L. Slembrouck (Leuven, Belgium)
Author Of 1 Presentation
- L. Slembrouck (Leuven, Belgium)
- I. Sebastian (Leuven, Belgium)
- A. Martinez (Leuven, Belgium)
- I. Vanden Bempt (Leuven, Belgium)
- H. Wildiers (Leuven, Belgium)
- A. Smeets (Leuven, Belgium)
- A. Van Rompuy (Leuven, Belgium)
- C. Weltens (Leuven, Belgium)
- K. Punie (Leuven, Belgium)
- E. Van Nieuwenhuysen (Leuven, Belgium)
- S. Han (Leuven, Belgium)
- I. Nevelsteen (Leuven, Belgium)
- P. Neven (Leuven, Belgium)
- G. Floris (Leuven, Belgium)
51P - The development of a predictive model for risk results of the 70-gene signature
Abstract
Background
Multigene signatures (MGS) select women with oestrogen receptor positive human epidermal growth factor receptor 2 negative (ER+/HER2-) breast cancer for whom adjuvant chemotherapy can be avoided. As MGS are expensive and require additional test material and time, models based on clinical-pathological features have been developed, mainly for OncotypeDX® but none exist for MammaPrint® (MP). We evaluated these models and the possibility of creating a new model for MP high/low genomic risk result.
Methods
This retrospective study, approved by the Ethics Committee of University Hospitals Leuven (S63323), included patients diagnosed at University Hospitals Leuven between 2013 and 2020 with primary operable ER+/HER2- lymph node negative or positive breast cancer. Tumour tissue of 143 patients was analysed by MP. Seven statistical models were computed: Magee equations (1, 2, 3 and average), Memorial Sloan Kettering simplified score, Breast Cancer Recurrence Score Estimator, OncotypeDXCalculator, MyMammaPrint.com, new Adjuvant! Online and PREDICT v2.1. The outcome of these models (N=10), clinical-pathological features (N=38) and the combination of both were tested as input for a new model. Patients were split (75/32/36) in train, test and validation sets. Feature selection was performed through an automated algorithm. Random forests (RF), support vector machines (SVM) and linear discriminant analysis (LDA) models were tested.
Results
The number of selected features and the accuracies of the machine learning models are shown in the table. The best performance was obtained with the combined feature set resulting in accuracies up to 75% with a sensitivity of 78% (14/18 true low risk) and a specificity of 72% (13/18 true high risk). These models had better MP risk prediction accuracies (75%) compared to the existing statistical models (<67%). Number of selected features and accuracy of the models
Accuracy (%) # features RF SVM LDA 19 75 75 69 5 69 70 68 25 75 75 72
Conclusions
Our predictive model, based on clinical-pathological features, can be used for patient selection for MP and therefore reduce cost, tumour tissue and time.
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
L. Slembrouck: Honoraria (self): Agendia. All other authors have declared no conflicts of interest.
Presenter Of 1 Presentation
- L. Slembrouck (Leuven, Belgium)
- I. Sebastian (Leuven, Belgium)
- A. Martinez (Leuven, Belgium)
- I. Vanden Bempt (Leuven, Belgium)
- H. Wildiers (Leuven, Belgium)
- A. Smeets (Leuven, Belgium)
- A. Van Rompuy (Leuven, Belgium)
- C. Weltens (Leuven, Belgium)
- K. Punie (Leuven, Belgium)
- E. Van Nieuwenhuysen (Leuven, Belgium)
- S. Han (Leuven, Belgium)
- I. Nevelsteen (Leuven, Belgium)
- P. Neven (Leuven, Belgium)
- G. Floris (Leuven, Belgium)
51P - The development of a predictive model for risk results of the 70-gene signature
Abstract
Background
Multigene signatures (MGS) select women with oestrogen receptor positive human epidermal growth factor receptor 2 negative (ER+/HER2-) breast cancer for whom adjuvant chemotherapy can be avoided. As MGS are expensive and require additional test material and time, models based on clinical-pathological features have been developed, mainly for OncotypeDX® but none exist for MammaPrint® (MP). We evaluated these models and the possibility of creating a new model for MP high/low genomic risk result.
Methods
This retrospective study, approved by the Ethics Committee of University Hospitals Leuven (S63323), included patients diagnosed at University Hospitals Leuven between 2013 and 2020 with primary operable ER+/HER2- lymph node negative or positive breast cancer. Tumour tissue of 143 patients was analysed by MP. Seven statistical models were computed: Magee equations (1, 2, 3 and average), Memorial Sloan Kettering simplified score, Breast Cancer Recurrence Score Estimator, OncotypeDXCalculator, MyMammaPrint.com, new Adjuvant! Online and PREDICT v2.1. The outcome of these models (N=10), clinical-pathological features (N=38) and the combination of both were tested as input for a new model. Patients were split (75/32/36) in train, test and validation sets. Feature selection was performed through an automated algorithm. Random forests (RF), support vector machines (SVM) and linear discriminant analysis (LDA) models were tested.
Results
The number of selected features and the accuracies of the machine learning models are shown in the table. The best performance was obtained with the combined feature set resulting in accuracies up to 75% with a sensitivity of 78% (14/18 true low risk) and a specificity of 72% (13/18 true high risk). These models had better MP risk prediction accuracies (75%) compared to the existing statistical models (<67%). Number of selected features and accuracy of the models
Accuracy (%) # features RF SVM LDA 19 75 75 69 5 69 70 68 25 75 75 72
Conclusions
Our predictive model, based on clinical-pathological features, can be used for patient selection for MP and therefore reduce cost, tumour tissue and time.
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
L. Slembrouck: Honoraria (self): Agendia. All other authors have declared no conflicts of interest.