College　students　face　increasing　stress　from　difficulties　with　studies,　employment,　and　social　interactions,　which,　if　left　unaddressed,　may　lead　to　depression　and　physical　illnesses.　Currently,　the　detection　of　stress　severity　relies　on　self-assessment　scales,　while　machine　learning　or　deep　learning-based　approaches　primarily　focus　on　classification.　This　study　proposes　an　approach　using　pulse　signals　containing　emotional　cues　and　deep　learning　to　automatically　detect　the　severity　of　stress　in　college　students.　Firstly,　pulse　signals　of　177　college　students　were　collected　using　photoplethysmography　(PPG)　during　they　watched　five　virtual　reality　(VR)　emotional　scenes,　including　calm,　sadness,　happiness,　fear,　and　tension.　Pulse　rate　variability　(PRV)　and　discrete　PPG　(dPPG)　were　extracted　from　these　signals　as　input　for　detecting　stress　severity.　Then,　the　proposed　stress　detection　framework,　1DCNN-BiLSTM　+　Cross-Attention　+　XGBoost,　was　employed　to　detect　stress　severity,　incorporating　an　emotional　Cross-Attention　mechanism.　The　impact　of　induced　emotions　on　stress　severity　detection　performance　was　examined.　The　results　indicated　that　stress　severity　detection　in　emotional　scenes　outperformed　in　calm.　Furthermore,　the　detection　performance　that　integrates　multiple　emotions　surpassed　single　emotions.　The　fusion　of　PRV　and　dPPG　signals　yielded　the　best　detection　performance.　This　study　provides　an　end-to-end　automated　approach　for　detecting　stress　severity　in　college　students.　　©　2010-2012　IEEE.