Scaling Biomedical Knowledge Graph Retrieval for Interpretable Reasoning: Applications to Clinical Diagnosis Prediction

Biomedical knowledge graphs (KGs) organize molecular mechanisms, biological pathways, and clinical concepts into structured representations that support diagnostic reasoning. As these graphs grow in scale and connectivity, scalable and interpretable multi-hop retrieval over deep biomedical graph structures has become a major computational bottleneck. We present LogosKG, a hardware-optimized retrieval system that enables efficient k-hop traversal over very large biomedical KGs using symbolic graph formulations and hardware-efficient execution. By integrating degree-aware partitioning, cross-partition routing, and on-demand caching, LogosKG scales to billion-edge graphs while preserving retrieval fidelity. Experiments demonstrate substantial efficiency improvements over CPU- and GPU-based baselines. Using diagnosis-oriented retrieval workloads as a downstream case study, we show that scalable access to deep, high-hop biomedical graph structures enables interpretable diagnostic evidence propagation. A clinician-aligned LLM-as-judge evaluation further indicates that high-hop KG retrieval improves reasoning quality in terms of accuracy, comprehensibility, and succinctness, underscoring the value of deep graph retrieval for diagnostic reasoning.