Abstract Keywords Metastatic breast cancer, Endocrine therapy, Targeted therapy, Oral chemotherapy, French population Context Oral anti-cancer treatments have been shown to be effective when followed carefully. Tamoxifen, for example, reduces the risk of relapse by half within 10 years of the diagnosis [1]. However, these treatments are frequently poorly adhered to. To determine the categories of patients at risk and the appropriate moment to contact them, we developed predictive models trained on anonymised reimbursement data extracted from the French Health Insurance database. Objective The primary objective is to model a metastatic breast cancer patient’s persistence and compliance to the treatment. We aim at detecting unwanted episodes (non persistence and non compliance) six months before they happen. The oncologist may then follow the patient more closely. Methods Patients data is extracted from the SNDS database, one of the largest structured databases of health data in the world. It contains reimbursement data of the French Health System, covering 98% of the French population (66 million persons). Useful data are, for example, hospitalisations, drug purchases or the patient’s age and city of residence. From this database, patients were selected on the basis of a diagnosis of metastatic breast cancer (if hospital stay) or on the basis of specific treatments for metastatic breast cancer. Men and patients under 18 are excluded from the study. We consider that a patient has a non persistent event if she has no treatment stock for 2 months (during a phase of targeted therapy or oral chemotherapy) or 3 months (during a phase of endocrine therapy) and if no change in treatment, palliative care entry or death is observed. The compliance is labelled through the MPR (Medical Possession Ratio): a patient is considered non-compliant if the MPR of her 3 nexts purchases is below 80%. The proposed models are trained to detect non-persistence and non-compliance events in the next 180 days. We created several groups of features describing the patient and her healthcare pathway. Results 250 000 patients were spotted with a breast cancer in the SNDS database. Amongst these, around 40 000 were spotted for a metastatic breast cancer between 2013 and 2018. 14% of the patients had at least one non persistence episode and 46% had at least one non compliance episode. For the persistence study, we used a logistic regression with a feature selection. This model has a Gini coefficient of 0.35. For the compliance study, we used a deep learning model based on a GRU model. This model has a Gini coefficient of 0.37. A multivariate analysis shows that the following features had a significative impact on both predicted risks (persistence and compliance) : age, previous compliance, type of oral treatment(s) currently followed (endocrine therapy, targeted therapy, or oral chemoterapy), number of different oral treatments followed in the past year. In both models, if the patient’s age is between 50 and 70 years it does not correlate with an increased risk. On the other hand, the more they deviate from this interval, the more likely they are to be non-compliant. Conclusion Both studies have models with quite the same interpretation. Patients younger than 50 or older than 70 are more likely to be non-persistent and non-prevalent. The past compliance is highly correlated to the future events. The consumption of oral chemotherapy in comparison to oral endocrine and targeted therapy is linked to an increased risk in both studies. Bibliographie [1]: E. Ekinci, S. Nathoo, T. Korattyil et al. (2018) Interventions to improve endocrine therapy adherence in breast cancer survivors: what is the evidence? J Cancer Surviv 12:348-356 Citation Format: Pierre Rinder, Théo Marcille, Paul Sinel–Boucher, Pierre Hornus, Pierre E. Heudel, Chantal Bernard-Marty, Christelle Levy, Luis Teixeira, Dorra Kanoun. Persistence and compliance of the French metastatic breast cancer population [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 P6-03-03.