CRGANNC-GD: Graph-Diffusion End-to-End Drug Class Recommendation for Stroke Decision Support from Symptom Profiles

Abstract

Stroke remains a leading cause of mortality and disability worldwide, requiring timely therapeutic decisions. Existing content-based drug recommendation approaches often rely on static similarity measures and multi-stage pipelines (e.g., clustering followed by instance-based retrieval), which may limit scalability and generalization. We propose CRGANNC-GD, an end-to-end recommendation framework that models patient profiles and medication classes as a heterogeneous graph and performs representation learning through graph diffusion to capture clinically meaningful relationships between symptoms, comorbidities, demographics, and drug classes. Unlike traditional KNN-centric pipelines, CRGANNC-GD learns a task-optimized latent space and outputs calibrated scores for multiple drug classes (antihypertensive, anticoagulant, fibrate). We further incorporate a ranking-aware loss to directly optimize top- k recommendation quality and introduce an explainability module that highlights influential clinical attributes and graph neighborhoods contributing to each recommendation. Experiments on a stroke dataset with 9691 records demonstrate that CRGANNC-GD improves predictive performance and recommendation quality over classical content-based baselines, while maintaining fast inference suitable for clinical decision support and telemedicine settings. Our results suggest that graph-diffusion representation learning provides a scalable and robust alternative for medication class recommendation in stroke management, with a clear pathway to extend the framework toward novel drug discovery via drug-drug and drug-target similarity integration.