Definition of septic shock

In order to retain algorithm comparability between LiSep LSTM and TREWScore, the definitions for sepsis, severe sepsis, and septic shock were based on the same criteria as Henry et al.¹⁴. In their study, sepsis was defined by the presence of any two of the SIRS criteria (shown below) in combination with the suspicion of an infection²⁶:

Body temperature: <36 °C or >38 °C

Heart rate: >90 BPM

Respiratory rate: >20 BPM or arterial CO₂ pressure (PaCO₂) < 32 mmHg

White blood cell count: <4,000/mL or >12,000/mL

Severe sepsis was defined as the presence of sepsis in combination with sepsis-related organ dysfunction. Sepsis-related organ dysfunction is specified in the surviving sepsis campaign guidelines as the presence of any of the symptoms listed below²⁷:

Systolic blood pressure: <90 mmHg

Blood lactate: >2.0 mmol/L

Urine output: <0.5 mL/kg over the last two hours despite adequate fluid resuscitation

Creatinine: >2.0 mg/dL without the presence of chronic dialysis or renal insufficiency as indicated by ICD-9 codes V45.11 or 585.9

Bilirubin: >2.0 mg/dL without the presence of chronic liver disease and cirrhosis as indicated by ICD-9 code 571 or any of its sub-codes

Platelet count: <100,000/μL

International normalised ratio (INR): >1.5

Acute lung injury with arterial O₂ pressure (PaO₂)/fraction of inspired oxygen (FiO₂) < 200 in the presence of pneumonia as indicated by an ICD-9 code of 486

Acute lung injury with PaO₂/FiO₂ < 250 in the absence of pneumonia

Septic shock was defined by the presence of severe sepsis and hypotension (systolic blood pressure <90 mmHg) despite adequate fluid resuscitation, defined as a fluid replacement over the past 24 hours ≥ 20 mL/kg or a total fluid replacement ≥1200 mL.

All patients were declared sepsis-negative or diagnosed with sepsis, severe sepsis, or septic shock on an hourly basis using these criteria. Patient and diagnosis statistics for our data set can be found in the ^{Supplementary Materials}.

When determining whether or not a patient should be considered septic shock-positive or negative Henry et al. considered an effect they called “censoring”. The reasoning was that when a patient receives treatment typical of that used to treat septic shock (e.g. fluid resuscitation), that specific patient’s condition gets censored in one of two ways. If the patient later fulfils the criteria for septic shock, the onset of the condition may have been delayed by the treatment. On the other hand, if the patient never fulfils the criteria for septic shock, the condition may have been pre-emptively treated. Henry et al. recognised that this might cause problems when fitting their model and they took special measures to deal with censored patients¹⁴. However, one of the benefits of LSTM networks is that they might be able to learn to recognise effects like censoring on their own, given that there are enough examples of it in the data set. Taking this into account, we decided not to explicitly consider censoring effects in our study.