This study was a secondary data analysis of the data of 65 366 participants in MIMIC-IV between 2008 and 2022. MIMIC-IV database is a comprehensive and publicly accessible repository that encompasses data on all patients admitted to the Beth Israel Deaconess Medical Center between 2008 and 2022. It captures extensive information, including patient length of stay, laboratory tests, medication treatments, vital signs and other pertinent details. To ensure patient privacy protection, personal information has been deidentified through the utilisation of random codes in place of patient identification. Consequently, informed consent and ethical approval from patients are not required.15 In the present study, participants aged ≥18 years old, information for early ondansetron usage assessment and sepsis assessment were included. Those with the length of ICU stay less than 48 hours, participants diagnosed as sepsis within first 48 hours after ICU admission and those without information on whether in-hospital mortality occurred were excluded.

Age, gender (women or men), race (White, Black, other or unknown), insurance (Medicare, Medicaid or other), marital status (married, unmarried or unknown), total urine output during the first 48 hours, diabetes (yes or no), trauma (yes or no), malignancy (yes or no), chemotherapy (yes or no), radiotherapy (yes or no), surgery (yes or no), Sequential Organ Failure Assessment (SOFA), Charlson Comorbidity Index (CCI), SIRS, weight, heart rate (beats per minute), mean arterial pressure (mm Hg), respiratory rate (breaths per minute), temperature (℃), white blood cell (WBC; K/µL), platelet (K/µL), haemoglobin (g/L), red blood cell distribution width (RDW), haematocrit (%), serum creatinine (mg/dL), prolonged prothrombin time (s), partial thromboplastin time (PTT; s), blood urea nitrogen (BUN; mg/dL), glucose (mg/dL), serum calcium (mmol/L), anion gap (mmol/L), machine ventilation (yes or no), vasopressor (yes or no), antibiotics (yes or no), insulin (yes or no), total ondansetron dose, drug frequency (no usage, once or ≥twice) and route (no usage, intravenous, oral, oral and intravenous or other including intravenous and sublingual (SL), oral and SL, and SL) were analysed.

Antibiotics included Ampicillin, Nafcillin, Oxacillin, Penicillin, Piperacillin, Azithromycin, Vancomycin, Clindamycin, Daptomycin, Erythromycin, Tobramycin, Gentamycin, Gentamicin, Amikacin, Ciprofloxacin, Levofloxacin, Moxifloxacin, Amphotericin, Doxorubicin, Cefazolin, Cefepime, Ceftazidime and Ceftriaxone. All the antibiotics were used within 48 hours after ICU admission. Surgery data were extracted from ‘procedures4’ table in the database. The laboratory data were extracted according to the corresponding item-id in the Labevents table of the database, and we used the first measurements within 48 hours after admission to the ICU.

The normal measurement data were expressed as mean±SD, and the comparison was performed by t-test. The non-normal measurement data were expressed as median and quartiles (M (Q1, Q3)), and Mann-Whitney U test was used to compare groups. The enumeration data were described as the number and ratio of cases (n (%)), and group comparisons were conducted by χ2 test or Fisher’s exact probability method. Missing values were shown in online supplemental table 1. Variables with missing values <20% were dealt with random forest imputation, and those with missing values ≥20% were deleted. The data before and after missing values imputation were compared in online supplemental table 2. The covariates selected by univariate cox stepwise regression model, along with those having the smallest Akaike’s information criterion (AIC), were regarded as confounding factors using stepwise Logit function in R StepReg package and setting select= ‘AIC’. The association between early ondansetron use and the risk of sepsis or in-hospital mortality was analysed by multivariate cox regression model. Estimates were expressed as HR and 95% CI. SAS9.4 software (SAS Institute Inc., Cary, North Carolina, USA) was used for statistical analysis, and forest plot was drawn using R version 4.2.3 (2023-03-15ucrt (R Foundation for Statistical Computing, Vienna, Austria)). P<0.05 was considered statistically significant.