Development and evaluation of human-friendly explanations for sepsis early-warning models

Project SummaryThe goal of this Phase I project is to render a proprietary artificial intelligence algorithm for sepsis prediction developed by Patchd, Inc., compatible with explainability capabilities, providing clinicians underlying reasons for increased disease risk, as opposed to a “black box” risk prediction alone. The algorithm is additionally innovative in that it imputes sepsis risk based on vital sign data gathered from wearable devices, providing constant real-time surveillance of high-risk patients at home without requiring in-patient monitoring. Sepsis has a high mortality rate of approximately 40% and often occurs in immunocompromised patients such as those undergoing cancer chemotherapy, organ transplant surgery, or treatment for autoimmune conditions. Catching the condition early remains the best way to reduce mortality as well as prevent the long-term health effects associated with post-sepsis syndrome, which occurs in up to 50% of sepsis patients. Artificial intelligence (AI) represents an intriguing means of identifying sepsis early. By compiling vital sign data from high-risk patients and analyzing any changes in real-time, machine learning algorithms have shown promise in identifying at-risk patients up to 8 hours before sepsis onset in hospital settings. While these predictive algorithms demonstrate the power of AI in patient monitoring, most high-risk patients are treated via out-patient care without constant vital sign monitoring. Patchd, Inc. has developed an AI algorithm for sepsis prediction which employs vital sign data generated using wearable devices, such as watches or patches which record patient data in real-time. Thus far, the algorithm has been shown to improve sepsis prediction accuracy and provide earlier warnings for sepsis onset compared to other scoring methods. Recently however, bioethicists and regulators have called for AI algorithms to pair their predictions with supporting explanatory information providing both clinicians and patients more transparency when considering treatment options. During this Phase I program, Patchd will employ its current algorithm architecture and add so-called explainability functions, revealing to the user which specific vital sign dynamics impacted a given prediction. Vital sign data will be analyzed for local explanation functionality using both kernel and deep SHapley Additive exPlanations (SHAP). Data will also be analyzed to evaluate the significance of data points over time, using attentional mechanisms to understand the importance of vital sign changes over time. Primary human data sets, comprising true-positives, true-negatives, false-positives, and false-negatives, will be tested using the algorithm to evaluate both its predictive and explainabilty capabilities. Given that sepsis impacts 1.7 million Americans each year, the Patchd approach to explainable sepsis prediction will address a large and growing market opportunity.