Implementation of an AI-Based Risk Stratification Analysis for Breast Cancer in Routine Healthcare

Introduction:Breast cancer (BC) remains a leading cause of death worldwide. Accurate pathological assessment is crucial for estimating individual prognosis and informing personalised treatment decisions. However, a substantial proportion of patients continue to face an uncertain risk of recurrence. Genomic tests such as the Oncotype DX can enhance prognostic accuracy, but they remain prohibitively expensive and widely inaccessible on a global scale.We have developed an artificial intelligence (AI)-based alternative to genomic testing, utilising digitised haematoxylin and eosin (H&E)-stained whole-slide images (WSIs). The analysis has been further improved and subsequently adapted to meet regulatory standards as a commercial product. The solution called Stratipath Breast leverages deep learning-based image analysis to extract prognostic information based on tumour morphology. Methods:The AI-model was trained and externally validated in population-representative datasets comprising 1,567 and 1,262 breast cancer patients, respectively. Prognostic performance has been validated in >2,700 patients and in 2,466 estrogen receptor-positive and HER2-negative (ER+/HER2-) patients without adjuvant chemotherapy in the Danish Breast Cancer Group Cohort. A new version of the model, adapted for use on preoperative core needle biopsies (CNB), enables risk stratification prior to neoadjuvant therapy. Results:Stratipath Breast has been proven to stratify patients with intermediate-grade BC into high-risk and low-risk groups based on H&E-stained WSIs. External validation demonstrated a hazard ratio of 3.88 (95% CI: 1.43-10.52) for progression-free survival between the lowest (reference) and highest risk groups. To date, the solution has been applied to over 1,000 resected BC tumours in routine healthcare, with observed reduction in diagnostic turnaround time by 6-10 days and per-patient costs by 70-80% cost compared to commercial genomic tests. Conclusion:Integrating AI-based risk stratification into routine pathology workflows holds the potential to improve patient outcomes and streamline healthcare systems, expanding access to precision diagnostics. The solution is currently implemented in several European hospitals. Stratipath Breast is to our knowledge the first regulatory approved AI-solution for BC risk stratification integrated into routine clinical practice. It enables reliable risk assessment directly from H&E-stained WSIs, with the potential to reduce both undertreatment and overtreatment of BC patients. The recent adaptation for CNB analysis further expands its clinical utility, supporting earlier and more informed treatment decisions.

Learning Objectives:

1. Understand the need of risk stratification in early-stage breast cancer
2. AI-based analysis as an alternative to genomic tests in breast cancer
3. Digitisation of pathology enables introduction of cost-effective solutions AI-based risk stratification