Timely　and　accurate　prediction　of　lane-changing　(LC)　risk　is　crucial　for　drivers　to　make　safe　LC　decisions.　This　study　proposes　a　spatiotemporal　attention　graph　neural　network　model　(STAG)　based　on　multivehicle　interaction　graph　modeling　to　characterize　the　dynamic　relationships　among　vehicles　in　a　connected　environment　and　predict　upcoming　LC　risks.　Specifically,　graph　theory　is　employed　to　model　the　interactions　among　a　LC　vehicle　and　its　surrounding　vehicles.　A　deep　learning　model　combining　a　graph　attention　network　(GAT),　gated　recurrent　unit　(GRU),　and　attention　mechanism　is　proposed　to　extract　the　spatiotemporal　features　of　multivehicle　interaction　graphs　for　LC　risk　prediction.　The　proposed　method　was　validated　using　the　highD　data　set.　The　results　show　that　(1)　compared　with　traditional　feature　input　methods,　using　multivehicle　interaction　graphs　can　improve　LC　risk　prediction　accuracy　by　1.5%;　and　(2)　the　STAG　model　accurately　extracts　the　spatiotemporal　features　of　multivehicle　interaction　graphs.　The　average　accuracy　of　LC　risk　prediction　was　4.4%　higher　than　that　of　baseline　models.　The　findings　of　this　study　provide　valuable　insights　for　traffic　safety　management　and　the　design　of　advanced　driver　assistance　systems　(ADAS).