Machine learning-driven screening of cosmetic allergens: Integrating consumer reviews of cosmetics

Summary: Introduction: As cosmetic use increases, adverse reactions pose growing clinical and public health concerns. This study leverages machine learning to predict risky cosmetics and ingredients by integrating data on ingredients and online-cosmetic consumer reviews describing adverse reactions. Method: The method involves collecting ingredient and review data, preprocessing to recognize fake reviews and extract aspects, and training models to identify risky cosmetics and ingredients associated with adverse reactions. Finally, suspect ingredients were primarily verified by patch test. Results: Among the machine learning models we used, the CatBoost model demonstrated the best predictive performance in suspected allergen screening, achieving an area under the curve (AUC) of 0.74. It identified 13 suspected allergens, among which 11 (84.6%) demonstrated positive reactions in a clinical cohort of 22 participants. Sorbitan sesquioleate exhibited the highest positive rate among participants at 31.8%, followed by xanthan gum, Pigment Red (CI16035) and propylene glycol, each with a positive rate of 22.7%. Conclusion: This study establishes a machine learning-driven framework for time-efficient and cost-effective identification of cosmetic allergens through large-scale analysis of real-world consumer reviews. Our approach leverages publicly accessible online consumer data of cosmetics, and represents the first instance of cosmetic adverse reaction risk assessments solely based on consumer reviews. Besides, primary verification confirms that the method effectively screened potential allergens in cosmetics. Nan Huang^{1, 3, 4}, Li Li^{1, 3}, Bo Fang², Shengjie Min², Lidan Xiong³, Wei Hua^{1, 3} 1. Department of Dermatology, West China Hospital of Sichuan University, Chengdu, Sichuan, China. 2. Zhiyuan Big Data Technology Co., Ltd., China Electronics Technology Group Corporation, Beijing, Beijing, China. 3. Cosmetics Safety and Efficacy Evaluation Center, West China Hospital of Sichuan University, Chengdu, Sichuan, China. 4. Laboratory of Dermatology, Clinical Institute of Inflammation and Immunology, Frontiers Science Center for Disease-related Molecular Network,, West China Hospital of Sichuan University, Chengdu, Sichuan, China. Bioinformatics, Computational Biology, and Imaging