Medical-AI trial: AI-assisted "blind-sweep" obstetric ultrasound by non-specialists — antenatal screening in rural DR Congo (SPAQ E-con AI) (NCT07677670)

1Prenatal ultrasound and the shortage of expertise

In antenatal care, ultrasound is central to knowing gestational age and the condition of the fetus and placenta, but acquiring and reading it requires trained specialists. Where specialists and equipment are scarce, the screening itself becomes hard to access. This is where the blind-sweep method draws interest: a non-specialist moves the probe in a set pattern to collect images, and AI handles the interpretation. Because a specialist does not need to aim the scan, the pool of operators can widen.

2Supporting blind sweeps by non-experts

The SPAQ E-con AI that this study evaluates supports exactly this non-specialist blind-sweep obstetric ultrasound with AI. Per the registry summary, the study is set in rural Democratic Republic of the Congo, where trained non-specialist health workers perform the scans, and it evaluates diagnostic accuracy and implementation feasibility.

Primary outcomes are gestational-age mean absolute error (trimester 2, 14–27 weeks; trimester 3, 28–36 weeks) and the calibration of the AI confidence (expected calibration error and Brier score). The accuracy reference is manual measurement by a reference reader (early ultrasound preferred; manual BPD, biparietal diameter, if unavailable; last menstrual period is not used).

It is multicenter and prospective, with a Hybrid Type 1 design that examines effectiveness and field implementation together, and it follows FDA PCCP (a pre-specified post-market change plan) and STARD-AI guidance, including a pre-specified model-update (a Batch 2 cut). The target is about 1,430 (IRB-approved ceiling 3,000), with early termination allowed once pre-specified analysis thresholds are met.

3Changing who holds the probe

The point of blind sweep is less accuracy than widening who can perform the imaging.

The primary endpoints are mean absolute error of gestational age and the calibration of the AI's confidence. Placing calibration — how confidently the AI answered — among the primary endpoints, and following the FDA PCCP and STARD-AI frameworks, embodies a modern discipline for using medical AI responsibly in the field. The target is about 1,430 cases.