GraD-IBD: Graph Representation Learning from Diagnosis Trajectories for Early Detection of Inflammatory Bowel Disease

Researchers propose GraD-IBD, a graph-based machine learning model that analyzes patient diagnosis histories encoded in ICD codes to detect inflammatory bowel disease risk earlier and more efficiently than existing sequential models. The approach reformulates longitudinal diagnostic trajectories as temporally directed graphs with a novel message-passing mechanism, demonstrating improved accuracy while reducing computational complexity.

GraD-IBD addresses a fundamental challenge in clinical AI: processing the irregular, hierarchical nature of patient diagnostic data. Traditional sequential models struggle with ICD code sequences because they impose n-dimensional lattice constraints unsuited to healthcare's branching diagnostic patterns. By reframing this problem as graph representation learning, the researchers create a more natural alignment between data structure and model architecture.

This work builds on growing recognition that healthcare data benefits from graph-based approaches. Medical diagnosis sequences are inherently non-linear—patients accumulate conditions over time in varied orders, and diagnoses relate hierarchically within ICD's classification system. Previous methods either oversimplified these relationships or created unnecessarily complex architectures to handle them. The context-aware, time-decay message passing mechanism elegantly captures temporal dependencies while respecting the temporal directionality of patient encounters.

For the broader healthcare AI ecosystem, this demonstrates how domain-specific model design beats generic sequential approaches. Clinical institutions and AI developers building risk prediction systems will find value in the reduced computational overhead—critical for resource-constrained healthcare settings. Earlier IBD detection translates to improved patient outcomes and reduced downstream treatment costs. The graph representation learning framework is generalizable to other chronic disease detection tasks beyond IBD, creating potential applications across gastroenterology, rheumatology, and other specialties managing complex diagnostic histories.

Future work should focus on deployment validation across diverse healthcare systems with varying ICD coding practices and dataset sizes. Integration with electronic health record systems and real-time clinical decision support represents the next frontier.

• →GraD-IBD reformulates diagnostic sequences as temporally directed graphs, better matching the actual structure of clinical data than sequential models
• →The model achieves superior IBD detection performance while substantially reducing computational complexity compared to state-of-the-art approaches
• →Graph representation learning from ICD codes demonstrates a generalizable framework applicable to other chronic disease risk prediction tasks
• →The time-decay message passing mechanism efficiently captures temporal dependencies in irregular patient diagnostic trajectories
• →Healthcare institutions can deploy this approach with lower computational overhead, enabling scalable disease risk prediction in resource-limited settings