Feature selection and classification approaches for biometric and biomedical applications

The focus of this thesis is to develop feature selection and classification models for biometric and biomedical problems. The models developed can be used in designing applications for predicting diseases and validate identity. The main challenge when dealing with features generated from biological systems are the variability in freatures with respect to the same class due to factures like age, gender, demographics. This dissertation addresses the problem in biological features and ways to improve classification models using feature selection or change in classification model. In this work three different biological data systems were studied. Two biomedical datasets, namely a Cardio Vascular Disease (CVD) prediction system using physiological features, a Stroke prediction model using retinal image features and a Biometric database using dynamics of handwritten signatures were analysed. In the CVD prediction model, an Australian population database (BMES) was used to predict 10 year CVD outcome using existing linear Farmingham model and compared with a new model based on non-linear Support Vector Machine (SVM) decision boundary which was able to improve the prediction accuracy. An SVM based feature ranking was also used to validate the features based in the Framingham model. It was observed that while Framingham features are relevant a non-linear classification model can improve the CVD prediction model. For retinal image based stroke prediction model, the commonly used green channel for retinal Bessel feature extraction was replaced with an adaptive Principal Component Analysis (PCA) based colour space. It was observed that while Green channel is generally considered to have good contract between vessels and background and PCA based colour space has a much better contract as it is adaptive to the image. Further extending the study of impace of adaptive feature selection on biological systems a signature database was used to extract dynamic global features. Dynamic features are based on dynamics of a signature like speed, pressure, velocityrather than the shape of a signature they are harder to forge. Freatures extracted were ranked using traditional global feature selection models and a class specific feature selection model developed as part of the research. It was observed that by using class specific feature selection model the classification accuracy can be improved when compared to a global feature selection model. This includes that in biological systems in the intra class variability can be used as a measure of feature stability. The absence of real world counter examples to simulate the forgery class also adds weight to the feature selection model based on intra class viability. The false rejection rate was significantly reduced using a class specific feature selection model without much reduction in the false acceptance rate.