Abstract P4-01-05: Machine learning to predict treatment response and tolerability in HR+, HER2– advanced breast cancer: German study AI4ANNA

Abstract Background: Predicting the probability of tumor progression and tolerability with sufficient accuracy remains a significant challenge in advanced breast cancer. The objective of AI4ANNA study was to assess the predictive potential of machine learning (ML) methods with respect to tumor control and safety outcomes using German study data (RIBECCA, RIBANNA) and to identify the most relevant baseline factors for prediction. Methods: Anonymized study data from two studies of ribociclib and endocrine therapy in patients with advanced HR+, HER2– breast cancer were used for predictive analysis. RIBECCA (N=487) was a multicenter, open-label, single-arm, phase 3b trial, and RIBANNA (N=1904) is an ongoing non-interventional study evaluating the real-world efficacy and safety of first-line ribociclib in combination with aromatase inhibitor/fulvestrant, endocrine monotherapy or chemotherapy. Study baseline features were used to develop prediction models for a variety of tumor control (including progression-free survival (PFS), overall response rate (ORR) at week 24, death) and safety outcomes (including general number of adverse events (AEs) as well as selected AEs belonging to blood system, cardiac, hepatobiliary, and gastrointestinal disorders). LASSO (Least Absolute Shrinkage and Selection Operator) and XGBoost (eXtreme Gradient Boosting) ML algorithms were employed to train prediction models. LASSO was selected as a representative of fully interpretable, linear models and XGBoost as a representative of highly flexible, nonlinear models. Predictive performance of these two algorithms was compared and predictive value of single baseline features was assessed using feature permutation importance method. Results were validated internally within the training study (10 times repeated 5-fold stratified cross-validation) as well as externally, ie, implementation and training of the prediction models on one study and validation on the other and vice versa. Results: Moderate predictive signal (at baseline) could be identified for the following two outcomes: ORR (area under the curve [AUC] mean 0.628 [RIBANNA] and 0.626 [RIBECCA]) and PFS (AUC mean 0.626 [RIBANNA], 0.604 [RIBECCA]). Model performance could be validated with very similar AUCs by cross-study evaluation. Patients could be assigned to one of three risk groups. The most important features for ORR prediction included the presence of locally advanced cancer and metastases presenting as bone only disease and for PFS the presence of liver metastases, histological grade and prior (neo)adjuvant treatment. For three safety endpoints, a predictive signal (AUC >0.6) was identified only in one study but not in the other (AEs “QT prolongation”, “leukopenia grade 3/4” and serious AE “vomiting”). However, insufficient predictive signals were found for all other outcomes. Conclusion: Prediction models for tumor control and safety outcomes were trained on a broad number of clinical baseline features. Moderate predictive signals could be identified for ORR and PFS. Even though the predictive performance (AUC) seems to be limited, patients could be assigned to one of three differently behaving risk groups at the baseline. The key predictive features for PFS included clinically known prognostic factors like liver metastasis, histological grade and prior (neo)adjuvant treatment. Citation Format: Peter A. Fasching, Achim Wöckel, Hans Tesch, Bernhard Volz, Uwe Pritzsche, Marc Bachmann, Asmir Vodencarevic, Julia Kreuzeder, Diana Lüftner. Machine learning to predict treatment response and tolerability in HR+, HER2– advanced breast cancer: German study AI4ANNA [abstract]. In: Proceedings of the 2022 San Antonio Breast Cancer Symposium; 2022 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2023;83(5 Suppl):Abstract nr P4-01-05.