The　surge　in　artificial　intelligence-driven　disease　prediction　technology　is　profoundly　transforming　the　educational　management　experience　of　medical　students　and　practitioners.　This　technological　advancement　not　only　provides　them　with　cutting-edge　knowledge　in　the　field　but　also　enhances　their　understanding　of　complex　medical　data　and　clinical　decision　support　systems,　thereby　integrating　advanced　technology　into　clinical　practice　and　further　advancing　the　medical　field.　However,　the　issue　of　feature　sparsity　in　electronic　medical　records　(EMRs)　has　become　a　major　constraint　on　the　advancement　of　current　technologies.　Graph　Neural　Network　(GNN)　can　model　the　graph　structured　data　and　infer　missing　features,　which　becomes　an　effective　solution　to　address　the　above　issue.　However,　if　two　nodes　are　far　from　each　other　(cross　more　than　three　sentences),　the　relation　among　them　is　difficult　to　be　inferred　by　GNN,　which　may　loss　some　useful　information　for　disease　prediction.　This　paper　presents　a　Knowledge-based　Attention　Network　(KAN)　for　disease　prediction,　which　aims　to　build　relations　between　long-distance　nodes　by　introducing　the　knowledge,　thus　fully　using　the　information　and　improving　the　disease　prediction　accuracy.　Specifically,　beside　linking　the　entity　nodes　based　on　the　extracted　inherent　relations,　a　medical-related　knowledge　graph　(KG)　and　a　trained　relation　completion　(RC)　model　are　leveraged　to　infer　more　potential　relations　of　nodes.　Then,　constructing　the　EMR　graphs,　thereby　learning　the　features　of　nodes　and　relations　to　generate　representative　embeddings　for　disease　prediction.　The　results　on　the　real-world　dateset　demonstrate　the　superiority　of　KAN.　The　advanced　principles　of　KAN　infuse　new　knowledge　into　the　education　and　management　of　healthcare　professionals,　enabling　them　to　learn　innovative　approaches　for　leveraging　cutting-edge　technologies　in　managing　complex　medical　data.　©　2024　IEEE.