Abstract:

To enhance the predictive performance of computer simulation-based alternative test methods for assessing the eye irritation potential of cosmetic ingredients, this study proposes a machine learning hybrid model that combines ensemble learning algorithms with similarity algorithms. The research collected data from six in vitro alternative test results, such as the fluorescein leakage test and neutral red uptake test, for 84 cosmetic ingredients, as well as data from in vivo test results. The predictive assessment effectiveness of the machine learning hybrid model was validated based on the collected test result data. The final experimental results indicate that the proposed machine learning hybrid model performs well in predicting the irritancy of cosmetic ingredients based on in vitro alternative test result data, with a predictive assessment accuracy rate of 100% in the validation experiment results. Furthermore, to improve the convenience, efficiency, and interpretability of the machine learning hybrid model in practical use, this study further proposes a method for evaluating the eye irritation potential of cosmetic ingredients based on a stratified combination of in vitro alternative tests, which is based on the hybrid model. This method includes a quick reference guide for predictive evaluation generated through computer simulation and a hierarchical combination evaluation strategy derived from experimental data analysis. By integrating the results of in vitro tests conducted under this tiered strategy, it enables the rapid prediction of eye irritation potential of cosmetic ingredients.

Key words: eye irritation, alternative methods, machine learning, stratified combination assessment