The　precise　dosage　of　insulin　plays　an　important　role　in　the　treatment　of　diabetes.　To　offer　accurate　dosage,　some　AI-based　auxiliary　dosing　systems　have　been　proposed.　Unfortunately,　these　schemes　demand　real-time　health　data,　which　is　highly　relevant　to　the　health　situation　of　the　diabetics.　The　traditional　personalized　drug　delivery　frameworks　for　accurate　dosing　of　insulin　always　collect　and　transmit　medical　data　in　plaintext,　which　may　cause　the　disclosure　of　user　privacy.　Therefore,　to　optimize　insulin　dosage　and　protect　privacy　simultaneously,　we　propose　a　framework　for　an　optimized　insulin　dosage　via　privacy-preserving　reinforcement　learning　for　diabetics　(OIDPR).　In　OIDPR,　both　the　additive　secret　sharing　and　edge　computing　are　deployed　to　complete　data　encryption　and　improve　efficiency.　The　user＇s　medical　data　is　divided　into　secret　shares　uniformly　at　random,　then　compute　separately　at　the　edge　servers.　During　the　computation　task　of　Q-learning,　data　is　stored　in　the　format　of　ciphertext　and　processed　using　the　proposed　additive　secret　sharing　protocol.　Finally,　comprehensive　theoretical　analyses　and　experiment　results　demonstrate　the　security　and　efficiency　of　our　framework.　©　2020　IFIP.