Estimating Treatment Effects for Time-to-Treatment Antibiotic Stewardship in Sepsis

Sepsis is a life-threatening condition with high in-hospital mortality rate. The timing of antibiotic (ATB) administration poses a critical problem for sepsis management. Existing work studying antibiotic timing either ignores the temporality of the observational data or the heterogeneity of the treatment effects. In this paper, we propose a novel method to estimate TreatmenT effects for Time-to-Treatment antibiotic stewardship in sepsis (T4). T4 estimates individual treatment effects (ITEs) by recurrently encoding temporal and static variables as potential confounders, and then decoding the outcomes under different treatment sequences. We propose a mini-batch balancing matching that mimics the randomized controlled trial process to adjust the confounding. The model achieves interpretability through a global-level attention mechanism and a variable-level importance examination. Meanwhile, we incorporate T4 with uncertainty quantification to help prevent overconfident recommendations. We demonstrate that T4 can identify effective treatment timing with estimated ITEs for antibiotic stewardship on two real-world datasets. Moreover, comprehensive experiments on a synthetic dataset exhibit the outstanding performance of T4 compared to the state-of-the-art models on ITE estimation.