Comparison of Data Mining Methods for Signal Detection of Targeted Therapy Related Adverse Events in Breast Cancer Patients

• Activity Number: 653 - Machine Learning and Other Statistical Methods in Clinical Trials
• Type: Contributed
• Date/Time: Thursday, August 1, 2019 : 10:30 AM to 12:20 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Efstathia Polychronopoulou* and Sharon Giordano and Lin-Na Chou and Xiaoying Yu and Yong-Fang Kuo
• Companies: UTMB and MD Anderson Cancer Center and The University of Texas Medical Branch and UTMB and The University of Texas Medical Branch
• Keywords: drug safety surveillance; pharmacovigilance; adverse events data mining; tree-based scan statistic; gamma poisson shrinkage; proportional reporting ratio

Abstract

Application of signal detection methods using claims data can improve post-marketing drug surveillance. The aim of the study is to compare two routinely used approaches, the proportional reporting ratio (PRR) and Gamma Poisson Shrinker (GPS) with a tree-based scan statistic (TBSS). Using data from Texas Cancer Registry and SEER linked to Medicare from 2010-2014 we identified 9076 patients with breast cancer treated with chemotherapy and 2559 patients treated with Herceptin in addition to chemotherapy. Inpatient and outpatient visits up to 2 years from start of therapy were used to identify adverse events (AE). Diagnosis codes were classified into 525 clinical categories. Separate analysis was conducted at patient and visit level. Across all methods we found a total of 38 signals associated with use of Herceptin. Clinical review determined that most signals identified by GPS and TBSS represented known AE or confounding, while ~25% of PRR signals were false positives. GPS on the highest signaling threshold failed to detect a well-established AE when time of follow-up was less than 6 months. Overall there was considerable agreement between methods with TBSS being more sensitive.