5%-5.9% of people worldwide and encompass more than 7,000 conditions. About 70%-80% are genetic disorders, and most begin during childhood. Yet reaching a diagnosis can be difficult because individual conditions are uncommon, specialist expertise may be limited, and information needed to identify them can be sparse or scattered across different sources.
These challenges can also make it difficult for conventional LLMs to identify the correct diagnosis in rare-disease cases. LLMs learn from enormous collections of existing information, but far more data are generally available for relatively common and well-characterized diseases than for ultra-rare disorders. This can bias LLMs toward more familiar diagnoses and make unusual presentations easier to miss, the researchers noted.
Researchers in China developed KGRD to overcome some of these limitations. The framework uses an LLM as its underlying AI model, but supplements it with structured rare-disease knowledge organized into a knowledge graph. This is essentially a large network that maps known relationships between diseases, genes, symptoms, biological pathways, and other relevant features.
Specialized AI agents combine genes, symptoms, and patient matches
KGRD draws on this information when evaluating a patient’s case and uses three specialized AI agents to generate or prioritize possible diagnoses. GeneAgent uses a patient’s genetic findings to identify and prioritize diseases that could explain the condition, PhenoAgent matches the patient’s symptoms with those associated with known diseases, and PatientAgent searches for patients with similar clinical features.
The findings are then considered by additional AI agents designed to mimic a multidisciplinary medical discussion. A Supervisor Agent coordinates this discussion, weighing the diagnoses and reasoning proposed by the specialized AI agents, and determining whether they have reached enough agreement or need further discussion.
A separate Verifier Agent evaluates how well a proposed diagnosis is supported by multiple sources of evidence, including the knowledge graph, gene-, symptom-, and disease-related evidence, and published research.
The team then tested KGRD on three datasets comprising 420 cases with confirmed rare-disease diagnoses. Genotype information was provided to the model for 150 cases, while the remaining 270 were evaluated without genotype-derived information.
The strongest results came from a version of KGRD that used DeepSeek-V3.1 as its underlying LLM. This version generated clinically useful candidate diagnoses for 81.9% of the 420 cases, compared with 73.6% when DeepSeek-V3.1 was asked to diagnose the same cases without the additional KGRD framework.
Gains persist when genetic data are withheld from the model
KGRD continued to outperform the underlying LLM when genotype-derived information was not provided to the model. Among the 270 cases evaluated using clinical features alone, it generated clinically useful candidate diagnoses in 75.2% of cases, compared with 64.4% for DeepSeek-V3.1 alone.
KGRD did not perform equally well in every setting. It struggled when tested on the Undiagnosed Diseases Network (UDN) dataset, which included particularly difficult-to-diagnose rare-disease cases. Nearly all of the models tested performed poorly on this dataset, with even the best-performing version of KGRD generating clinically useful candidate diagnoses in only 34.8% of cases.
Overall, the findings suggest that AI may benefit rare-disease diagnosis when it combines and verifies multiple sources of evidence, rather than relying on the reasoning of a single AI model alone, the researchers wrote. Continued progress, they said, will require improving the knowledge these systems rely on and testing them in ways that better reflect the complexity of real-world diagnosis.
The goal is ultimately to develop automated systems “that assist clinicians in making safer, more informed, and more reproducible diagnoses and counselling decisions,” the researchers concluded.
Several study authors are inventors on a Chinese patent application related to aspects of the KGRD framework.
Leave a comment
Fill in the required fields to post. Your email address will not be published.