Artificial intelligence to accurately identify breast cancer patients with a pathologic complete response for omission of surgery after neoadjuvant systemic therapy: An international multicenter analysis.

Abstract

565 Background: Neoadjuvant systemic treatment elicits a pathologic complete response (pCR) in up to 80% of women with breast cancer. In such cases, breast surgery, the gold standard for confirming pCR in the breast, may be considered overtreatment. So far, no approach alone – e.g. imaging, vacuum-assisted biopsy (VAB) – has accurately detected and excluded residual disease without surgery in multicenter prospective trials. We evaluated the ability of Artificial Intelligence algorithms to securely identify patients with residual tumor in the breast to safely select patients who might be spared from surgery. Methods: We collected multicenter, international data from 570 women who were included in prospective trials with initial stage I-III breast cancer of all biological subtypes and at least partial response on imaging, undergoing VAB before guideline-adherent surgery. We trained an ensemble of algorithms (including Regularized Regression, Support Vector Machines, and Neural Network) using 27 patient, tumor and VAB variables. Data were randomly partitioned into training and test sample with a 3:1 ratio and developed with 10-fold cross-validation. Primary endpoint was the sensitivity to diagnose residual disease by algorithm compared to surgery. Diagnostic performance of the algorithm was further evaluated on an external, independent dataset. Results: The algorithm was able to reliably identify women with residual disease before surgery (see table): Sensitivity for the internal test set was 96.9% (94 of 97; 95%CI 91.2-99.4%) and for the external, independent dataset 96.2% (26 of 27; 95%CI 80.4-99.9%). Most informative predictor of residual disease were tumor cells diagnosed in the VAB specimen, DCIS in the initial diagnostic biopsy, grading, and largest diameter on imaging after neoadjuvant treatment. Conclusions: Safely selected patients without residual disease as assessed by our algorithm may be spared by breast surgery in future trials. [Table: see text]

Development of Machine Learning Algorithms for the Prediction of Financial Toxicity in Localized Breast Cancer Following Surgical Treatment

“…The development of our algorithms was informed by guidelines on how to use ML in medicine 13 and diagnostic tests 22 as well as previously published research by our group. 16,23,24 We provide a detailed description of the algorithm development according to these guidelines in the Data Supplement. We used the Prediction Model Study Risk of Bias Assessment Tool (PROBAST) to ensure high accuracy when these models were used elsewhere.…”

“…Four algorithms were selected based on their demonstrated high performance on similar predictive tasks. 23,24 Algorithms are introduced briefly below, and details on the algorithm development are published elsewhere and in the Data Supplement. 16 Regularized regression using the elastic net.…”