The　medical　service　provider　establishes　a　heart　failure　prediction　model　with　deep　learning　technology　to　provide　remote　users　with　real-time　and　accurate　heart　failure　prediction　services.　Remote　users　provide　their　health　data　to　the　health　care　provider　for　heart　failure　prediction　through　the　network,　thereby　effectively　avoiding　the　damage　or　death　of　vital　organs　of　the　patient　due　to　the　onset　of　acute　heart　failure.　Obviously,　sharing　personal　health　data　in　the　exposed　data　sharing　environment　would　lead　to　serious　privacy　leakage.　Therefore,　in　this　paper,　we　propose　a　privacy-preserving　heart　failure　prediction　(PHFP)　system　based　on　Secure　Multiparty　Computation　(SMC)　and　Gated　Recurrent　Unit　(GRU).　To　meet　the　real-time　requirements　of　the　PHFP　system,　we　designed　a　series　of　data　interaction　protocols　based　on　additional　secret　sharing　to　achieve　lightweight　outsourcing　computing.　Through　these　protocols,　we　can　protect　the　user’s　health　data　privacy　while　ensuring　the　efficiency　of　the　heart　failure　prediction　model.　At　the　same　time,　to　provide　high-quality　heart　failure　prediction　services,　we　also　use　the　new　mathematical　fitting　method　to　directly　construct　the　safety　activation　function,　which　reduces　the　number　of　calls　to　the　security　protocol　and　optimizes　the　accuracy　and　efficiency　of　the　system.　Besides,　we　built　a　security　model　and　analyzed　the　security　of　the　system.　The　experimental　results　show　that　PHFP　takes　into　account　the　safety,　accuracy,　and　efficiency　in　the　application　of　heart　failure　prediction.　©　2020,　Springer　Nature　Switzerland　AG.