Resumo de eventos cient??ficos
Breast cancer subtypes diagnostic via high performance supervised machine learning
Registro en:
0000-0001-7404-9606
0000-0002-0029-7313
Autor
FAROOQ, SAJID
DEL-VALLE, MATHEUS
SANTOS, MOISES O. dos
NASCIMENTO, SOFIA
BERNARDES, EMERSON S.
ZEZELL, DENISE M.
INTERNATIONAL CONFERENCE ON CLINICAL SPECTROSCOPY, 12th
Resumen
Aim: Breast cancer molecular subtypes are being used to improve clinical decision. The Fourier
transform infrared (FTIR) spectroscopic imaging, which is a powerful and non-destructive technique,
allows performing a non-perturbative and labelling free extraction of biochemical information towards
diagnosis and evaluation for cell functionality. However, methods of measurements of large areas of
cells demand a long time to achieve high quality images, making its clinical use impractical because
of speed of data acquisition and dearth of optimized computational procedures. In order to cope with
these challenges, Machine learning (ML) technologies can facilitate to obtain accurate prognosis of
Breast Cancer (BC) subtypes with high action ability and accuracy.
Methods: Here we propose a ML algorithm based method to distinguish computationally BC cell lines.
The method is developed by coupling K neighbors Classifier (KNN) with Neighborhood Component
Analysis (NCA) and NCA-KNN methods enables to identify BC subtypes without increasing model
size as well additional parameters.
Results: By incorporating FTIR imaging data, we show that using NCA-KNN method, the classification
accuracies, specificities and sensitivities improve up to 97%, even at very low co-added scan (S_4).
Moreover, a clear distinctive accuracy difference of our proposed method was obtained in comparison
with other ML supervised models.
Conclusion: For confirming our model results performance, the cross validation (k fold = 10) and
receiver operation characteristics (ROC) curve were used and found in great agreement, suggest a
potential diagnostic method for BC subtypes, even with small co-added scan < 8 at low spectral
resolution (4 cm-1).