Hardware-Efficient LSTM-Based Wearable Neural Network for Predicting Blood Glucose Levels in Diabetic Patients

P. Lokanadham ,Assistant Professor ,Dept of Information Technology ,SV College of Engineering ,Tirupati,India.
Basabathina Muni Roopa Sri ,B. Tech, Dept of Information Technology, SV College of Engineering,Tirupati, India.
Sarvepalli Sai Uday Kiran ,B. Tech, Dept of Information Technology, SV College of Engineering, Tirupati,India.
Thambisetty Nandhini, B. Tech, Dept of Information Technology, SV College of Engineering, Tirupati,India.
Bommisetti Sunil, B. Tech, Dept of Information Technology, SV College of Engineering, Tirupati, India.
Email: lokanadham.p@svce.edu.in, roopabasabathina@gmail.com, sarvepallisaiuday@gmail.com,
nandinithambisetty@gmail.com, bommisettisunil241@gmail.com.

Abstract-Blood glucose levels in diabetes are critical for managing health and preventing complications. Diabetes patients monitor their blood glucose levels regularly to adjust diet, medication, and lifestyle for effective control and better health outcomes. The existing LSTM FPGA based blood glucose prediction system uses traditional sigmoid and tanh activations, resulting in high energy use (76,594 nJ), longer latency (260 µs), and large FPGA area consumption, making it less ideal for long-term wearables. The computationally intensive activationfunctions require additional DSPs, increasing complexity and energy overhead, while the system's structure is not optimized for hardware efficiency, limiting battery life and real-time performance. More recent approaches focus on low-power FPGA implementations with optimized activation functions and neural network quantization to improve energy efficiency and maintain accuracy for wearable devices.The proposed system improves this by usinghardware-friendly hard sigmoid and hard tanh functions,  applying quantization, and reducing FPGA resources by two-thirds. This leads to a 15.5-fold reduction in energy consumption (4,950 nJ), 18.57 times faster predictions (14 µs), and elimination of DSPs, while maintaining clinical accuracy, enhancing device autonomy, real-time monitoring, and enabling edge computing for better privacy and personalization Keywords:Blood glucose,clinical quantization, FPGA, edge computing