Human-AI Collaborative Intelligence for Improving Fetal Flow Management

Human-AI Collaborative Intelligence for Improving Fetal Flow Management: A Randomized Trial

This randomized controlled study evaluates the effectiveness of explainable AI (XAI) in improving clinicians' interpretation of Doppler ultrasound images (UA and MCA) in obstetrics. It involves 92 clinicians, randomized into intervention and control groups. The intervention group receives XAI feedback, aiming to enhance accuracy in ultrasound interpretation and medical decision-making.

Objectives:

1. To develop an interpretable model for commonly used Doppler flows, specifically the Pulsatility Index (PI) of the umbilical artery (UA) and middle cerebral artery (MCA), with the aim to provide quality feedback on Doppler spectrum images and suggest potential gate placements.
2. To test the effects of providing Explainable AI (XAI)-feedback for clinicians compared with no feedback on their accuracy in ultrasound interpretation and management.

Currently, Doppler ultrasound velocimetry serves as a crucial tool in obstetric practice, particularly for assessing the umbilical artery (UA) and middle cerebral artery (MCA) in uteroplacental-fetal circulation. While Doppler ultrasound is valuable for detecting conditions like fetal anemia and placental insufficiency, its accuracy relies on operator expertise. Artificial intelligence (AI) offers potential enhancements, especially in high-risk pregnancies. However, existing AI applications in fetal ultrasound often lack transparency, leading to user distrust. This study aims to address these limitations by developing an explainable AI model to assist clinicians in interpreting Doppler ultrasound images of UA and MCA for improved management of high-risk pregnancies.

The study's objectives are:

1. To develop an interpretable model for commonly used Doppler flows, specifically the Pulsatility Index (PI) of the umbilical artery (UA) and middle cerebral artery (MCA), with the aim to provide quality feedback on Doppler spectrum images and suggest potential gate placements.
2. To test the effects of providing Explainable AI (XAI)-feedback for clinicians compared with no feedback on their accuracy in ultrasound interpretation and management.

All participants will be instructed to provide gate placement for flow images of the umbilical artery and the MCA, and to evaluate the quality of the resulting flow curves. Each participant will be required to evaluate a total of 40 unique images (10 flow images for UA and MCA, 10 spectral doppler images for UA and MCA, respectively). From the four groups (UA-flow, UA-spectrum, MCA-flow & MCA-spectrum) the investigators will provide matched sets of 40 images that are provided to participants, who are matched for their level of experience within each hospital (PGY 1-2; PGY 3-5; board certified Obstetricians). For flow images, the participants will be instructed to identify the most appropriate gate placement. For the spectral flow curves, participants will be asked to evaluate whether the flow curves were of sufficient quality to inform medical management decisions.

The inclusions criteria for MCA and UA images will be images taken from the third trimester (>= week 28).

Study Design: Randomized controlled trial

Data Source: 1840 unique ultrasound scans including umbilical artery (UA) and middle cerebral artery (MCA) measurements. The 1840 unique images includes: 460 images of UA-flow images, 460 UA-spectrum images, 460 MCA-flow images and 460 MCA-spectrum images.

Participants: 92 clinicians with varying competence levels across four different university hospitals.

Intervention: XAI feedback on MCA/UA Doppler spectral curves and gate placement suggestions.

Control Group: No XAI feedback.

Procedure: Clinicians will be divided into two groups of 46 each, matched for experience levels across hospitals. The control group will place a gate on MCA/UA images and evaluate the Doppler spectrum without AI feedback, while the intervention group will perform the same tasks with access to AI feedback.

Outcome Measures: The participants' responses in the two groups are reviewed by two fetal medicine sonographers who evaluate the participants' answers independently of each other. In a disagreement, the two sonographers reach a consensus after discussion.

Inclusion Criteria:

• The inclusion criterion is the use of ultrasound at least once per week

Exclusion Criteria:

• The exclusion criterion is the absence of experience in ultrasound scanning.