AI Blind-Sweep Ultrasound for Antenatal Screening by Non-Specialist Health Workers in Rural DR Congo

Overview

FS2 evaluates the diagnostic accuracy and implementation feasibility of an AI-assisted blind-sweep obstetric ultrasound (SPAQ E-con AI), operated by trained non-specialist health workers, for antenatal screening in rural Democratic Republic of the Congo. Primary outcomes are gestational age mean absolute error (Trimester 2 and Trimester 3) with 95% confidence intervals and AI confidence calibration. The reference standard is manual measurement by a reference reader (early ultrasound first; manual BPD if unavailable; last menstrual period is not used). Planned enrollment is approximately 1,380: Batch 1 (closed, 80 enrolled) and Batch 2 (approximately 1,300), within an IRB-approved ceiling of 3,000. Early termination is permitted upon achievement of pre-specified analysis-plan thresholds. The study is a multi-center prospective Hybrid Type 1 Effectiveness-Implementation design and includes a pre-specified adaptive model-update (Batch 2 cut) plan following FDA PCCP and STARD-AI guidance.

Study conduct is organised in Batches. Batch 1 (Centre de Sante CBCO, Kenge health zone, Kwango province) enrolled 80 participants, of whom 78 had complete datasets, and is closed. Batch 2 is planned at approximately 1,300 participants. The confirmatory analysis cut is approximately 130 to 160 cases acquired with a model frozen before collection, comprising a core set of 80 to 100, an early-gestation enrichment of 25 to 30, and an upper-gestational-age enrichment of 25 to 30. Enrolment may stop early once the pre-specified confidence-interval criteria for the two primary gestational-age outcomes are met. The registered enrolment ceiling of 3,000 is unchanged. Sites are facilities of the Kenge health zone, Kwango province, within the perimeter already approved in the protocol (up to 6 facilities plus 2 to 5 village outreach sites). Eligibility follows the protocol: pregnant women aged 18 years or older. Multiple pregnancies are eligible. Two sub-cohorts address the bounds of the gestational-age range. The early sub-cohort (crown-rump length 7 to 84 mm, approximately under 14 weeks) draws on the same source population as Cohort B-2 and is recruited through community health worker (RECO) village outreach and neighbouring health facilities; it is reported descriptively. The upper-bound sub-cohort (35 weeks or more) is a separate validation subset, recruited in parallel and analysed separately; sensitivity and specificity are reported, and positive and negative predictive values are not reported because prevalence in this subset is artificial. The registered Trimester 3 estimation window (28 to 36 weeks) is unchanged. In Batch 2, point estimates are reported for 28 to 34 weeks, and 35 to 36 weeks is handled as "to be confirmed / refer". This is an operational reporting restriction, not a change to the registered window. The AI model is identified and frozen before each Batch and recorded in a model and application version registry under a pre-determined change control plan. Results are reported by frozen version. AI inference is not used for clinical decision-making.