The Causal Impact of Natural Language Processing-Driven Clinical Decision Support on Sepsis Mortality in England: An Augmented Synthetic Control Analysis of NHS Trust-Level Data

Background: Sepsis remains a leading cause of preventable hospital mortality in England, with NHS England reporting over 48,000 sepsis-related deaths annually. Natural language processing (NLP)-driven clinical decision support systems (CDSS) have been deployed in several NHS Trusts to enable automated early detection of sepsis from unstructured clinical notes, yet causal evidence of their effectiveness at the hospital level remains limited. Objective: To estimate the causal effect of implementing NLP-driven CDSS for sepsis detection on 30-day in-hospital sepsis mortality rates across NHS Trusts in England, using the augmented synthetic control method (ASCM) to account for selection bias and time-varying confounders. Methods: We conducted a quasi-experimental study using Hospital Episode Statistics (HES) and NHS Digital mortality data from 2017 to 2024 across 142 NHS acute Trusts. Twelve Trusts that implemented NLP-CDSS for sepsis detection between April 2020 and March 2022 constituted the treated group. The ASCM was employed to construct a weighted combination of donor Trusts that closely approximated the pre-intervention trajectory of each treated Trust. Ridge-augmented bias correction was applied to improve pre-intervention fit. We estimated average treatment effects on the treated (ATT) using staggered adoption designs with heterogeneity-robust estimators. Inference was conducted via conformal permutation-based p-values and placebo tests. Results: NLP-CDSS implementation was associated with a statistically significant reduction in 30-day sepsis mortality of 3.7 percentage points (95% CI: -6.1 to -1.3; p = 0.003), corresponding to a 14.2% relative risk reduction. Effects emerged approximately six months post-implementation (ATT at 6 months: -2.1 pp, 95% CI: -4.0 to -0.2) and strengthened over time (ATT at 24 months: -5.3 pp, 95% CI: -8.4 to -2.2). Placebo tests across 130 donor Trusts confirmed the implausibility of chance findings (placebo p = 0.008). Heterogeneity analysis revealed larger effects in Trusts with higher baseline mortality ({beta} = -0.41, SE = 0.12, p = 0.001) and those that integrated NLP-CDSS with existing electronic health record systems ({beta} = -1.8 pp, SE = 0.6, p = 0.003). Conclusions: NLP-driven CDSS deployment for sepsis detection was associated with clinically meaningful and statistically robust reductions in hospital sepsis mortality across NHS Trusts. Augmented synthetic control methods provide rigorous causal evidence in the absence of randomized trials, supporting the broader adoption of NLP-enabled surveillance systems in national healthcare settings.