2.3. Extraction and coding of variables related to cardiometabolic function and inflammation

Custom-built natural language processing (NLP) software (Cardinal, 2017) was used to extract numerical cardiometabolic and inflammatory marker data from unstructured text, e.g. medical notes. Blood results were included only if they were recorded around the start date of the EIS involvement (±100 days), and if more than one test was recorded within this time frame, the closest to the EIS start date was chosen.

Cardiometabolic markers were available for a variable proportion of the sample within the first 100 days from EIS involvement start, and included systolic/diastolic blood pressure (mm/Hg), body mass index (BMI) (kg/m²), glucose (mmol/L), glycated haemoglobin (HbA1c) (mmol/mol), and a full lipid profile (total cholesterol (mmol/L) and triglycerides (mmol/L)). Some of the NLP tools for the extraction of cardiometabolic marker data were developed specifically for this study. Accuracy and reliability for all cardiometabolic markers were satisfactory, as measured by recall (probability of retrieving a record given it was relevant; >0.75 for all) and precision (probability of a record being relevant, given it was retrieved; >0.90 for all) statistics (see (Osimo et al., 2018) for how these were calculated). High-density lipoprotein (HDL) cholesterol levels were not always recorded, so we used the Friedewald equation (Tremblay et al., 2004) to derive this parameter from other lipid levels in patients lacking a direct measure.

We extracted measures of CRP and differential cell counts using methods described previously (Osimo et al., 2018). In keeping with previous analyses and data availability, CRP levels were categorised as follows: ≤3 mg/L (“non-inflamed”); >3 CRP ≤ 10 mg/L (“low-grade inflammation”); >10 mg/L (“suspected infection”). See also the Supplementary Methods.

In addition, we calculated NLR and basophil to lymphocyte (BLR) ratios.