Chemometrics analysis for the detection of dental caries via ultraviolet absorption spectroscopy / Katrul Nadia Basri

The prevalence of dental caries is still high and this has raised a major concern to the society and government. Dental caries is a progressive disease that belongs to the group of non-communicable diseases (NCDs). Currently, dental caries has become the first ranking among the NCDs due to its high prevalence. Methods such as visual inspection by the dentist, imaging and illumination techniques have been widely used by the dentist to diagnose the patient’s severity of dental caries. These conventional methods required expert assistance, reagent and were mostly used for diagnostic purpose. Ultraviolet (UV) spectroscopy has a great potential to be an early screening tool for the detection of dental caries. Chemometrics analysis needs to be coupled with UV spectroscopy to correlate with the UV spectra and the caries score based on International Caries Detection and Assessment System (ICDAS). The UV spectra collected in the range 200 – 350 nm shows the absorption at 260-310 nm due to the presence of certain bacteria that cause dental caries. The spectra were split into calibration and validation data using stratified random sampling for each of ICDAS class by a ratio of 80:20. Different preprocessing methods (mean centre, autoscale and Savitzky-Golay smoothing) were applied to the spectra to reduce the noise embedded in the spectra. Classification algorithms such as K-nearest neighbour (KNN), logistic regression (LR), linear discriminant analysis (LDA) and decision tree (DT) were employed to classify the spectra into its ICDAS score. The best performance obtained using Savitzky-Golay smoothing and LDA algorithms after the wavelength selection with the accuracy reported of 0.90. The precision, sensitivity, specificity obtained for the model were 1.00, 0.86 and 1.00 respectively. Artificial neural network (ANN) was performed on the spectra to investigate its feasibility to predict the dental caries. The ANN architecture was optimized by tuning the hyperparameter. The best result of ANN model obtained were 0.85, 0.8, 0.57 and 0.92 for accuracy, precision, sensitivity and specificity. Dimension reduction algorithm such as LDA and CNN were applied on the spectra to reduce the number of variables to be trained. The result obtained has revealed that the combination of LDA-ANN did not improve the performance of the model. The accuracy obtained using CNN was 0.85 for the overall performance of calibration and validation. Model 2 of CNN has maximum performance of validation in terms of accuracy, precision, sensitivity and specificity but the calibration model requires more optimization. The accuracy of the CNN model is comparable with the accuracy of the previous work that utilizing CNN for the imaging data to detect caries (diagnostic tool).