HER-France is a French national database focused on HER2 status in breast cancer and provided by 125 pathology laboratories (PL) since 2011. PL used to compare their HER2 positivity (HER2+) rate with the calculated national average. To provide a more sensitive monitoring tool evaluating their practices quality through indicators, a new strategy considering a predicted estimate of HER2+ rate by PL instead of national average was investigated. Model to predict probability of HER2+ on core-needle biopsies (CNBs) was developed using penalized logistic regression on tumor characteristics. PL HER2+ rate estimations were obtained by averaging individual HER2+ probabilities. A PL having included more than 100 CNBs between January 2014 (ASCO-CAP recommendations) and April 2016 is considered as outlier when its averaging HER2+ predicted rate is outside a 99% confidence interval (CI) limits of its observed rate. Other indicators are absolute and relative percentage differences between observed and predicted HER2+ rates. Prediction accuracy (AUC) on 30,777 CNBs was higher than 0.77. Among the 56 PL with at least 100 CNBs (i.e. 95% of all the analysed CNBs), 6 (10.7%) were identified as outliers. Illustration with 4 PL: 2 not outliers: PL1 (n=676): observed rate: 13.6%; 99%CI: [10.2; 17.0]; predicted rate: 10.9%; absolute (relative) difference: 2.7 (20%) PL2 (n=2,062): observed rate: 9.1%; 99%CI: [7.4; 10.7]; predicted rate: 10.1%; absolute (relative) difference: 1.0 (11.8%) 2 outliers: PL3 (n=1,468): observed rate: 13.5%; 99%CI: [11.2; 15.8]; predicted rate: 10.9%; absolute (relative) difference: 2.6 (19%) PL4 (n=1,854): observed rate: 9.6%; 99%CI: [7.8; 11.4]; predicted rate: 12.0%; absolute (relative) difference: 2.4 (25%) PL tumor characteristics provide better accuracy in quality assessment practices than comparison to national average. Data mining models implemented in the HER-France monitoring web-based tool will help PL to assess their own rate through consistency indicators.

The aim of the real world data study HERABLE in gastric (GC) and gastroesophageal junction adenocarcinoma (GEJC) was to assess the quality of HER2 testing, which is part of routine assessment to decide target treatment initiation. In addition, exploratory analysis has been performed to identify variables influencing discordance between a local and central evaluation of HER2 overexpression determination. From July 2012 to February 2014, this observational study in France tested tumor samples from patients with GC or GEJC, regardless of disease stage. HER2 positive status was defined as immunohistochemistry (IHC) 3+ or IHC2+/In Situ Hybridization (ISH)+. The true concordance was assessed between local and centralized HER2 status using a Kappa coefficient. To identify variables influencing the discordance, supervised data mining models such as decision tree and random forest were used. The targeted outcome measure was the false negative specimens. 394 specimens from 367 pts were analyzed by 19 local laboratories. Mean age was 66±13 years, 69 % male. The true concordance between HER2 status assessed by local and centralized laboratories was acceptable with a kappa of 0.69 (95CI [0.60-0.78]). The discordance rate was 9% (95CI [6-12]), with 27% (20/73) of false negative and 5% (16/320) of false positive. Decision tree showed that 3 variables were modeling the 20 false negative specimens: percentage of stained cells, tumor heterogeneity and IHC results. 15 of cases had <60% of stained cells and were all heterogeneous, whereas the 5 remaining cases with ≥60 % of stained cells were all IHC2+. The main reason for HER2 status discordance between local and centralized analyses was the tumor heterogeneity in gastric cancer. Given the short history of the HER2 scoring system developed for gastric cancer, a retesting of specimens by a centralized ISH laboratory for heterogeneous tumors could be advisable to ensure the best therapeutic decision.

HER-France is a French national web-based database focused on HER2 status in breast cancer and provided by 125 Pathology Laboratories (PL) since October 2011. It was developed for a local and national monitoring of HER2 but contains also additional data like the SBR grade, ER and PR status, and Ki67 index. The objective of this study is to allow PL to compare their actual HER2 rate, not only with the national data (obtained by aggregation of data of all participating labs) but also to their own expected HER2 rate estimated through the additional collected data. To guarantee best pre-analytical conditions, only results on core-biopsies were taken into account. For the prediction model 30,777 sample results from 2014 were used. Different models were built such as penalized regression and random forest to predict HER2 positivity. To evaluate the performance of the models and choose the most suitable, the database was divided in two subsets: the training (70% of the database) and the test one (30%). The models were built on the training data and evaluate on the test dataset. An accuracy measure (AUC) was used to compare models, with AUC=1 indicating a perfect fitting, whereas AUC under 0.5 indicating a random guess. HER2 positivity rate was 12%. The most accurate model found was the penalised regression as the prediction accuracy (AUC=0.78) proved good ability for HER2 modeling. The significant factors associated with HER2 positivity were the higher SBR grade and Ki67, lower RO and RP status and ductal histological subtype. This study added Ki67 as a strong factor of HER2 positivity nearby histological subtype, grade and hormonal status as previously mentioned by Ruschoff et al. In the future, these results should be used to develop a tool integrated within Her-France functionalities in order to allow pathologists to better monitor their practices.