Combined methods of optical spectroscopy and artificial intelligence in the assessment of experimentally induced non-alcoholic fatty liver

Background and Objective: Due to the existing prevalence of nonalcoholic fatty liver disease (NAFLD) and its relation to the epidemic of obesity in the general population, it is imperative to develop detection and evaluation methods of the early stages of the disease with improved efficacy over the current diagnostic approaches. We aimed to obtain an improved diagnosis, combining methods of optical spectroscopy -diffuse reflectance and fluorescence- with statistical data analysis applied to detect early stages of NAFLD. Methods: Statistical analysis scheme based on quadratic discriminant analysis followed by canonical discriminant analysis were applied to the diffuse reflectance data combined with endogenous fluorescence spectral data excited at one of these wavelengths: 330, 365, 385, 405 or 415 nm. The statistical scheme was also applied to the combinations of fluorescence spectrum (405 nm) with each one of the other fluorescence spectra. Details of the developed software, including the application of machine learning algorithms to the combination of spectral data followed by classification statistical schemes, are discussed. Results: Steatosis progression was differentiated with little classification error (≤1.3%) by using diffuse reflectance and endogenous fluorescence at different wavelengths. Similar results were obtained using fluorescence at 405 nm and one of the other fluorescence spectra (classification error ≤1.0%). Adding the corresponding areas under the curves to the above combinations of spectra diminished errors to 0.6% and 0.3% or less, respectively. The best results for the compounded reflectance-plus-fluorescence spectra were obtained with fluorescence spectra excited at 415 nm with a total classification error of 0.2%; for the combination of the 405nm-excited fluorescence spectrum with another fluorescence spectrum, the best results were achieved for 385 nm, for which total relative classification error amounted 0.4%. The consideration of the area under the spectral curves further improved both classifiers, reducing the error to 0.0% in both cases. Conclusion: Spectrometric techniques combined with statistical processing are a promising tool to improve steatosis classification through a label free approach. However, statistical schemes here applied, might result complex for the everyday medical practice, the designed software including machine learning algorithms is able to render automatic classification of samples according to their steatosis grade with low error.