The de facto standard in machine learning architecture is to rely on cloud computing todo all the needed data processing from Learning to classication, while edge nodes or IoT (Internet of Things) devices provide the raw data without formatting or ltering. Although this approach shows its ecacy in many areas, it suers lots of drawbacks. Such as privacy, massive data/storage management, and inecient power management in limited resources edge nodes, which is our dissertation's concern. We are listing and testing dierent architectures that can improve edge node power consumption and eciency in data handling, enhancing privacy, and data management on cloud computing. We have applied dierent architectures to the Internet of Medical Things (IoMT) eld, which drastically improves real-time monitoring and recording of electrocardiogram (ECG) signals. We have used various communication protocols to show the impact on overall power consumption. Also, we have presented results that can guide architecture and communication selection according to application usage. Our experiments and proling were done on actual targets from IoT market. We used Bluetooth Low Energy (BLE) Nordic nRF52 Development Kit and SLWSTK6020B with EFR32BG Bluetooth SoC Starter Kit from Silicon Labs.