Learning　unsupervised　representations　from　multivariate　medical　signals,　such　as　multi-modality　polysomnography　and　multi-channel　electroencephalogram,　has　gained　increasing　attention　in　health　informatics.　In　order　to　solve　the　problem　that　the　existing　models　do　not　fully　incorporate　the　characteristics　of　the　multivariate-temporal　structure　of　medical　signals,　an　unsupervised　multi-Context　deep　Convolutional　AutoEncoder　(mCtx-CAE)　is　proposed　in　this　paper.　Firstly,　by　modifying　traditional　convolutional　neural　networks,　a　multivariate　convolutional　autoencoder　is　proposed　to　extract　multivariate　context　features　within　signal　segments.　Secondly,　semantic　learning　is　adopted　to　auto-encode　temporal　information　among　signal　segments,　to　further　extract　temporal　context　features.　Finally,　an　end-to-end　multi-context　autoencoder　is　trained　by　designing　objective　function　based　on　shared　feature　representation.　Experimental　results　conducted　on　two　public　benchmark　datasets　(UCD　and　CHB-MIT)　show　that　the　proposed　model　outperforms　the　state-of-the-art　unsupervised　feature　learning　methods　in　different　medical　tasks,　demonstrating　the　effectiveness　of　the　learned　fusional　features　in　clinical　settings.　©　2020,　Science　Press.　All　right　reserved.