Portable Early Prediction of Sepsis from Clinical Data on Intel Myriad X

Sepsis is a life-threatening condition where microbes present in the blood stream cause an unregulated immune response from the body which can result in tissue damage, multi-organ failure and eventually death. It affects 30 million people worldwide and causes 6 million deaths. Studies have indicated that every hour sepsis goes undetected, patient mortality increases 4-8%, thus early detection of the disease is necessary to decrease mortality rates and provide better patient outcomes. This project is derived from this year’s 2019 Computing in Cardiology Challenges, which focuses on the early detection of sepsis using machine learning algorithms. The two primary focuses of the project include (1) developing a model which can successfully detect sepsis early and (2) demonstrating the clinical viability of this deep learning model by implementing the model on Intel’s Neural Compute Stick 2 (Myriad X VPU processor) which is a portable USB device that can implement and deploy deep learning models. While developing the model, the data was preprocessed in the following ways: (1) linear interpolation was used to fill all missing values and values not collected were 0, (2) synthetic minority oversample technique (SMOT) was used to balance the data set, (3) patient file lengths were standardized to have only 3 and 17 hours of patient data, (4) all 40 parameters were used (although a variety of subsets were tested). Data preprocessing was done using MATLAB. The model was developed in python using Tensorflow and Keras. The model itself was a seven layer dense neural network (DNN) with Leaky Rectified Linear Unit (ReLU) activation functions for all hidden layers and a sigmoid activation function for the output layer. This was paired with binary cross entropy loss. The resulting model detected sepsis well, with a 76% sensitivity (correctly identified sepsis) and 80% specificity (correctly identified healthy patients). The model was successfully implemented on Intel’s Neural compute stick 2 (Myriad X VPU processor) with an average of 10ms processing time to determine whether the patient had sepsis, demonstrating clinical viability. Further testing including tuning hyperparameters to further increase sensitivity and specificity is currently being done