Accurate diagnostic coding of sepsis is important for surveillance, resource allocation, and health policy planning. Claims-based data such as ICD-10 codes are frequently used for epidemiologic monitoring, but their agreement with clinical criteria-based definitions is uncertain. The study evaluated a fully automated electronic health record (EHR)-based algorithm that applies Sepsis-3 clinical criteria to classify sepsis, comparing algorithm-derived classifications with diagnostic coding.
The primary objective was to assess concordance between ICD-10 diagnostic coding of sepsis and classification using a previously validated fully automated EHR-based algorithm that applies Sepsis-3 criteria. The study also examined time trends and the association of EHR tool introduction with coding patterns.
This was an observational study including adult in-hospital admissions. Study periods differed by site: 2016–2024 at Karolinska University Hospital and 2022–2024 at three hospitals in Region Västerbotten, Sweden. A common study period (2022–2024) was used for direct comparison across sites.
The Sepsis-3 algorithm identified sepsis when there was suspected infection combined with an increase in Sequential Organ Failure Assessment (SOFA) score of ≥2 points. Diagnostic codes for sepsis were extracted and categorised as explicit or specific. Explicit and specific code examples provided in the source included R65.1 and R57.2; the study grouped codes accordingly for analysis.
Concordance between the algorithm classification and coding was evaluated using descriptive statistics and kappa statistics. Time trends in coding and the effect of introducing an automated SOFA calculator within the EHR at Karolinska were also analysed.
During the common study period (2022–2024), the dataset included 174,343 admissions. Using the Sepsis-3 algorithm, sepsis was classified in 11.4% of admissions at Karolinska University Hospital and 6.2% of admissions in Västerbotten.
By contrast, sepsis identified by diagnostic coding was markedly lower. At Karolinska, explicit codes were present in 2.5% of admissions and specific codes in 1.4%. In Västerbotten, explicit codes occurred in 1.2% and specific codes in 0.6% of admissions.
Agreement between the EHR-based algorithm and code-based identification was low. Kappa statistics reported in the source were:
These kappa values indicate poor concordance between algorithm-detected sepsis using Sepsis-3 criteria and sepsis captured by diagnostic coding in both settings.
Across the Karolinska study period 2016–2024, the source reports that the introduction of an automated SOFA calculator in the EHR system was associated with increased usage of specific sepsis codes. Despite this increase, coding frequencies remained considerably lower than algorithm-based classifications of sepsis. The abstract notes a temporal association but does not provide additional quantitative details on the magnitude of the coding change beyond the broad observation.
The authors conclude that fully automated EHR-based algorithms applying Sepsis-3 criteria allow for data-driven sepsis surveillance and may support more reliable diagnostic coding practices. The algorithm identified substantially more sepsis cases than claims-based ICD-10 coding and showed low concordance with diagnostic codes, suggesting that reliance on routine coding alone may substantially undercount sepsis incidence compared with clinical-criteria-based automated methods.
Automated identification could inform surveillance efforts, help allocate resources more accurately, and prompt improvements in clinical documentation and coding workflows. The reported association between an EHR SOFA calculator and increased specific code usage suggests integration of clinical decision support tools may influence coding behavior, although coding still lagged behind algorithm detection.
The abstract summarizes study design, methods, key results, and a temporal association with the EHR SOFA tool. Specific limitations such as chart review validation, sensitivity/specificity metrics of the algorithm versus manual adjudication, potential differences in case-mix across sites, and detailed quantitative time-trend data were not reported in the abstract. Any additional methodological details, subgroup analyses, or external validity considerations were not provided in the source text quoted here.
Overall, the study indicates that a fully automated EHR-based Sepsis-3 algorithm captures more sepsis cases and shows low agreement with ICD-10 coding, and that EHR tool implementation may increase code usage though coding remains far below algorithm-derived incidence.