Development and validation of an algorithm to identify severe sepsis onset from electronic medical records

Objective: To develop and evaluate an automated algorithm to identify sepsis onset from ethe electronic health records (EHR), referred to as time-zero (t₀), to enable more accurate surveillance, quality improvement, and training related to SEP-1 aligned care. Materials and methods: We developed an algorithm to construct a comprehensive timeline of systemic inflammatory response syndrome (SIRS) criteria and organ dysfunction (OD) using structured data, and documentation of infection (DOI) using both structured data and unstructured clinical notes. The algorithm scans each timeline to detect the co-occurrence of SEP-1 components within a 6-hour window to determine t₀. Algorithm performance was assessed using 2,030 manually abstracted adult sepsis cases from a large academic health system in southeast Michigan. Results: Using a subset of 516 abstracted cases, we show that including clinical notes achieves higher concordance of DOI with abstractors t₀ (41.9%) compared to using antibiotic (27.9%) or culture order proxies (34.7%). Combining all three sources achieves the highest DOI concordance (44.4%). On average, the algorithm DOI time was significantly earlier than abstractors (mean: -0.30 h, 95% Cl: -0.51 to -0.09) across all 2,030 cases, resulting in a significantly earlier t₀ (mean: -0.51 h, 95% CI: -0.65 to -0.36). Discussion: Automated approaches to analyze EHR data offer a scalable framework for SEP-1 monitoring, research, and quality improvement. Conclusion: Incorporating unstructured clinical notes improves detection of infection suspicion and enhances concordance with manual abstraction of sepsis onset.