An Optimized Hardware Implementation of Deep Learning Inference for Diabetes Prediction

Diabetes Mellitus (DM) is a chronic disease characterized by the reduced metabolic action of glucose in the human body that triggers the failure of proper functions of various organs. Diagnosis of diabetes is challenging due to the complex correlation with other diseases and initial asymptomatic behavior. Efficient hardware design that incorporates the evaluation of risk factors of the disease can improve early diagnosis, regular monitoring, and clinical decision-making. In this paper, we present an implementation of a deep learning (DL) inference in Field Programmable Gate Array (FPGA) to predict DM. The model hyperparameters have been tuned to obtain the DL model that ensures acceptable performance in low-power and miniaturized silicon area with reduced storage requirements. We trained a four-layer Fully Connected Neural Network (FCNN) with RMSProp optimizer and binary cross-entropy loss function. The best-learned model among 200 epochs is used to extract the weights and biases for hardware implementation of the inference module. We achieved an accuracy of 91.15%. The outcome of this research can be integrated with the system-on-chip platform to develop smart diabetic monitoring and management tools.