Data fusion of multiple‐information strategy based on Fourier transform near infrared spectroscopy and Fourier‐transform mid infrared for geographical traceability of Wolfiporia cocos combined with chemometrics

Abstract

Owing to the widespread concern relating to herb safety and quality, there is a momentum to discriminate the geographical traceability of fungus with multiple‐information technologies. In this study, we attempted to evaluate the fusion strategy of multiple‐information for the geographical traceability of this fungus based on Fourier transform near infrared spectroscopy (FT‐NIR) and Fourier‐transform mid infrared spectroscopy (FT‐MIR) with chemometrics. From all results, (1) comparative visualization of t‐distributed stochastic neighbor embedding (t‐SNE) emerged that the superior effect of FT‐NIR spectroscopy was better than FT‐MIR, which were separated according to the samples of different regions; (2) in view of optimized spectrum, the geographical traceability of extreme learning machine (ELM) model was better than the partial least squares discriminant analysis (PLS‐DA) and support vector machine (SVM) models that the accuracy of training set for all models was 100%; (3) the accuracy of train and test for ELM model were 100% with single spectrum, low‐level, and mid‐level fusion of FT‐NIR and FT‐MIR based on principal components (PCs) and competitive adaptive weighting algorithm (CARS), which was suitable for geographical traceability and evaluation of this fungus. (4) The three models based on successive projection algorithm (SPA) were not dissatisfied for the geographical traceability of this fungus. (5) The result of hierarchical cluster analysis (HCA) was consistent with the above results, which could distinguish samples according to different regions. All results showed that models could be developed as superior model on the foundation of single spectrum based on FT‐NIR, low‐level or mid‐level (except SPA) fusion of ELM, which could be employed in geographical traceability of Wolfiporia cocos, effectively.

Classification of colorectal primer carcinoma from normal colon with mid‐infrared spectra

“…When it comes to diagnosing cancer, these numbers are simply too low not to mention that other studies dealing with colorectal cancer data showed a stable higher than 90% accuracy for as the result of classification, albeit with different machine learning methods 41 . Meanwhile, projects revolving around FT‐IR techniques and spectrum level classification via traditional models or modified neural networks managed to reach similar or higher efficiency scores as well 42–47 . However, in some cases, these studies put a lot more focus on the preprocessing of the spectra or even using the different derivatives as input for the machine learning models, which could definitely be a viable option for us to better our results.…”

“…41 Meanwhile, projects revolving around FT-IR techniques and spectrum level classification via traditional models or modified neural networks managed to reach similar or higher efficiency scores as well. [42][43][44][45][46][47] However, in some cases, these studies put a lot more focus on the preprocessing of the spectra or even using the different derivatives as input for the machine learning models, which could definitely be a viable option for us to better our results. The accuracy can be improved by adding new slides to the cohort and increasing the number of spectra.…”