Image-only ultrasound AI for dating-independent delivery-date prediction externally validated in four Chilean hospitals

Technical Summary: Image-Only Ultrasound AI for Dating-Independent Delivery-Date Prediction

Problem Statement
Accurate estimation of the delivery date (EDD) is fundamental to obstetric care, governing gestational age assignment, fetal growth interpretation, and critical clinical decisions regarding induction and neonatal planning. However, conventional dating methods—relying on the last menstrual period (LMP), first-trimester crown–rump length (CRL), or later biometry—perform poorly in populations lacking reliable early ultrasound or menstrual history. This includes public-sector and migrant populations who often present late in pregnancy. Furthermore, the standard 280-day convention for gestational length is known to overestimate duration in many non-European populations, introducing systematic bias. Consequently, there is a persistent equity gap where the patients most in need of accurate timing (those without reliable early dating) receive the least accurate estimates.

Methodology
This study presents an independent, multicentre external validation of a prospectively deployed, image-only artificial intelligence (AI) system.

• Study Design: A retrospective cohort analysis of a prospectively deployed system across four public hospitals in Chile (Barros Luco Trudeau, Padre Hurtado, El Pino, and Dr. Luis Tisné Brousse). The analysis was conducted independently of the model developer.
• Data: The cohort comprised 3,788 singleton ultrasound studies from 2,550 pregnancies performed between March 2021 and November 2025. The population included a high-risk mix with significant representation of migrants (34.4%, notably Venezuelan and Haitian) and comorbidities such as fetal growth restriction and gestational diabetes.
• AI System: The model (Ultrasound AI, Inc.) accepts standard-of-care ultrasound image frames as input. Crucially, it operates without gestational age, biometric measurements, demographics, or operator annotations. It learns directly from pixel-level patterns across the entire frame, trained against the calendar date of birth.
• Comparator: The AI was compared head-to-head against the conventional clinical estimate (the established due date projected to the scan date, calculated as 280 days minus gestational age at scan).
• Statistical Analysis: Performance was quantified using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and bias. Comparisons utilized a non-parametric bootstrap resampling of pregnancies (3,000 resamples) to account for clustering and the Wilcoxon signed-rank test for paired differences. Calibration of prediction intervals was assessed by comparing empirical coverage against nominal coverage.

Key Contributions

1. Dating Independence: The study demonstrates that an image-only model can predict delivery timing without any input regarding gestational age or dating history, effectively bypassing the errors associated with uncertain menstrual history or late-entry scans.
2. External Validation: It validates the "PAIR" study findings (originally from a single US center) in a distinct, independent health system (Chile) with heterogeneous equipment and a high-risk, diverse population.
3. Calibration: Unlike previous iterations, this work explicitly validates the calibration of the model's predicted distribution, showing that the output can be reliably reported as a delivery-date range rather than a single point estimate.
4. Equity Focus: The study specifically evaluates performance in the subgroup of pregnancies lacking reliable early dating, where conventional methods are structurally weakest.

Results

• Overall Accuracy: Across all 3,788 studies, the AI predicted the delivery date with a Mean Absolute Error (MAE) of 6.9 days (95% CI 6.7–7.0).
• Performance in Uncertain Dating: In the subgroup of 456 studies with late/uncertain dating (no first-trimester CRL), the AI maintained an MAE of 6.4 days. In contrast, the conventional due date deteriorated significantly to an MAE of 13.2 days.
• Head-to-Head Comparison: In 69.2% of all pregnancies, the AI provided a more accurate estimate than the conventional due date. This advantage widened to 73.2% in the uncertain-dating subgroup.
• Bias Correction: The conventional estimate exhibited a systematic bias of −10.8 days (birth occurred ~11 days earlier than predicted), reflecting the population's shorter gestational length compared to the 280-day convention. The AI showed a negligible bias of −1.5 days. Sensitivity analyses revealed that a large portion of the AI's superiority over the conventional method stems from correcting this population-specific bias, with a smaller residual advantage in dispersion.
• Preterm Birth: For preterm births (<37 weeks), the conventional method failed completely (MAE 37.5 days, as it assumes 40 weeks), whereas the AI maintained an MAE of 10.8 days and was more accurate in 99.6% of cases.
• Calibration: The model's high-coverage prediction intervals were well-calibrated. The approximately 94% interval (quantiles 0.03–0.97) covered 96.5% of actual delivery dates.
• Robustness: Performance remained consistent across all four hospitals, various gestational ages at scan, and major comorbidity subgroups, including obesity.

Significance and Claims
The paper claims that this dating-independent, calibrated estimate addresses a critical equity gap in obstetric care. By relying solely on image patterns, the AI provides accurate predictions for patients who lack reliable early dating—precisely the group for whom conventional methods fail. The authors argue that this capability is not merely a technical improvement but a clinical and equity-relevant advancement, as it offers a reliable alternative when the standard "due date" is unavailable or untrustworthy.

The study positions the AI as a potential "second read" adjunct to established dating, particularly useful for late-presenting or migrant patients. The authors emphasize that the AI's value lies in its ability to deliver accurate, calibrated predictions for patients without a trustworthy due date, rather than simply outperforming an optimally recalibrated due date (which requires retrospective population data unavailable in real-time). The work supports the continued development of ultrasound AI as a practical tool for delivery-date estimation, moving the technology closer to clinical adoption through rigorous, independent validation.