Predicting fishiness off-flavour and identifying compounds of lipid oxidation in dairy powders by SPME-GC/MS and machine learning

This study examined the correlation between untargeted solid-phase micro-extraction gas chromatography/mass spectrometry (SPME-GC/MS) data and sensory fishiness of dairy powders fortified with long-chain polyunsaturated fatty acids and iron. A machine learning approach for sensory prediction from raw CG/MS data is discussed and its potential for determining key contributing compounds shown. To find peak correspondence and to correct retention time shifts, GC/MS raw data of different samples were aligned using dynamic programming. Sensory modelling and prediction was done without prior peak identification in the mass spectral library. Regression was achieved by multiple classification tasks using a Random Forest model. The obtained sensory predictions showed good accuracy both in leave-one-out evaluation and on a separate powder sample test set. GC/MS peaks suggested by Random Forest to significantly contribute to fishiness were identified to be from the chemical classes of alcohols, ketones, aldehydes and furans.