Severe　depression　often　exhibits　suicidal　tendencies,　making　early　identification　and　intervention　crucial　to　prevent　its　further　progression.　This　study　focuses　on　developing　a　high-performance　method　and　device　for　early　detection　of　depression.　We　propose　a　hybrid　framework　for　early　depression　detection　that　integrates　multiple　deep　learning　techniques　and　ensemble　learning.　This　framework　features　a　dual　bidirectional　temporal　convolutional　network　(BiTCN)　to　encode　both　local　and　global　causal　relationships,　a　bidirectional　long　short-term　memory　(BiLSTM)　network　to　capture　long-term　dependencies　and　contextual　relationships,　an　emotional　cross-attention　(eCA)　module　to　encode　the　significance　of　different　emotions,　a　multimodal　feature　cross-attention　(MFCA)　mechanism　to　prioritize　various　feature　modalities,　and　an　ensemble　learning　method　to　decode　and　infer　depression　detection.　The　input　signals　include　pupil　waves　and　pulse　rate　variability　(PRV)　signals,　measured　during　both　calm　(non-emotional)　and　emotional　states　such　as　sadness,　happiness,　fear,　and　tension.　To　enhance　the　generalization　capability　of　the　model,　data　augmentation　techniques　are　applied　to　the　training　dataset.　Test　results　show　that　detection　performance　based　on　emotional　cues　(ECs)　is　superior　to　that　based　on　non-ECs　(calm).　Notably,　the　fusion　of　pupil　waves　and　PRV　signals　with　ECs　has　achieved　state-of-the-art　(SOTA)　performance　in　depression　detection.　These　findings　highlight　the　crucial　role　of　emotional　signals　in　improving　the　performance　of　depression　detection.　The　end-to-end　high-performance　automatic　detection　of　early　depression　(ADED)　device　developed　in　this　study　can　serve　as　an　early　detection　tool,　thereby　promoting　the　potential　application　of　artificial　intelligence　technology　in　mental　health　screening　and　clinical　practice.　©　1963-2012　IEEE.