Abstract
Implementing AI prioritization can accelerate care timelines for patients requiring additional workup, while maintaining the efficiency of delayed interpretation for most participants. Reducing diagnostic delays could contribute to improved patient adherence, decreased anxiety and addressing disparities in access to timely care.
Introduction
In the United States, breast cancer screening typically follows a "batch" reading model. Patients undergo mammography and return home, while radiologists interpret the images later. If an abnormality is detected, the patient is recalled for diagnostic imaging. This delay often spans days or weeks, causing significant patient anxiety and increasing the risk of "loss to follow-up," particularly among underserved populations. While immediate "same-day" results are ideal, they are logistically challenging for most clinics. This study, a collaboration between Northwestern University and Google Health, investigates whether artificial intelligence (AI) can bridge this gap by triaging high-risk cases for immediate review.
Methods
The researchers conducted a prospective, randomized, unblinded implementation study at the Lynn Sage Comprehensive Breast Center in Chicago, involving 1,000 participants aged 40–89.
AI System and Workflow
A cloud-based AI model developed by Google Health generated malignancy risk scores (0 to 1). Participants were randomized into two groups:
- Control Group: Followed the standard batch-reading workflow.
- Experimental Group: Mammograms were processed by AI in real-time. Cases exceeding a risk threshold were "Prioritized" for radiologist review within 30 minutes. If the radiologist confirmed a finding, the patient was offered same-visit diagnostic imaging.
Primary Metrics
The study evaluated two primary endpoints:
- $T_A$ (Time to Additional Imaging): The duration from the initial screening to the follow-up diagnostic imaging.
- $T_B$ (Time to Biopsy Diagnosis): The duration from the initial screening to the final biopsy result.
Results
The AI-modified workflow significantly accelerated the diagnostic process without increasing the overall radiologist recall rate.
Operational Efficiency
- Reduced Delays: The experimental group saw a 25% reduction in mean $T_A$ (19.1 days vs. 25.6 days) and a 30% reduction in mean $T_B$ (39.2 days vs. 55.9 days).
- Prioritized Subgroup: For patients specifically flagged by AI, the mean $T_A$ dropped to just 3.5 days.
- Patient Acceptance: 86% of patients offered same-visit follow-up accepted it, indicating high motivation for immediate resolution.
Clinical Safety
The system achieved 100% sensitivity, correctly flagging all six cancers detected during the study for prioritization. Importantly, the radiologist recall rate remained stable (17.0% experimental vs. 14.3% control), suggesting the AI did not lead to over-diagnosis.
Discussion
This study highlights AI's potential as a workflow optimizer rather than just a diagnostic aid. A notable finding was the "real-world gap": the AI prioritized 29% of cases in practice, compared to 14% in retrospective testing, emphasizing the importance of prospective validation.
Limitations and Future Directions
The study's primary limitation is the small number of detected cancers (n=6), which limits definitive conclusions regarding diagnostic accuracy. Additionally, the study was conducted at a single high-resource academic center. Future research should explore implementation in community settings and address radiologist "discomfort" through AI explainability tools to foster better human-AI collaboration.
Summary updated