Study population

This study utilized data from the Medical Information Mart for Intensive Care (MIMIC)-IV database, a large, publicly available dataset containing de-identified ICU patient records from Beth Israel Deaconess Medical Center¹⁵. We used MIMIC-IV version 3.0, which includes data from 2008 to 2022, covering over 94,000 patient records with detailed information on demographics, vital signs, laboratory results, medications, procedures, and diagnoses.

The study focused on patients admitted to the ICU who experienced septic shock and received fluid resuscitation with either normal saline or Ringer’s lactate, or a combination of both. Patients with underlying end-stage kidney disease (ESKD) or those already receiving kidney replacement therapy at the onset of sepsis were excluded from the analysis. Sepsis was defined according to Sepsis-3 criteria, which included a Sequential Organ Failure Assessment (SOFA) score increase of ≥ 2 points and suspected or confirmed infection¹⁶. Suspicion of infection was identified through positive blood cultures or antibiotic administration by clinicians. Septic shock was defined as sepsis patients presenting with mean arterial pressure (MAP) ≤ 65 mmHg, systolic blood pressure (SBP) ≤ 90 mmHg, vasopressor use, or lactate levels ≥ 2 mmol/L¹⁶.

For the training of a deep learning-based causal inference model, the dataset was split into training (80%), validation (5%), and test (15%) datasets.

Exposures, outcome, and variables

The exposure was defined based on the type of fluid administered within a 12-hour window before or after the time sepsis was identified, which was marked by a SOFA score increase of two points or more. Fluids were categorized into six groups: (1) normal saline only, (2) Ringer’s lactate only, (3) normal saline combined with Ringer’s lactate, (4) Ringer’s lactate combined with albumin, (5) normal saline combined with albumin, and (6) normal saline, Ringer’s lactate, and albumin administered together. The 12-hour window was selected to account for the possibility that hemodynamic instability may precede the recorded rise in the SOFA score and to accommodate variations in the duration of fluid resuscitation among healthcare providers.

The primary outcomes were in-hospital mortality and kidney outcomes. Kidney outcomes were defined as either a doubling of serum creatinine—relative to the baseline creatinine value measured closest to sepsis onset within the first 24 h of onset—occurring more than 24 h after sepsis onset, or the initiation of kidney replacement therapy more than 24 h after sepsis onset^17,18.

We included various variables in the analysis, including demographic data, laboratory results, underlying medical history, and patient status. Patients with missing albumin and creatinine values were excluded from the analysis. Among the remaining patients, variables with more than 10% missing values were excluded from the final set of analyzed variables. Missing data were handled using Multivariate Imputation by Chained Equations to reduce bias¹⁹. Variables excluded due to missing data included serum protein, C-reactive protein, and height data (Supplementary Table S1). The final included variables comprised demographic data (age, sex, weight), initial vital signs (systolic blood pressure [SBP], diastolic blood pressure [DBP], mean arterial pressure [MAP], heart rate, peripheral oxygen saturation [SpO₂], temperature), and initial patient status (fraction of inspired oxygen [FiO₂], SOFA score, Glasgow Coma Scale [GCS] score, mechanical ventilator). Comorbidities were defined using International Classification of Diseases [ICD] diagnostic codes, including chronic kidney disease [CKD], myocardial infarction [MI], congestive heart failure [CHF], peripheral vascular disease [PVD], cerebrovascular disease, chronic pulmonary disease, chronic liver disease, diabetes, hypertension, cancer, and metastatic cancer. The focus of infection was also identified using ICD codes, including lung infection, peritoneal infection, abdominal infection, genito-urinary infection, septicemia, heart infection, fungal infection, bacteremia, central nervous system (CNS) infection, soft tissue infection, and upper respiratory infection²⁰. Laboratory variables included albumin, white blood cell count [WBC], hemoglobin, platelets, calcium, pH, anion gap, bicarbonate, blood urea nitrogen [BUN], creatinine, Chronic Kidney Disease Epidemiology Collaboration [CKD-EPI] estimated glomerular filtration rate [eGFR], sodium, potassium, chloride, prothrombin time international normalized ratio [PT INR], activated partial thromboplastin time [aPTT], and lactate. Medication variables included the infusion rates of norepinephrine, dopamine, epinephrine, vasopressin, and dobutamine.

Development of deep learning-based causal inference model

The deep learning-based causal inference model used in this study is based on Dragonnet, which was originally designed for binary treatment scenarios²¹. The model was modified to handle multiple treatments by incorporating the Targeted Regularization (TARNet) technique²². Dragonnet typically addresses binary treatment causal inference by predicting outcomes and applying adjustments (epsilon) to reduce model bias for two treatment options: receiving or not receiving a treatment. For this study, the model was adapted to handle multiple treatment options, allowing for additional regularization for each treatment. (Supplementary methods)

The treatment variable was defined based on the six types of fluid combinations described earlier. For the analysis of treatment effects, normal saline was chosen as the baseline treatment. The model was trained using early stopping based on the validation dataset, and hyperparameter tuning was performed by assessing the area under the receiver operating characteristic curve (AUROC) in the validation dataset. Models were trained separately for in-hospital mortality and kidney outcomes. Although our model is intended to estimate treatment effects rather than serve solely as a predictive tool, we assessed its fit and discrimination on the test dataset using AUROC, accuracy, and F1score. Calibration plots were analyzed to evaluate the model’s calibration performance.

Using the trained model, we calculated the treatment effect of each fluid combination on in-hospital mortality and kidney outcomes, comparing them to normal saline alone. The treatment effect was quantified as the probability of outcome occurrence in the overall population. Through this process, the average treatment effect (ATE) was calculated, which represents the mean of the treatment effects. To compare the magnitude of the treatment effects between different fluid combinations, t-tests were performed.

Logistic regression analysis to identify patient characteristics associated with better outcomes for ringer’s lactate or albumin infusion

To identify patient characteristics associated with better outcomes for Ringer’s lactate or albumin infusion, multivariable logistic regression was used. The treatment effects on in-hospital mortality and kidney outcomes were calculated for the total population using the respective models, comparing to normal saline only. Two fluid comparisons were made using treatment effects: normal saline only versus Ringer’s lactate only, and normal saline only versus normal saline combined with albumin. The characteristics analyzed included demographic data (age, sex, weight), initial vital signs (MAP, heart rate, SpO₂, temperature), and patient status (SOFA score, GCS score, FiO₂, mechanical ventilator). Additional comorbidities included CKD, MI, CHF, PVD, cerebrovascular disease, chronic liver disease, diabetes, hypertension, and cancer. The analysis also considered infection sites, including lung infection, peritoneal infection, abdominal infection, genito-urinary infection, septicemia, heart infection, fungal infection, bacteremia, CNS infection, soft tissue infection, and upper respiratory infection. Laboratory results included albumin, WBC count, hemoglobin, platelets, calcium, pH, anion gap, bicarbonate, eGFR, sodium, potassium, chloride, PT INR, aPTT, and lactate. The analysis also factored in vasopressor use. Continuous variables were standardized, and standardized odds ratios were calculated. Since a large number of variables were used in the analysis, the p-value significance threshold was adjusted using the Bonferroni correction, rather than the standard cutoff of 0.05.