Article

How Cognitive Agents (Bots) May Help "Solve" Sepsis

Subtle Signs, Potentially Dire Consequences

We ask Google for directions. We ask Siri to dial a friend, Alexa for the weather forecast. Typically, we rely on smart assistants for convenience and ease. But might they also have an application during moments when human lives hang in the balance?

Could a chatbot support a clinician who has to sift and winnow through mounds of data in an electronic medical record (EMR) when every minute counts?

Take for example sepsis.  It affects approximately 1.7 million adults in the United States and is a leading cause of mortality in hospitals, potentially contributing to more than 250,000 deaths annually.1 In spite of numerous performance improvement initiatives aimed at increasing awareness of sepsis—as well as advances in the availability of patient electronic medical record (EMR) data—recognizing sepsis deterioration in its early stages remains a significant challenge for clinicians and medical staff. Not only are the signs and symptoms often nonspecific (such as fever and tachycardia), many patients in critical care settings have underlying diseases with comorbidities that may produce symptoms common to sepsis.2d Recognizing those patients most at risk—and responding at the right time with the proper treatment before symptoms become deadly—is the conundrum that sepsis presents.

How Cognitive Agents Can Assist the “Natural Language” Clinical Conversation

Data researchers conducted a study of over 100,000 ICU patients. Their work demonstrated that when methods of machine learning and natural language processing (NLP) were applied retrospectively to information already routinely collected in electronic health records - and enriched by insights from the critical care physicians' notes – algorithms could significantly improve a forward-looking prediction model for sepsis mortality in the ICU.  In the study, those results were then compared with other approaches that used only abnormal vital signs and laboratory values to predict outcomes.

Natural language processing refers to any interaction between computers and human language. In a related way, a chatbot – also known as a cognitive agent – leverages the same natural language processing for the purpose of building intuitive human-computer interface.  

Taking the concept one step further, in order to assist clinicians in the process of identifying and responding to patients most at risk for sepsis deterioration, the cognitive agent (or chatbot), might be able to perform several tasks. For example,

  • Understand the questions the clinicians ask about the patient data (intent)
  • Apply a healthcare context of the dialogue to help “connects the dots” of data already within the EMR
  • Provide relevant insights that might help inform the clinician’s own critical thinking for the patient's treatment

Envision an Intelligent, Interactive Partner in Managing the Signs of Sepsis

In the future world of the ICU or emergency department, may evolve to an ongoing exchange between the clinician and cognitive agent with natural language processing. This “virtual” interaction might address clinicians’ questions in context and provide rapid access to personalized patient information. When there is little time to waste—and mounds of data to sift through— cognitive agents may be designed to compare or relate lab results to clinical events and notify clinicians of the patient’s status. This would require data insights to be shared in a meaningful and clinically relevant way.

Other potential chatbot “skills” may include the ability to understand multiple inputs, and answer questions based on context. A cognitive agent may someday become an integral component of treatment teams; listening, answering questions and providing key information rather than merely alerting the clinician of a change in the patients’ status.

Beyond weather forecasts and music requests of today’s virtual assistants, it may be that clinicians equipped with clinical chatbots may more readily take informed and appropriate actions to act on patient deterioration.

References

  1. Prevalence, Underlying Causes, and Preventability of Sepsis-Associated Mortality in US Acute Care Hospitals, JAMA Critical Care Medicine, February, 2019. https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2724768 Accessed September 13, 2019.
  2. Validation of Prediction Models for Critical Care Outcomes Using Natural Language Processing of Electronic Health Record Data, Original Investigation, JAMA Critical Care Medicine, December 2018. https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2719128 Accessed September 13, 2019.

 

Access related articles, videos, webinars at PRECISELY, by GE Healthcare and Roche