Content uploaded by David Pau
Author content
All content in this area was uploaded by David Pau on Jun 29, 2022
Content may be subject to copyright.
Evaluation of breast cancer HER2 status accuracy using a predictive
model based on HER-France real world national database.
Caroline Egele1, David Pau2, Justine Rabut3, Dominique Fetique1, Jessica Martin2, Julien Dupin4, Jean-Pierre Bellocq1
1AFAQAP and Hôpitaux Universitaires de Strasbourg, Strasbourg, France ; 2Roche, Boulogne-Billancourt, France; 3Lincoln for Roche; 4ITM stat for Roche
BACKGROUND
HE R-France is a French national database focused on HER2
status in breast cancer and provided by 125 Pathology
Laboratories (PL) since October 2011.
This web-based data collection has been developed by
the AFAQAP (French national quality control association
in histo- and cyto-pathology) for a local and national
monitoring of HER2 status and key related factors like
the SBR grade, estrogen receptor (ER), progesterone
receptor (PR), histological type and Ki67 index.
OBJECTIVE
Provide description and model the HER2 status with the
key related factors.
Allow PL to compare their actual HER2 rate to their own
expected HER2 rate estimated through the collected data.
RESULTS
REFERENCES
¹ Chenard MP, and al. Presented at the 2016 USCAP annual meeting, Seatle, USA
² Wolff AC, and al. J Clin Oncol 2013, 31:3997-4013
³ Penault-Lorca F, and al. Annales de pathologie (2014) 34, 352—365
4
Tufféry S, Data mining et statistique décisionnelle, 2012, Technip
5 Rüschoff J, and al. Presented at the 2015 ASCO annual meeting, Chicago, IL, USA. P11062
CONCLUSION
HERFrance is a rich database to predict HER2 status using datamining methods.
CNB characteristics are useful to describe and model the HER2 status, for this reason laboratories should
be encouraged to fill in overall data such as Ki67 index or tumor size.
In the future, these results will be used to develop a tool in order to predict the HER2 positivity rate and
allow pathologists to better monitor their quality control process.
Overall population
55,571 CNB
20,518 CNB in the
training sample
10,259 CNB in the
validation sample
11,613 CNB in the
training sample
5,807 CNB in the
validation sample
Population selected after
publication of ASCO/CAP
Guideline 2013
with HER2 status available
30,777 CNB
Population selected after
publication of ASCO/CAP
Guideline 2013 with Ki67
and HER2 status available
17,420 CNB
Model 1
Model 2
Pop 1
Pop 2
Population with HER2
status available
50,602 CNB
Figure 1: Population flow-chart
Overall population
(n=55,571)
Population 1¹
(n=30,777)
Population 2²
(n=17,420)
HER2+ 6,089 (12.0%) 3,669 (11.9%) 2,045 (11.7%)
HER2- 44,160 (87.3%) 26,804 (87.1%) 15,165 (87.1%)
Equivocal 353 (0.7%) 304 ( 1.0 %) 210 (1.2%)
Missing 4,969
¹ Population from ASCO/CAP Guideline 2013 with HER2 status available
² Population from ASCO/CAP Guideline 2013 with Ki67 and HER2 status available
Table 1: Description of the HER2 status
n (%) HER2+
(n=6,089)
HER2-
(n=44,160)
Equivocal
(n=353)
SBR grade
I 302 (5) 11,594 (28) 38 (11)
II 3,192 (56) 24,351 (58) 223 (66)
III 2,238 (39) 6,148 (15) 79 (23)
Missing 357 2,067 13
Histological tumor type
Ductal 5,511 (93) 34,770 (80) 316 (92)
Lobular 267 (5) 6,491 (15) 20 (6)
Other 153 (3) 2,082 (5) 9 (3)
Missing 158 817 8
Tumor size
pT1a/b 41 (28) 483 (37) 3 (50)
pT1c 43 (29) 467 (36) 0 (0)
pT2/3 62 (42) 340 (26) 3 (50)
Missing 5,943 42,870 3 47
ER positive¹ 3,775 (63) 38,560 (88) 297 (86)
PR positive¹ 2,576 (43) 32,083 (73) 231 (67)
Ki67 positive² 2,294 (85) 13,259 (53) 140 (77)
¹ French classification: positive if tumor cells stained ≥ 10%
² French classification: positive if tumor cells stained > 10%
Table 2: Characterics of CNB according to HER2 status
AUC=1 indicating a perfect fitting, whereas AUC under 0.5 indicating a random guess
The prediction accuracy proved a good ability for HER2 modelling (AUC=0.76 and
0.78, figure 2a and 2b).
The multivariate analysis confirmed that the significant factors associated
with HER2 positivity status were the higher SBR grade and Ki67, lower ER
and PR status and ductal histological subtype.
This study presented Ki67 as an additional significant factor of HER2 positivity nearby
histological subtype, grade and hormonal status as previously mentioned in literature 5.
This graph represents the estimation of the HER2 rate according to the HER2 rate observed in population selected after the publication of ASCO/CAP Guideline
2013 with HER2 status available for laboratories having recorded more than 100 CNB. The point size shows the number of CNB analyses by the laboratories.
The farther the point is from the bisecting line, the farther is the difference between the actual and estimated HER2 overexpression rate.
Figure 2a: Model 1
ROC curve on validation dataset
Figure 2b: Model 2
ROC curve on validation dataset
Figure 3 : Accuracy of the HER2 overexpression estimation rate by laboratory.
The HER-France database included 55,571 CNB and the HER2 status
was fully evaluated for 50,602 CNB (Figure 1).
The HER2 positivity rate was 12.0% and remained stable in the population
selected after the publication of ASCO/CAP Guideline 2013 (Table 1).
Patients with HER2 overexpression presented more elevated SBR grade,
larger tumor size, ductal histological subtype, higher Ki67 index and lower
PR and ER status than HER2 negative patients (Table 2).
METHODS
Methodology and primary results of HERFrance database was presented at
USCAP 20161.
To guarantee best pre-analytical conditions, only results on core needle biopsies
(CNB) were taken into account.
HER2 status considerations
For the prediction model, results from 2014 were used in order to be in line with
the guideline of the American Society of Clinical Oncology/College of American
Pathologists² (ASCO/CAP) and GEFPICS³ to test Human Epidermal Growth Factor
Receptor 2 (HER2) in breast cancer.
Data-mining methodology
The estimated rate of HER2 positivity have been obtained using datamining
methods which are usually applied on large and complex datasets to extract
unknown information.
Different methods were used such as penalized regression and random forest to
predict HER2 positivity. We present here the results of the most efficient method:
the penalized regression4.
The modelling consists to forecast the actual HER2 status (HER2+ vs HER2- and
equivocal) according to predictive factors. As the Ki67 index was collected in 53%
of CNB, two models were implemented to address this issue (model 1 and 2,
Figure 1). Tumor size was only available for CNB linked to surgical specimen (3%
of completed data), then it was not considered in the model.
Model validation and performance assessment
To evaluate the performance of the models, we compared estimate HER2 status to
actual status. Thus, the database was divided in two random subsets: the training
(70% of the database) and the validation one (30%). The models were built on
the training data and evaluated on the validation dataset (Figure 1).
To compare models, Receiver Operating Characteristics (ROC) curve was plotted
and Area Under the Curve (AUC) was evaluated as an accuracy measure (Figure
2a and 2b).
A= 728 CNB in this laboratory,
estimated HER2+ rate=14.7%,
actual HER2+ rate=14.6%.
B= 113 CNB in the laboratory,
estimated HER2+ rate=8.8%,
actual HER2+ rate=16.8%.
#PMD4
POSTER_HERFrance_90x120.indd 1 19/10/16 09:56