Optimizing risk-based breast cancer screening policies with reinforcement learning

Area of Science:

• Medical Imaging and Artificial Intelligence
• Oncology and Public Health
• Machine Learning in Healthcare

Background:

• Screening programs aim for early disease detection but face challenges balancing benefits against overscreening costs.
• Personalized screening strategies are needed to optimize resource allocation and patient outcomes.
• Current breast cancer screening protocols may not fully leverage individual risk stratification.

Purpose of the Study:

• To introduce and evaluate Tempo, a novel reinforcement learning-based framework for personalized cancer screening.
• To assess the efficacy of Tempo in breast cancer screening using diverse datasets.
• To demonstrate the adaptability of Tempo to various clinical preferences and risk models.

Main Methods:

• Development of a reinforcement learning framework (Tempo) for personalized screening policies.
• Training and validation of Tempo using large-scale screening mammography datasets from multiple international institutions (MGH, Emory, Karolinska, CGMH).
• Integration of Tempo with image-based artificial intelligence (AI) risk models and comparison with clinical risk models.

Main Results:

• Tempo, combined with an AI risk model, demonstrated significantly higher efficiency in simulated early detection per screen frequency compared to current clinical practices across all test sets.
• The Tempo policy proved adaptable to different screening preferences, allowing clinicians to adjust the trade-off between early detection and screening costs without retraining.
• Tempo policies utilizing AI-based risk models outperformed those based on less accurate clinical risk models.

Conclusions:

• Pairing AI-based risk models with AI-designed screening policies like Tempo offers a promising approach to enhance early cancer detection.
• Tempo has the potential to significantly improve screening program efficiency by reducing overscreening while advancing early detection.
• The adaptability of Tempo supports personalized screening tailored to individual patient needs and clinical settings.