Linking Chemical Parameters to Sensory Panel Results through Neural Networks To Distinguish Olive Oil Quality

Abstract

A wide variety of olive oil samples from different origins and olive types has been chemically analyzed as well as evaluated by trained sensory panelists. Six chemical parameters have been obtained for each sample (free fatty acids, peroxide value, two UV absorption parameters (K232 and K268), 1,2-diacylglycerol content, and pyropheophytins) and linked to their quality using an artificial neural network-based model. Herein, the nonlinear algorithms were used to distinguish olive oil quality. Two different methods were defined to assess the statistical performance of the model (a K-fold cross-validation (K = 6) and three different blind tests), and both of them showed around a 95-96% correct classification rate. These results support that a relationship between the chemical and the sensory analyses exists and that the mathematical tool can potentially be implemented into a device that could be employed for various useful applications.

Machine Learning and Feature Selection Applied to SEER Data to Reliably Assess Thyroid Cancer Prognosis

“…ANNs -training and optimizing MLPs. MLPs are the most employed type of ANN 45 , and as any supervised model, they require each data point to be labelled ("0 s" for alive cases and "1 s" for COD-TC cases). Inside every MLP there is a set of weighted parameters (or weights) that connect every unit (nodes and neurons) from one layer with all units in neighbouring layers.…”

“…The MLP uses the training set to modify the weights, and the verification set to evaluate the performance of the model intrinsically with data not employed to change these weights. In other words, the verification dataset is a group of samples that the MLP utilizes to ensure it avoids overfitting for the training dataset and is able to generalize for external data 45 .…”

“…The transfer function restricts the range of the values given by every neuron. In this case, the non-linear sigmoid function has been employed, which limits data between 0 and 1 45 .…”

“…On the other hand, a high NHN could lead to overfit systems that are not able to generalize well for data that is external to the learning dataset. A heuristic method has been employed to optimize the NHNs, testing all possibilities within a logical window that would never lead to models with NHNs lower than 3 or less than a 50-to-1 sample-to-weight ratio (to avoid overfitting) 45 .…”

“…For this process, the databases are randomly divided into three: training, verification, and test ("blind") datasets, containing approximately 70%, 20%, and 10% of the samples, respectively. Furthermore, this process was carried out three times for each model to ensure robustness and flexibility (three different random divisions of data) and the final reported statistical performances result from the averages of these three tests (Table 3) 45 . Also, ROC curves for each classifier have been depicted for further evaluation (Fig.…”