Machine　learning　provides　a　potential　modeling　framework　for　better　revealing　the　cognitive　state　of　the　human　brain　during　mental　arithmetic,　but　how　to　choose　the　features　of　mental　arithmetic　for　model　training　remains　to　be　studied.　However,　many　regions　in　the　cerebral　cortex　are　closely　related　to　mental　arithmetic,　and　they　are　important　seeds　for　distinguishing　cognitive　states　of　the　mental　arithmetic.　In　this　paper,　with　the　network　analysis　on　the　task-evoked　fMRI　data,　the　brain　regions　with　significant　differences　on　the　brain　functional　connectivity　properties　in　mental　arithmetic　are　found.　These　brain　regions　are　used　as　the　features　for　mental　arithmetic　cognitive　state　classification,　and　the　trained　model　is　constructed　by　using　multi-voxel　pattern　analysis.　The　results　show　that　the　performance　of　the　model　trained　by　the　brain　regions　with　significant　differences　in　network　properties　as　the　modeling　features　is　significantly　higher　than　that　of　the　model　constructed　by　the　brain　regions　without　significant　differences　in　network　properties.　After　further　reducing　the　dimensions　of　these　seeds　with　linear　discriminant　analysis,　the　average　ac-curacy　of　the　trained　models　for　all　subjects　reached　60.2%,　which　was　statistically　significantly　higher　than　the　chance　level.　Therefore,　the　trained　model　can　be　effectively　used　to　classify　mental　arithmetic　cognitive　states.　©　2019　IEEE.