ISSUE 025 · Oct 27, 2026 · 8 min read

What if an undated pregnancy did not have to wait for an ultrasound appointment?

Do not use an AI blind-sweep estimate as the care decision. Test a completion-guaranteed pathway that moves late-entry patients faster to definitive ultrasound and obstetric care.

01Clinical problem · Evidence

When dates are uncertain, the clock for needed care becomes uncertain too.

For patients entering prenatal care late or unsure of their last menstrual period, gestational age anchors screening, counseling, and high-risk referral. When scanners and skilled staff are distant, a referral may never become a completed appointment, leaving only an uncertain estimate.

02Technology stack · Proposal

Join handheld ultrasound, blind-sweep AI, and referral operations.

A trained nurse or midwife acquires protocolized sweeps; the model returns an estimate and quality state. The system does not translate that output into treatment.

It uses the signal to trigger risk review, a definitive scan slot, transport support, and follow-up for noncompletion.

03Research question · Hypothesis

Not just “is AI accurate?”—does the patient complete definitive testing?

Ask whether AI sweeps plus referral navigation increase completion of definitive obstetric ultrasound within seven days versus usual referral, without increasing missed urgent findings or inappropriate management decisions.

04First study · Proposal

Twelve safety-net clinics, 600 patients, a stepped-wedge cluster trial.

Enroll adults with uncertain dates and a clinically presumed singleton intrauterine pregnancy of 16–37 weeks. Control periods use usual referral; intervention periods add same-day blind sweeps, a quality gate, human review, a scan guaranteed within seven days, and transport support.

The primary outcome is definitive-ultrasound completion within seven days. Measure error versus final dating, time to scan, prenatal-care linkage by 30 days, repeat visits, patient cost, subgroup disparities, and missed urgent findings.

All prenatal and delivery decisions remain based on definitive evaluation and clinician judgment.

05Defensible data · Hypothesis

Link why the referral completed—or broke—not only the image.

Connect raw sweeps and quality failures, model uncertainty, operator and device, human review, slot availability, contact and transport, definitive imaging, final dating, and 30-day linkage in one event stream. The moat is learning site-specific bottlenecks and equity—not merely another accuracy model.

06Venture thesis · Assumption

Start with safety-net systems and regional care networks—the product is completed imaging.

Bundle device leasing, operator training, quality audit, human review, and referral navigation into a per-clinic monthly fee plus an operations fee per completed definitive scan. The initial value is fewer lost referrals, repeat visits, long-distance transfers, and administrative hours—not replacing sonographers.

Outcomes contracts with Medicaid plans are a later hypothesis contingent on better seven-day completion, 30-day prenatal linkage, and reduced disparities.

Key numbers

The delivery gap left by accuracy evidence

4,695 people

Initial AI development and validation cohort in the US and Zambia

0.2 days

Difference in mean absolute error between AI and standard assessment in a 400-person 2024 study

4.2 days

AI mean absolute error in 2026 Chicago–Nairobi validation; 4.5 days for standard care

16–37 weeks

FDA-cleared range—not for prenatal management or delivery planning

Evidence and interpretation

Accuracy is supported; improved access is not yet proven

Established: a 4,695-person 2022 study and a 400-person prospective study registered as NCT05433519 and published in 2024 found that AI dating from novice-acquired blind sweeps had error similar to standard fetal biometry. In 385 participants across new Chicago and Nairobi settings in 2026, the adapted model achieved a 4.

2-day mean absolute error and was noninferior to the 4. 5-day clinical standard.

FDA cleared K252148 on March 27, 2026, but only for a presumed 16–37-week singleton intrauterine pregnancy, as adjunctive information not intended for prenatal management or delivery planning. The 2026 Zambia PIKABU paper is a protocol evaluating acceptability, feasibility, and fidelity of care integration; it does not yet report patient outcomes.

Whether the proposed pathway improves definitive-scan completion, prenatal linkage, or cost is therefore an untested research and venture hypothesis.

2026 prospective cross-setting validation 2024 prospective integrated-device study Clinical trial registration NCT05433519 FDA 510(k) K252148 FDA indications and performance summary 2026 Zambia implementation protocol

Failure conditions and safety, ethics, and regulatory risks

Clinical and business failure

Do not scale if seven-day definitive-scan completion fails to improve by at least 20 percentage points, 30-day prenatal linkage does not improve, or avoided costs do not cover the service. AI error alone is not success.

Clinical decisions outside the cleared scope

Never finalize a due date or automate medication, testing, transfer, or delivery timing from the AI output. Label it as an adjunctive estimate and block downstream orders until definitive imaging and clinician confirmation.

Wrong population and urgent findings

This is not a tool to rule out or diagnose multiples, ectopic pregnancy, fetal demise, placenta previa, or anomalies. Screen bleeding, pain, hypertension, reduced fetal movement, and other red flags before AI and route them directly to standard urgent evaluation.

Uncertainty and quality failure

Do not present low-confidence sweeps, out-of-distribution inputs, or connectivity failures as valid estimates. Suppress the output, repeat acquisition or escalate to definitive imaging, and monitor operator-level failure rates.

Equity and automation bias

AI-mediated prioritization can still push patients with transport, language, or coverage barriers to the back. Audit completion and error by race, BMI, geography, language, and operator, and include interpretation and transport support in the intervention.

Privacy, liability, and regulatory change

Minimize and encrypt fetal images, location, and contact data with role-based access. Current FDA clearance is tied to specified Butterfly systems and trained health professionals.

A new probe, consumer self-use, model update, or automated care recommendation may require separate validation and 510(k), De Novo, or investigational review.

Primary sources

References

Established facts are separated from research and venture hypotheses that still require validation.

01
Willis A, Lee C, Krogue J, et al. Generalization of AI-Based Gestational Age Assessment Using Blind Sweep Ultrasonography. JAMA Netw Open. 2026;9(7):e2622484.
Across 385 participants in new Chicago and Nairobi settings, the adapted model’s 4.2-day mean absolute error was noninferior to 4.5 days for standard biometry. This tested diagnostic performance, not clinical outcomes.DOI 10.1001/jamanetworkopen.2026.22484
02
Stringer JSA, Pokaprakarn T, Prieto JC, et al. Diagnostic Accuracy of an Integrated AI Tool to Estimate Gestational Age From Blind Ultrasound Sweeps. JAMA. 2024;332(8):649-657.
In a 400-person prospective study, AI and standard assessment had mean absolute errors of 3.2 and 3.0 days in the 14–27-week window, meeting the prespecified equivalence criterion.NCT05433519 · DOI 10.1001/jama.2024.10770
03
Pokaprakarn T, Prieto JC, Price JT, et al. AI Estimation of Gestational Age from Blind Ultrasound Sweeps in Low-Resource Settings. NEJM Evid. 2022;1(5).
Developed and tested the model in 4,695 participants in the US and Zambia; a separate novice, low-cost-device test set showed error similar to standard biometry.PMID 36875289 · DOI 10.1056/EVIDoa2100058
04
U.S. Food and Drug Administration. 510(k) Premarket Notification K252148: Butterfly Gestational Age Tool.
Class II 510(k) clearance for adjunctive GA output in a presumed 16–37-week singleton intrauterine pregnancy—not for prenatal management or delivery planning. The independent performance evaluation included 111 participants at four US sites.K252148 · Decision Mar 27, 2026
05
Gazzetta E, Matenga TFL, Martin S, et al. A pilot study to implement artificial intelligence-enabled, point-of-care obstetric ultrasound for gestational age estimation in Zambia: an evaluation protocol. PLoS One. 2026;21(9):e0357962.
The PIKABU protocol evaluates acceptability, feasibility, and fidelity when nurses, midwives, and other nonexperts integrate AI POCUS into antenatal care. It is not yet an effectiveness result.PMID 42726706 · DOI 10.1371/journal.pone.0357962
06
ClinicalTrials.gov. Diagnostic Accuracy of a Novel Machine Learning Algorithm to Estimate Gestational Age.
Registry record and statistical plan for the prospective diagnostic-accuracy study published in JAMA in 2024.NCT05433519

All issues