(891) Beyond Biometry: AI-Driven Delivery-Date Prediction Outperforms Hadlock-derived Estimates in Singletons

Ultrasound-derived EDD estimation early in the 1st trimester attempts to define an optimal duration of pregnancy; however, additional data to fetal biometry can be utilized to predict pregnancy duration. Our aim was to determine if a Delivery Date AI model can predict actual days-until-delivery with non-inferiority (≤7 days margin) versus the gold-standard Hadlock method after the 1st trimester and assess superiority.

Study Design:

Retrospective multi-site analysis of singleton pregnancies (≥14 weeks) treated at University of Kentucky and 12 satellite clinics from 2017-2021. Individual patient ultrasound studies with ≥3 de-identified 2-D grayscale ultrasound images were eligible. AI generated delivery dates; whereas Hadlock GA was obtained from US machine text. Actual date of delivery came from delivery records. Absolute error (AE) was calculated as the difference between actual delivery date and the estimate. Mean AE was the primary endpoint. Median AE and R² were calculated. McNemar and two-sided paired t-tests were used.

Results:

2350 studies met inclusion criteria, 1061(45.1%) 2nd and 1289(54.9%) 3rd trimester. Delivery Date AI mean AE was 11.1 days (95 % CI 10.7–11.5) vs Hadlock 15.7 days (14.9–16.5); Δ = –4.6 days (95 % CI –5.3 to –3.9); non-inferiority met; superiority was identified (p < 0.0001). R² improved with AI-incorporation into the estimate: 0.900 vs 0.728. Median AE 8.3 vs 9.0 days was similar; however, data was skewed with AI demonstrating superiority at the upper bounds. Both predictions within 14 days (73.0% vs 66.3%) and within 21 days (88.3% vs 79%) were improved (P< 0.0001); but no difference was identified within 7 days (Figure 1). Performance advantage was observed across trimesters; most BMI subgroups; differing ages and ethnicities; and across ultrasound platforms.

Conclusion:

Relative to the Hadlock standard, Delivery Date AI produced a 4–5 day improvement in predicting date of delivery. Consistent gains across patient subgroups in differing practice locations and across ultrasound platforms supports generalizability of this observation.