Novel machine learning techniques based hybrid models (LR-KNN-ANN and SVM) in prediction of dental fluorosis in groundwater

Abstract

Studies have shown that excessive intake of fluoride into human body from drinking water may cause fluorosis adversely affects teeth and bones. Fluoride in water is mostly of geological origin and the amounts depend highly on many factors such as availability and solubility of fluoride minerals as well as hydrogeological and geochemical conditions. Chemical methods usually accomplish fluoride analysis in drinking water. The chemical methods are expensive, labor-intensive and time-consuming in general although accurate and reliable results are obtained. An alternative cost-effective approach based on machine learning (ML) technique is investigated in this study. Furthermore, most effective input parameters are selected via proposed Simulated Annealing (SA) search scheme. Selected subset (SAR, K+, NO3-, NO2-, Mn, Ba and Fe) by SA algorithm exhibited high correlation coefficient values of 0.731 and strong t test scores of 5.248. On the other hand, most frequently selected individual features were identified as NO3-, NO2-, Fe and SAR by vote map. The results of experiments revealed that selected feature subset improves the prediction performance of the learning models while feature size is reduced substantially. Thus it eventually enabled determination of fluoride in a cheap, fast and feasible way.