Reinforcement Learning-Based Temporal Knowledge Graph Reasoning for Predicting Chronic Gastritis Diagnosis and...

The clinical progression of chronic gastritis involves intricate temporal dependencies, which makes it difficult to capture both the dynamic trajectory of the disease and the underlying relationships among medical events using conventional methods.

This study aims to dynamically predict chronic gastritis diagnosis. We propose RL4TKGR, a reinforcement learning-based temporal knowledge graph (TKG) reasoning model, to predict a chronic gastritis diagnosis.

RL4TKGR incorporates a reinforcement learning policy network that uses a dual-path encoding module to model historical and nonhistorical diagnostic information separately. RL4TKGR further integrates a dual-channel reward function with a dynamic weight allocation mechanism, which adaptively balances the two information sources. This design addresses the strategic bias problem and enables interpretable reasoning. The chronic gastritis temporal knowledge graph (CG-TKG) was constructed from chronic gastritis diagnosis and treatment records of 17,906 patients comprising 38,360 visits from March 2009 to October 2022.

Experiments on the self-constructed CG-TKG resulted in RL4TKGR achieving the highest mean reciprocal rank (MRR) on CG-TKG-4 (51.66, SD 0.76), CG-TKG-8 (53.18, SD 0.79), and CG-TKG-12 (56.95, SD 0.84). After adding recent TKG reasoning baselines, adaptive path-memory network (DaeMon) achieved the strongest baseline MRR for all 3 subsets, with MRR values of 43.26 (SD 0.95), 51.76 (SD 1.12), and 54.87 (SD 1.17), respectively. Compared with DaeMon, the overall paired t tests across MRR and Hits@1/3/10 yielded 𝑃<.001 for all 3 subsets. Ablation experiments and case analyses further supported the contribution of the historical modeling components and the practical utility of the model for predicting disease subtypes and therapeutic medications.

This study improves chronic gastritis diagnosis and treatment prediction by integrating a history-aware dual-path policy network with a dynamic event balance factor in reinforcement learning-based TKG reasoning.