Sepsis, an autoimmune response to infection, is one of the deadliest and most expensive conditions treated in US hospitals, affecting approximately 750,000 Americans each year. Early detection and prevention could dramatically save lives, money and resources, yet there is no reliable way to diagnose sepsis quickly. Dr. Shamim Nemati and Dr. Ashish Sharma ,from the Department of Biomedical Informatics at Emory University’s School of Medicine, are taking an innovative approach to this challenge: using anonymised electronic health records collected from 30,000 patients in Emory’s intensive-care units (ICUs), Dr. Nemati created an AI engine to analyse 65 relevant variables, including vital signs, patient demographics and lab results. By continuously monitoring a patient’s data stream at five-minute intervals, the sepsis prediction engine constructs a composite score in real time, which predicts the likelihood of developing sepsis and displays its findings on a dashboard for clinicians to evaluate. Since early detection is key, clinicians can see the score and its rationale when treatment with antibiotics is most effective.
Emory University researchers use Google Cloud to predict sepsis in intensive care patients
By combining clinical data, machine learning and the scalable infrastructure of Google Cloud, Emory University’s sepsis prediction engine uses real-time analytics in an effort to provide better care for at-risk patients while also controlling medical costs.
"By converting our TensorFlow-based sepsis prediction algorithm into an app and running it on the Google App Engine, we were able to completely abstract away infrastructure requirements for running and scaling up the deployment and instead focus only on improving our algorithm."Shamim Nemati, Assistant Professor, Department of Biomedical Informatics, Emory University, USA
A critical-care solution
The engine has three crucial components: the incoming and stored data sets, the AI algorithm that analyses the data and a front-end user interface for clinicians. The data input and storage are particularly complex: tens of megabytes of high-resolution data such as blood pressure and respiration rates for each patient must be time-stamped, kept private and secure and be processed instantaneously in order to produce timely results under high-stakes conditions. The engine then produces a composite Sepsis Risk Score displayed on a dashboard that is designed to be easy for clinicians to read at a glance. An alarm system notifies clinicians when any patient reaches a threshold likely to contract sepsis, making it easier for busy carers to respond quickly.
Dr. Sharma designed the engine on Google Cloud using an integrated set of Google Cloud and open-source tools, such as TensorFlow and a set of containerised microservices, resulting in a smooth and almost instantaneous processing of data input, predictive analysis and output to the front end interface – all in real time. By building a Fast Healthcare Interoperability Resources (FHIR) database on Google Cloud, Nemati and Sharma are ensuring that the engine can scale and interoperate across institutions on a reliable, secure and private platform that also integrates with other projects on cloud technologies, like the wearable monitoring devices already in use at Emory hospitals.
Scaling through Google Cloud
So far, Nemati, Sharma and their team at Emory have partnered with the Emory eICU Center to validate the engine on data hosted on local servers, testing different time frames before achieving an impressive 85% accuracy at predicting sepsis four to six hours before onset. In order to deploy the programme at other sites they turned to App Engine. "By converting our TensorFlow-based sepsis prediction algorithm into an app and running it on the Google App Engine, we were able to completely abstract away infrastructure requirements for running and scaling up the deployment and instead focus only on improving our algorithm,” Nemati asserts.
Now that they know the engine works, they are planning to test it with more users, both patients and clinicians. They are also porting the algorithm over to Google Cloud Machine Learning Engine and TPUs for greater performance and scalability, and incorporating end-to-end encryption to minimise the potential exposure of patient data. By conducting a broad distributed study on Google Cloud, they can ask a new set of questions: What is the ideal time frame to make accurate predictions or optimise treatments? Will the engine help doctors to help patients better? How does a Risk Score affect treatment across different hospitals with their own local workflows and cultures?
In the end, what matters most is improving medical outcomes for real patients in ICUs, and Sharma is aware of that. “The reason why this algorithm is doing such a fantastic job is because it’s providing information in the actionable window when physicians can make meaningful interventions for a patient. Also, the algorithm opens up the deep-learning black box and informs the physician why it thinks the patient is at risk.” Nemati agrees, “A 2017 NEJM article showed that for each hour sepsis treatment is delayed, a patient’s risk of death increases by 4 percent. So what percentage of lives can we save if we could catch sepsis this way and put patients on antibiotics in time? We don’t know yet, but that’s what we’re currently testing at Emory, and we need to show generalizability elsewhere."