OBJECTIVES: This study aimed to develop and validate algorithms for identifying metastasis/recurrence of high-risk HR+/HER2 EBC in hospital-based databases.

METHODS: An electronic medical record database (NHO Clinical Data Archive [NCDA]) and a claims database (Medical Information Analysis [MIA] databank) were used (data available from Apr 2015 to Mar 2021). Female breast cancer patients, who had the first breast surgery between Apr 2015 and Mar 2019, received any hormone therapy and no anti-HER2 therapy, and had ≥1-year follow-up period or died within 1 year after the breast surgery date (Index Date), were eligible. Patients classified as ≥N1 or ≥T3 of TNM stage during the Index Hospitalization (hospitalization including Index Date) were assumed as having high risk of metastasis/recurrence; among them, patients from two hospitals were selected for this validation study. At each hospital, two independent specialists confirmed true outcomes and stage classification in the institutional medical records. Two models, 1) combination of diagnosis, medication, procedure (Rule-based model), and 2) the Lasso regression model (Lasso-model) with diagnosis, medication, procedure as independent variables, were used to develop algorithms to define metastasis/recurrence events. Sensitivity and positive predictive value (PPV) of the algorithms were calculated by comparison with true outcomes.

RESULTS: A total of 166 patients were evaluated (55±13 years old [mean±SD]; number of patients by stage: T1=32, T2=93, T3=26, T4=15 and N0=9, N1=129, N2=10, N3=18). Sixteen metastasis/recurrence events were identified as true outcomes from medical records. Sensitivity and PPV of the algorithms (defined by diagnosis of metastasis/recurrence and use of drugs for metastatic breast cancer) were 68.8% and 68.8% (Rule-based model), and 87.5% and 73.7% (Lasso-model). Either the clinical or pathological stage retrieved from each hospital matched 83.7% of overall stage, 95.8% of T-classification, 89.8% of N-classification in the database.

CONCLUSIONS: Algorithm developed with Lasso-model resulted in relatively high sensitivity and PPV.