KGRD: a knowledge-graph-augmented automated reasoning framework for diagnosis and counselling of paediatric rare genetic disorders.

Abstract

Diagnosis and counselling for paediatric rare diseases remain constrained by the sparsity of structured patient-level data and fragmented genetic knowledge, which can induce 'common-attention' bias in conventional large language models (LLMs). Here, we describe KGRD, a knowledge-graph-augmented diagnostic support framework empowered by knowledge-driven and data-driven inference over patient-level genomic and phenotypic data. KGRD consists of three specialised inference agents for deductive reasoning regarding disease aetiology, as well as a collective decision-making module that integrates multidisciplinary deliberation with multi-source verification. In a validation benchmark of 420 rare disease cases, KGRD(DS) achieved the strongest overall performance, increased the mean Bond score of the top-ranked diagnosis from 3.27 to 3.85 and raised CIE from 73.6 to 81.9%, corresponding to 35 additional cases with candidate-diagnosis Bond score ≥4. Together, these results indicate that KGRD provides effective diagnostic support for paediatric rare diseases.