Interpretable machine learning-based real-time sepsis diagnosis

Sepsis is among the top global health issues, leading to serious health risks with high mortality and morbidity. Recent statistics revealed around 50 million Sepsis cases and 11 million deaths annually. Sepsis arises due to the body’s abnormal response to an infection, leading to severe health conditions such as organ failure. As such, failure to detect Sepsis timely may be fatal. However, early detection of Sepsis onset remains a great challenge. This paper leverages lightweight models and develops an efficient approach and system for accurate, easily affordable, and early diagnosis of Sepsis. The diagnosis uses seven non-invasive vital signs, namely, heart rate, body temperature, systolic, diastolic, and mean arterial blood pressure, oxygen saturation level, and end-tidal carbon dioxide. The proposed system will improve public health and relieve the burden on the ICU by preventing severe Sepsis cases. Our proposed system could also be implemented using low-cost sensors and interfaces, making the system suitable for remote monitoring with the Internet of Things (IoT). Furthermore, a novel method is adopted for handling class imbalance in the dataset. Additionally, Shapley Additive Explanations (SHAP) and Locally Interpretable Model-agnostic Explanations (LIME) were leveraged to gain insight into the system. The system was implemented and evaluated using our evaluation platform consisting of basic sensors and a Raspberry Pi as a point-of-care Sepsis detection tool proof of concept. Our model obtained a utility score of around 46% (set by the Physionet challenge, the database source). The model also achieved better performances of 86.49% accuracy and an AUROC of 0.94 compared to the related Sepsis detection works.