A fuzzy regression model for the prediction of oral cancer susceptibility

Abstract

Precise and accurate predictive models are very important in cancer screening initiatives. The need for new approaches and philosophies in modeling prediction in disease susceptibility studies are influenced by the recent advances in soft computing as well as the questionable accuracy and inapplicability to individual prediction of commonly used statistical analysis techniques. Soft computing especially fuzzy concept is highly suitable for dealing with vague, ambiguous and complex information. The purpose of this research is to develop a computer-prototype using fuzzy linear regression for the prediction of dichotomous response variable in general and disease susceptibility in particular. The objective of this thesis is to present the development,testing and validation of an adaptive fuzzy regression algorithm that can be used for the prediction of binary response variable. In this research a machine learning algorithm is developed to estimate an unknown dependency between a set of given input variables and its corresponding response variable which is binary in nature. Thus the general aim of this study is to develop a computer-prototype using fuzzy regression concept that can be used to predict dichotomous outcome in general and disease susceptibility at individual and group prediction level in specific. The proposed adaptive fuzzy regression prediction model was experimented on an oral cancer data set and validated on a hypertension data set. The model’s prediction performance was measured based on its calibration and discrimination abilities. The model’s prediction performance, interpretation abilities and variable selection ability were then validated against four validation models including the oral cancer clinicians’ predictions, fuzzy logic, fuzzy neural networks and statistical logistic regression predictions. The adaptive fuzzy regression, fuzzy neural network and statistical logistic regression found to have better prediction abilities compared to oral cancer clinicians and fuzzy logic prediction model. IV The adaptive fuzzy regression, fuzzy neural network and statistical logistic regression have compatible prediction performances. Similarly, the three models are capable of finding the ‘optimal’ input predictor set by utilization of different variable selection techniques. Both the adaptive fuzzy regression prediction model and statistical logistic regression prediction model produced transparent prediction equation by which associations between the input predictors and the predicted outcome are clearly projected, a feature that is lacking in fuzzy neural network prediction model. However, the adaptive fuzzy regression prediction model is capable of handling vague, ambiguous relationship between independent and dependent variables which statistical logistic regression failed to address. In conclusion, the good results obtained in this application suggest that the proposed adaptive fuzzy regression prediction model is highly reasonable, desirable and effective in producing a valid and transparent intelligent exploratory predictive model in predicting dichotomous response variable.