Home > Research > Publications & Outputs > A Self-Learning Fuzzy Classifier with Feature S...
View graph of relations

A Self-Learning Fuzzy Classifier with Feature Selection for Intelligent interrogation of mid-IR spectroscopy data from exfoliative cervical cytology using selflearning classifier eClass.

Research output: Contribution to journalJournal article


<mark>Journal publication date</mark>2008
<mark>Journal</mark>International Journal of Computational Intelligence Research
Issue number4
Number of pages10
Pages (from-to)392-401
<mark>Original language</mark>English


Abstract: The development of a predictive tool for the diagnosis of cancer is required to be both specific and sensitive, providing new information to increase understanding of the disease. We set out to determine if we could achieve this, and improve the current correct diagnosis rate of cervical cancer by combining ATR-FTIR spectroscopy with a self-learning fuzzy classifier, eClass. Cytology samples were acquired from normal, lowgrade squamous intraepithelial and high-grade squamous intraepithelial patients. Interrogation of normal and precancerous lesions was performed by ATR-FTIR spectroscopy to obtain 10 spectra from each sample. Following pre-processing (baseline correction and normalization) the data were analyzed using eClass which is characterized by being automatic, datadriven, transparent, computationally-efficient, and effectively providing high classification rates for complex non-linear and multivariate problems. An important characteristic of eClass is its ability to select features automatically based on the accumulated contribution of each of a large set of initial features (wavenumbers of the spectra). In this way, the classifier structure (a fuzzy rule base) can evolve gradually to exclude confounding factors such as inter-individual variation, and develop according to the changing requirements of a data stream i.e., identify risk biomarkers of progression towards transformation. The structure of the proposed evolving fuzzy classifier consists of three two-class classifiers connected in a cascade fashion, which provide a classification rate of 77% by using ten-fold cross-validation with unknown data.