Development and Validation of the VIOSync Sepsis Prediction Index: A Novel Machine Learning Model for Sepsis Prediction in ICU Patients

Abstract

Background: Sepsis is the third leading cause of death worldwide and the main cause of in-hospital mortality. Despite decades of research, sepsis remains a major challenge faced by patients, clinicians, and medical systems worldwide. Early identification and prediction of patients at risk of sepsis and adverse outcomes associated with sepsis are critical. In this work, we aimed to develop an artificial intelligence algorithm that can predict sepsis early. Materials and Methods: We developed a predictive model for sepsis using data from the Physionet Cardiology Challenge 2019 ICU database. Our cohort consisted of adult patients who were admitted to the ICU. Sepsis diagnoses were determined using the Sepsis-3 criteria. The model, built with the XGBoost algorithm, was designed to anticipate sepsis prior to the appearance of clinical symptoms. An internal validation was conducted using a hold-off test dataset to evaluate the AI model's predictive performance. Results: We have developed the VIOSync Sepsis Prediction Index (SPI), an AI-based predictive model designed to forecast sepsis up to six hours before its clinical onset, as defined by Sepsis-3 criteria. The AI model, trained on a dataset comprising approximately 40,000 adult patients, integrates variables such as vital signs, laboratory data, and demographic information. The model demonstrated a high prediction accuracy rate of 97%, with a sensitivity of 87% and a specificity of 98% in predicting sepsis up to 6 hours before the onset. When compared to the established qSOFA score, which has a specificity of 89% for sepsis prediction, our VIOSync SPI algorithm significantly enhances predictive reliability, potentially reducing false positive rates by a factor of 5.5. Conclusions: The VIOSync SPI demonstrated superior prediction performance over current sepsis early warning scores and predictive algorithms for sepsis onset. To validate the generalizability of our method across populations and treatment protocols, external validation studies are essential.