A　comprehensive　safety　evaluation　framework　has　been　proposed　incorporating　digital　twins　(DTs)　with　deep　learning.　The　physical　entity　and　the　DT　model　of　a　bridge　form　the　DT　body,　and　the　information　transmission　within　the　body　is　realized　by　an　information　interaction　medium.　Three　successive　DT　modeling　levels,　named　multilevel　integrated　DT　modeling,　are　defined　for　safety　evaluation.　The　first　two　levels　are　synchronously　updated　to　reflect　the　current　state　of　a　cable-stayed　bridge.　The　updating　is　carried　out　by　using　cable　force　influence　matrices　to　update　the　perceptual　data.　A　deep　learning　algorithm,　named　the　gated　recurrent　unit　neural　network,　is　used　as　the　information　interaction　medium　for　the　real-time　cable　force　predictions,　thereby　enabling　the　hyper-real-time　DT　modeling　at　the　third　level　for　safety　prognosis.　The　validation　on　a　cable-stayed　bridge　has　illustrated　the　application　to　the　different　functions　covering　the　state　assessment,　safety　warning　and　safety　prognosis.　©　2025　Elsevier　Ltd