Abstract
BACKGROUND: Human epidermal growth factor receptor 2 (HER2) is a treatment target in breast cancer (BC), driving therapeutic strategies. Changes over time in HER2 expression have been described and understanding of these fluctuations is crucial for personalized medicine. We aimed to assess HER2 expression dynamics using real-world data and natural language processing (NLP) from electronic health records (EHRs). MATERIAL AND METHODS: RosHER is a retrospective, observational, longitudinal, population-based, multicenter study (NCT05217381). An NLP tool extracted HER2 information from EHRs of adult patients with early, locally advanced, or de novo metastatic BC, who were initially diagnosed between 2005 and 2021. The primary endpoint was to evaluate HER2 dynamics in HER2 status and expression between initial diagnosis and recurrence or progression using NLP. The secondary endpoints were description of baseline clinicopathological characteristics and treatment patterns. RESULTS: Between January 2022 and November 2023, 18 533 patients were selected from seven Spanish sites. A cut-off of ≥6 months was established between initial determination and relapse or progression. The artificial intelligence (AI)-based tool identified 510 patients with two documented HER2 determinations and 209 with HER2 expression by immunohistochemistry/in situ hybridization. Overall discordances were 10.6% in HER2 status and 34.0% in HER2 expression. HER2-zero expression switched to HER2-low (23.2%), but not HER2-positive (0%); HER2-low expression converted to HER2-zero (32.0%) and HER2-positive (7.0%); finally HER2-positive expression switched to HER2-low (20.8%) and HER2-zero (15.1%). CONCLUSIONS: This is the first study using NLP to evaluate HER2 discordances, which need to be further investigated. Improving AI methods and implementing similar EHR structures among hospitals would increase the success in clinical data extraction.