Artificial Intelligence Inference of Community Clinic Maternal Sepsis Risk Under Unmeasured Traditional Birth Attendant Referral Delay: A Healthcare Instrumental Variable Evaluation

Community clinic algorithms that infer maternal sepsis risk from routine antenatal and intrapartum records cannot observe the delay between a traditional birth attendant's decision to refer and the woman's arrival at the clinic. This unmeasured referral delay is a confounder of the clinic-level risk model, because it is associated both with the severity of infection on arrival and with the probability that the woman is recorded as a clinic patient at all. The strategy treats the distance from the woman's household to the nearest traditional birth attendant as an instrument for observed referral delay, conditional on clinic catchment and household covariates. Power was estimated across 2000 design cells over cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design requires 54 clusters to detect a 5 percentage point effect at 80% power. The instrumental variable design is the appropriate inferential form for this problem, and that the before-after-control-impact estimator is the appropriate evaluation design for the clinic-level intervention that the risk model supports.