Development and optimisation of chemometric techniques for the evaluation of meat freshness

Muscle foods such as meat, fish and poultry are an integral part of human diet. Over time, such food succumbs to spoilage, resulting from various intrinsic and extrinsic factors, the most significant of which is microbial activity. Spoilage changes the organoleptic properties of meat, rendering it unacceptable to the consumer, and may ultimate result in the food becoming toxic. Spoilage is therefore of major commercial and public health interest. This thesis describes the development and application of a novel suite of software tools designed to support novel instrumental approaches for the accurate, rapid and inexpensive evaluation of meat freshness. A pipeline was built for the analysis of highly heterogeneous data obtained by a diverse range of high-throughput techniques across four three-class case studies. As a first step, PCA was applied for dimensionality reduction, feature extraction and exploratory analysis. PLS-DA and SVMs were employed as classifiers, and classification ensembles implemented as a means of improving classification accuracy. Rigorous validation and evaluation methods based on bootstrapping and permutation testing were applied to ensure that the performance metrics are representative of real-world application, and to ascertain the statistical significance of the results. This was made possible by the development of an advanced optimisation approach, which reduced the computational demands of SVM tuning by up to ~ 90× times. The functionality of the pipeline was further enhanced by exploiting GPA and CPCA as data fusion techniques, to evaluate whether better classification accuracy is achieved when integrated as opposed to standalone datasets are used. SVM ensembles proved to be the most powerful and accurate classification method since they produced consistently higher prediction rates ( ) than PLS-DA. Among the analytical techniques, HPLC was established as the most diagnostic method for the assessment of meat freshness, with a of 80%. Among the two data fusion techniques, CPCA outperformed GPA. However, CPCA only exceeded standalone HPLC in a minority of cases, presenting an overall of 82%.