Diagnosis　prediction　is　becoming　crucial　to　develop　healthcare　plans　for　patients　based　on　Electronic　Health　Records　(EHRs).　Existing　works　usually　enhance　diagnosis　prediction　via　learning　accurate　disease　representation,　where　many　of　them　try　to　capture　inclusive　relations　based　on　the　hierarchical　structures　of　existing　disease　ontologies　such　as　those　provided　by　ICD-9　codes.　However,　they　overlook　exclusive　relations　that　can　reflect　different　and　complementary　perspectives　of　the　ICD-9　structures,　and　thus　fail　to　accurately　represent　relations　among　diseases　and　ICD-9　codes.　To　this　end,　we　propose　to　project　disease　embeddings　and　ICD-9　code　embeddings　into　boxes,　where　a　box　is　an　axis-aligned　hyperrectangle　with　a　geometric　region　and　two　boxes　can　clearly　＇include＇　or　＇exclude＇　each　other.　Upon　box　embeddings,　we　further　obtain　patient　embeddings　via　aggregating　the　disease　representations　for　diagnosis　prediction.　Extensive　experiments　on　two　real-world　EHR　datasets　show　significant　performance　gains　brought　by　our　proposed　framework,　yielding　average　improvements　of　6.04%　for　diagnosis　prediction　over　state-of-the-art　competitors.　©　2024　Copyright　held　by　the　owner/author(s).　Publication　rights　licensed　to　ACM.