Machine learning-driven identification of key chemical determinants of tobacco leaf sensory irritation.

INTRODUCTION: To identify the key chemical components affecting the sensory irritation of tobacco leaves, a binary classification model for high and low irritation was constructed based on 78 chemical components and sensory evaluation scores of 353 tobacco leaf samples. METHODS: First, the median absolute deviation (MAD) method was applied to remove outliers from the high- and low-irritation samples. Then, the ReliefF algorithm was applied for dimensionality reduction, selecting 33 core features to eliminate data redundancy. Using the selected features, a random forest (RF) algorithm was employed to build the classification model, and the optimal number of decision trees was determined to be 60. RESULTS: The ReliefF-RF model achieved an accuracy of 84.38% on an independent test set, with precision, recall, and F1-score all at 86.49%, outperforming the original RF model as well as other machine learning models such as support vector machine (SVM) and k-nearest neighbors (KNN). Through feature importance evaluation, eight key chemical indicators were identified: total nitrogen, total alkaloids, cryptochlorogenic acid, oleic acid + linolenic acid, reducing sugar, sugar-nitrogen ratio, Fru-Asp, and neochlorogenic acid. DISCUSSION: SHapley Additive exPlanations (SHAP) analysis revealed that higher levels of nitrogenous compounds were strongly associated with increased irritation, whereas elevated levels of sugar components, specific organic acids, and amino acid derivatives were associated with reduced irritation. Notably, Fru-Asp exhibited a complex non-linear response, where both extremely high and low levels contributed to higher irritation. This study provides a useful reference and data support for the targeted regulation of cigarette irritation.