Background: Five-year breast cancer survival is driven largely by stage at diagnosis, varying from >99% for localized to <35% for distant/metastatic disease. In the U.S.-Affiliated Pacific Islands (USAPI), screening mammography is largely absent outside Guam, radiologists are scarce, cost of travel is prohibitive, and on-island pathology services are minimal or nonexistent, resulting in advanced-stage diagnosis rates >50%. The World Health Organization’s Global Breast Cancer Initiative (GBCI) presents three goals for breast cancer early detection: >60% of cancer diagnosed at non-advanced stage; imaging, pathology, and diagnosis complete within 60 days; and >80% treatment completion. Achieving these targets in the USAPI requires strategies that do not rely on mammography or specialty interpretation. Artificial intelligence (AI) may function as a clinical companion to healthcare workers, enabling them to perform specialty tasks.

Objectives: To increase breast cancer survival using an AI-enabled framework, accessible imaging, and biopsy technology in the USAPI. We call this paradigm Makawalu. We present the overall paradigm, pilot data, and AI solutions.

Methods: The Makawalu paradigm uses AI to augment local resources for education, risk assessment, detection, and diagnosis. An AI chatbot will help educate local women on resources, risk factors, diagnostic procedures, and treatments. Following clinical breast examination and questionnaire-based risk assessment, symptomatic and high-risk women will proceed to AI-informed breast ultrasound imaging. One AI solution will focus on detection of suspicious lesions from imaging. A second will estimate probability of cancer. In women without suspicious findings, AI-derived breast density will be added to their risk assessment. Women with suspicious findings will be recommended for fine-needle aspiration biopsy under AI-informed imaging guidance. The collected breast cells will undergo rapid onsite evaluation and staining, bedside digitization, and AI estimation of both specimen adequacy and likelihood of true-negative status. Specimen will then be preserved and shipped to a regional facility. This paradigm makes use of AI for five separate tasks,

Results: We are evaluating AI chatbots, such as askellen.ai, for breast cancer education. The AI models for detection and classification have been developed using over 8,854 breast ultrasound images from the Hawai‘i and Pacific Islands Mammography Registry (HIPIMR). On a withheld test dataset, the classification model achieved an AUROC of 0.88. For inclusion in clinical risk models, AI-derived mammographic density was estimated using 405,120 negative breast ultrasound images. An AUROC of 0.85 was achieved across all density categories. Using a public dataset of 482 breast cytology images and transfer learning from a histology model (Virchow2) malignancy status was predicted with an AUROC of 0.96. Current studies include adaptation of these models to point of care ultrasound systems and to bedside biopsy procedure. Application of AI is expected to contribute to achieving the GBCI’s goals.

Conclusions/Implications: In settings without adequate mammography, radiologists, or pathologists, AI may provide a practical pathway to all aspects of breast cancer education, risk assessment, detection, and diagnosis. The Makawalu paradigm integrates multiple AI solutions to operate together with local healthcare workers. This framework may be generalizable to other remote and resource-constrained populations globally.