The PRECISIO Pipeline: Artificial Intelligence–Informed Precision IV Fluids in Sepsis

Randomized controlled trials in patients with sepsis have found no differences in outcomes between liberal and restrictive IV fluid resuscitation strategies. In this study, we aimed to use counterfactual inference and machine learning to predict optimal IV fluid strategies for individual patients with sepsis in the emergency department (ED).

We developed a counterfactual inference decision-support pipeline (PRECISIO), using observational data from Sepsis-3 defined encounters at two large academic hospitals, between January 1, 2016, and December 31, 2018. This pipeline used an XGBoost architecture to predict in-hospital death, adjusting for confounding, under restrictive (<2L) or liberal (≥2L) fluid strategies (during first 6 hours following sepsis diagnosis). We used an 80/20 data split for model development and validation, balanced by treatment and outcome. Model input variables included demographics, comorbidities and clinical variables (e.g. labs, vital signs). PRECISIO recommended the fluid strategy with the lowest predicted risk of death. We defined clinical concordance as agreement between observed care and the PRECISIO-recommended strategy. We performed model-based policy evaluation to estimate the average treatment effect (ATE), defined as the absolute risk difference between usual care and PRECISIO-recommended care.

We included 8,039 ED encounters with sepsis (56.8% male, median age 61years [IQR: 49-73]). In the overall cohort, the rate of in-hospital death was 7.1%, with 28.4% (n=2,283) receiving liberal treatment. Model performance was similar for predicting in-hospital mortality in the liberal (AUC: 0.85) and restrictive (AUC: 0.82) models. In the validation cohort, 53% (852/1,608) of patients received concordant care; observed in-hospital death was lower among patients receiving concordant vs discordant care (5.5% vs 9.1%, p=0.003). In our policy evaluation analysis, we estimated that mortality would have dropped from 7.1% (usual care) to 4.3% (PRECISIO) had all patients received PRECISIO-recommended care (mortality risk difference, 2.9% [95% CI 1.7% to 4%]; relative mortality reduction, 40% [95% CI 24% to 56%], p<0.001).

Counterfactual inference models may help predict heterogeneous treatment responses and inform precision IV fluid resuscitation strategies for patients with sepsis.