The predominant components of natural fish oil are glycerides, mainly triglycerides (TGs). One rapid way to measure glycerides is rapid evaporative ionization mass spectrometry (REIMS): a technique originally developed for the testing of tumours during surgery, sometimes called the “ion knife”. This study (purchase required, $USD 31.50) used an electric soldering iron ion source (ESII) coupled with REIMS to develop a classification model for fish oil species, then used explainable AI to identify the species-dependent triglyceride markers driving the model.
Six fish oil types (Trichiurus lepturus, Larimichthys crocea, Larimichthys polyactis, Pampus argenteus, Scomber scombrus, Thunnus obesus) were analyzed, yielding 143 lipid ions. The authors found that their lipid profiles showed substantial similarities among different biological origins, highlighting the challenge of biological origin authentication.
The authors applied Machine Learning algorithms and explainable AI to the datasets (SHapley Additive exPlanations (SHAP) framework, based on cooperative game theory, enabling quantitative interpretation of individual feature contributions and providing insight into model decision mechanisms).
They found that the CatBoost model gave the best classification performance. SHAP-based interpretation enabled interpretation of model decision mechanisms and identification of key lipid discriminant markers. Twelve key discriminant markers selected by global feature importance enabled the construction of a simplified CatBoost model, improving the classification accuracy still further.
The authors conclude that ESII-REIMS combined with SHAP-guided interpretable machine learning provides a rapid and effective workflow for fish oil authentication and the identification of chemically interpretable markers from complex lipid fingerprints.
Photo by Alek Newton on Unsplash
Comments