Diagnostic Accuracy of an Integrated AI Tool to Estimate Gestational Age from Blind Ultrasound Sweeps

Information Science Artificial Intelligence Obstetrics and Gynecology Medicine Medical Imaging
integrated AI tool gestational age estimation diagnostic accuracy blind ultrasound sweeps low-resource settings pregnancy care obstetrical sonography

Diagnostic Accuracy of AI Tools for Estimating Gestational Age Based on Blind Ultrasound Scanning

Background

Accurate estimation of gestational age (GA) is the foundation of good prenatal care, typically achieved through ultrasound examinations. However, many low-resource areas lack sufficient ultrasound equipment, making accurate GA estimation challenging. In recent years, advancements in hardware and artificial intelligence (AI) for medical image analysis have provided opportunities to use this diagnostic tool more widely. This study developed a low-cost, battery-powered device that does not require high-end configurations, based on a deep learning AI model, to assess its accuracy in estimating gestational age in the hands of non-specialist users.

Research Origin

This study was authored by Jeffrey S. A. Stringer, MD, and his team, who are affiliated with the University of North Carolina, the University of Zambia, and other institutions. The study was published online in JAMA on August 1, 2024.

Research Process

  1. Study Design:

    • The study was designed as a prospective diagnostic accuracy study.
    • Conducted in two locations: Lusaka, Zambia, and Chapel Hill, North Carolina.
    • A total of 400 individuals with viable, singleton, uncomplicated early pregnancies were recruited.

  2. Methods:

    • Participants first had their gestational age determined by crown-rump length (CRL) using traditional methods, serving as the “reference standard.”
    • In follow-up, untrained general users utilized AI-assisted devices (blind scanning), and certified sonographers used high-end machines for fetal measurements.
    • The primary assessment window was set between 14 0/7 weeks and 27 6⁄7 weeks.
    • The device used was a Butterfly IQ+ handheld ultrasound with modified software, integrating a deep learning model.

  3. Experimental Steps:

    • Initial verification of patient location and medical records was followed by the blind scanning procedure.
    • Each user underwent a brief one-day training to learn basic operations, such as software navigation, patient positioning, probe movement, etc.
    • A 10-second blind scan was performed using the battery-powered device, with real-time processing and image analysis.

  4. Data Analysis:

    • Monte Carlo simulation was used to determine sufficient sample size to ensure a 95% confidence level within the preset margin of error.
    • The primary tolerance range was set at ±2 days, with secondary outcomes including root mean square error, accuracy within 7 days, and 14 days.

Main Research Results

  1. Primary Window Results (14-27 Weeks):

    • The mean absolute error (MAE) for the two methods within the primary assessment window was 3.2 days (AI tool) and 3.0 days (traditional method).
    • Approximately 90.7% of assessments were within ±7 days, with both methods performing comparably.
    • Accuracy was consistent across high BMI subgroups and different geographical locations.

  2. Secondary Window Results (28-36 Weeks and 37-40 Weeks):

    • During the 28-36 week period, the AI tool had an MAE of 6.07 days, compared to 7.12 days for the traditional method, with the AI tool performing better.
    • During the 37-40 week period, the AI tool performed worse than the traditional method, with a significant drop in accuracy.

Conclusion

Between 14 and 27 weeks, non-specialist users with no ultrasound training using a low-cost, AI-assisted portable ultrasound tool achieved GA estimation accuracy comparable to that of trained sonographers using high-end equipment. This has significant implications for obstetric care in low-resource settings, advancing the World Health Organization (WHO) goal of widespread access to prenatal ultrasound.

Research Highlights

  • Innovation: The integration of AI into portable ultrasound devices significantly reduces equipment costs and eliminates the need for professional operation.
  • Applicability: In low-resource settings, widespread adoption of such AI-assisted devices could improve the coverage and accuracy of prenatal care.
  • Effectiveness: The study results demonstrate the feasibility and immediate potential of this technology.

Importance

This study showcases the effective application of AI in medical devices, not only improving the accuracy of GA estimation but also providing a viable prenatal care solution for low-resource regions. In the future, this technology could be further validated and promoted among more high-risk populations, bringing broader health benefits.

The study was funded by the Bill and Melinda Gates Foundation, with equipment provided by Butterfly Systems. This research not only expands the application of GA estimation but also provides empirical support for future medical technology innovations.

Closing Remarks

The study highlights the application of AI technology in estimating GA during pregnancy, demonstrating its accuracy and practical significance in resource-poor areas. This is part of the digital health transformation, contributing to improving the quality of prenatal care globally and advancing the goals of universal and equitable healthcare.