Abstract

Health data captured in Electronic health records (EHRs) have enabled the development of computational approaches to improve patient management and treatment, including early diagnosis of severe conditions such as sepsis. The validity of these efforts, however, largely relies on which sepsis definition is used and the quality of the underlying data. Here we tested different sepsis definitions to better understand how phenotyping approaches may impact the classification accuracy of sepsis prediction algorithms. To assess the extent to which sepsis definitions (dis)agree with each other, we have analised a large cohort of patients admitted to the ICU (over 22,000) from MIMIC-IV. Cases were classified as septic and non-septic using the Sepsis-3 definition as a standard and compared with different ICD-10-based sepsis phenotyping criteria. Most of administrative sepsis definitions agreed with each other when identifying positive sepsis cases. At the same time, we identified considerable disagreement between Sepsis-3 and administrative definitions. This discrepancy affected machine learning algorithms’ predictive performance. Two algorithms out of three built on Sepsis-3 outperformed models based on other phenotypes. Experiments demonstrate that phenotype definitions can significantly influence a predictive model performance. This highlights the importance of consistent and validated digital phenotyping criteria.