External validation of a data-driven algorithm to identify breast cancer recurrences from administrative healthcare data

INTRODUCTION: Population-based data on the incidence of breast cancer (BC) recurrence are scarce, because tracking the relevant information is resource-consuming. We report the external validation of a data-driven case-finding algorithm to identify selected treatment and procedure codes relating to BC recurrences from administrative healthcare data. METHODS: We linked a cancer registry-based cohort of BC patients from the Emilia-Romagna (northern Italy) regional hospital discharge and outpatient care databases, and we ran the algorithm. We compared the algorithm-determined incidence and temporal pattern of recurrences during 10 years of follow-up with those obtained with a manual clinical chart review. RESULTS: The algorithm had a 92.3% (95% confidence interval: 87.2%; 95.8%) sensitivity, a 98.0% (97.0%; 98.8%) specificity, a 88.1% (82.4%; 92.5%) positive predictive value and a 98.8% (97.9%; 99.3%) negative predictive value. The two curves of subhazard rates of total recurrences showed a good degree of overlapping and both exhibited the expected bimodal temporal pattern, with a peak between the 2nd and the 3rd year of follow-up and another between the 8th and the 9th year. The algorithm-based 10-year cumulative incidence function of recurrence was 15.8% (13.7%; 18.1%) versus 15.0% (13.0%; 17.3%). The algorithm captured with good precision the temporal pattern of recurrence by molecular profile. The design of the algorithm excluded the recurrences (11.1% of the total) detected during the first 12 months (HER2-negative BC) and 24 months (HER2-positive BC) after the diagnosis. CONCLUSIONS: This validation study demonstrates that identifying breast cancer recurrences from administrative healthcare data is feasible and sufficiently reliable.