Desarrollo de una escala predictiva en pacientes con shock séptico refractario mediante un modelo híbrido de aprendizaje automático y regresión

Abstract

Objective: To develop a scale to predict refractory septic shock (SS) based on clinical variables recorded during initial evaluations of patients.

Methods: Multicenter retrospective study of data for patients with suspected infection registered in the Marketplace for Medical Information in Intensive Care (MIMIC-IV). These data were used for the development and internal validation of the refractory SS scale (RSSS). For external validation, we used retrospective data for 2 cohorts: 1) patients diagnosed with SS in an emergency department (ED cohort) whose data were registered in a Korean SS registry, and 2) patients diagnosed with SS in 6 hospital intensive care units (ICU cohort). A machine-learning automatic clinical scoring system (AutoScore) was used in the development phase. The performance of the RSSS in the validation cohorts was assessed with the area under the receiver operating characteristic curve (AUROC) for each. The primary outcome was the development of refractory SS within 24 hours of ICU admission. Refractory SS was defined by the need for a norepinephrine-equivalent dose greater than 0.5 µg/kg/min.

Results: We collected data for 29 618 patients from the MIMIC-IV registry, 3113 patients for the ED cohort, and 1015 for the ICU cohort. The RSSS had 6 predictors: serum lactate level, systolic blood pressure, heart rate, temperature, arterial pH, and leukocyte count. The scale's AUROCs were as follows: 0.873 (95% CI, 0.846-0.900) in the internal validation, 0.705 (95% CI, 0.678-0.733) in the ED cohort on arrival, 0.781 (95% CI, 0.757-0.805) in the ED cohort at the moment of diagnosing hypoperfusion or hypotension, and 0.822 (95% CI, 0.787-0.857) in the ICU cohort. Calibration was acceptable in all the cohorts.

Conclusions: The RSSS had adequate diagnostic accuracy in multiple cohorts of patients diagnosed in the ED and ICU.