Disease　analysis　using　multimodal　physiological　signals　is　currently　a　hot　research　area.　Aiming　at　the　problem　that　current　multimodal　feature　fusion　approaches　neglect　the　correlation　between　different　modalities,　this　paper　proposes　a　hybrid　multimodal　fusion　model　based　on　the　attention　mechanism　for　depression　recognition.　The　model　selects　two　modalities,　EEG　and　speech　signals,　for　feature　fusion.　In　the　feature　extraction　stage　of　EEG　signals,　the　correlation　between　features　is　analyzed　using　Spearman’s　rank　correlation　coefficient　to　achieve　dimensionality　reduction.　In　the　feature-level　modal　fusion　stage,　the　attention　mechanism　is　applied　to　automatically　learn　the　contribution　of　different　features　and　the　complementarity　between　different　modalities.　In　the　model　fusion　layer,　a　multi-layer　long　and　short-term　memory　network　(ML-LSTM)　is　adopted,　which　makes　full　use　of　the　complementarity　between　the　features　of　EEG　signals　and　audio　signals.　Finally,　the　outputs　of　feature-level　fusion　and　model-level　fusion　are　weighted　and　combined　to　obtain　the　final　prediction　results.　Experimental　results　on　the　MODMA　dataset　show　that　the　accuracy　of　the　proposed　multimodal　fusion　model　for　depression　recognition　reached　88.54%.　©　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.　2025.