Application of Machine Learning for the Determination of Damaged Starch Ratio as an Alternative to Medcalf and Gilles Principle

Abstract

As an alternative to the conventional amperometric method used for the determination of damaged starch ratio in wheat flour, two machine learning techniques were applied to a database constructed of 6264 voltammetric data obtained at two different electrodes, two different potassium iodide concentrations, and three different damaged starch ratios. Lift maps were extracted using association rule mining from the voltammetric database to describe electrode behavior and sensitivity to Chopin Dubois units (UCD) values. K-nearest neighbors (KNN) algorithm applied to the voltammetric experiments was able to predict UCD with higher accuracy when KI concentration was low. In addition, current quartiles, scatter, and shifts in lift maps and distinct regions after KNN classification showed that higher sensitivity towards damaged starch ratio is achieved on GC electrode and at low KI concentration.