A STACKED ENSEMBLE MODEL FOR DIAGNOSIS OF CRANIOCERVICAL (HEAD AND NECK) CANCER

Abstract

Head and neck cancers (HNC) otherwise known as craniocervical cancers are diseases that grow out of control in the head and neck regions. 5-50% of all cancers globally and are often associated with facial disfigurement that causes cosmetic embarrassment due to delay in referral to the specialist concerned (ENT surgeons). They are of various types and as such characterized as a heterogeneous disease. The incidence of HNC is on the increase owing to several factors. There is often late presentation that can result in loss of lives (mortality) especially in Africa due to paucity of specialists. These challenges prompted the development of a stacked ensemble model for diagnosis of HNC to facilitate prompt referral. The data used for this work were collected from the ENT/Head and Neck and pathology departments in University Medical Sciences Teaching Hospital, Akure, Federal Medical Centre, Owo and Obafemi Awolowo University Teaching Hospital, Ile Ife. The dataset consists of 1473 instances with 18 features. Three filter based feature selection methods: Consistency, Information Gain and Chi Square were used to select the relevant features from the HNC dataset. Three supervised learning algorithms were deployed for the base learners: Decision Tree (C4.5), K-Nearest Neighbors and Naïve Bayes. The predictions of the base learners were combined and passed to meta learners: Multinomial Logistic Regression (MLR) and Logistic Model Tree (LMT). The results showed that Consistency, Information Gain and Chi Square feature selection methods with stacked MLR were 94.90%, 95.38% and 95.38% respectively. Consistency, Information Gain and Chi Square feature selection methods with stacked LMT were 94.77%, 95.11% and 94.91%. It was deduced that both Information Gain and Chi Square with stacked MLR produced highest results.