Discordance between actual care and recommendations by a reinforcement learning model for sepsis

Abstract

Fluid and vasopressor management of sepsis requires a complex set of sequential treatment decisions tailored to individual responses. We examined the frequency and implications of disagreement between reinforcement-learning (RL), using the AI Clinician as a test example, and actual clinical decisions. A novel discordance score was defined as the Euclidean distance between clinical team actions and RL-recommended treatments (range 0 to 5.66), across 25 treatment options of vasopressors and intravenous (IV) fluids. Among 30,678 ICU patients with sepsis cared for across 309,415 time-epochs, discordance was more common among epochs of patients with SOFA > 6 than those with SOFA < 6, and for intravenous fluid choices than for vasopressors. After multivariable adjustment, greater discordance was associated with 90-day mortality (adjusted odds ratio (OR) 1.11, 95%CI: 1.05,1.18) in the SOFA strata > =6 but not for SOFA < = 4 (adjusted odds ratio (OR): 1.02, 95%CI: 0.97,1.08).