A method for classifying the results of cosmetic human patch tests suitable for sensitive skin

Abstract

The present application belongs to the technical field of cosmetic safety evaluation, and discloses a human patch test result typing method for cosmetics suitable for sensitive skin, which comprises the following steps: collecting comment data and performing deduplication and noise reduction processing, analyzing the comment data set using a trained RoBERTa-large language model, identifying adverse reaction comments, and calculating adverse reaction scores and adverse reaction rates; performing a patch test and setting a typing line according to the patch test results; extracting product characteristics of the cosmetics; establishing a generalized linear model; testing the likelihood ratio chi-square and significance of each generalized linear model, and screening the best typing of the cosmetics corresponding to the patch test. The method of the present application can effectively evaluate the safety risk of cosmetics for target users.

Description

Technical Field

[0001] This invention belongs to the field of cosmetic safety assessment technology, and in particular, relates to a method for classifying the results of human patch tests for cosmetics suitable for sensitive skin. Background Technology

[0002] Currently, the cosmetics market is experiencing rapid growth. To meet the increasing demand for cosmetics, both the variety and quantity of products have increased significantly. At the same time, cases of skin discomfort or adverse reactions caused by cosmetics are also on the rise, making cosmetic safety a key indicator that consumers increasingly value. Patch testing is one of the most commonly used safety assessment methods in the field of cosmetic technology. It can effectively detect whether cosmetics cause allergic reactions or screen for and identify low-allergenic ingredients, thereby ensuring cosmetic safety.

[0003] Interpreting test results and predicting risks are crucial for patch testing. Currently, the interpretation of patch test results has significant limitations. On the one hand, although studies have confirmed a significant correlation between positive patch test results and adverse reaction reports monitored by provincial adverse reaction monitoring systems, there is a lack of reference samples for the potential risks of cosmetics corresponding to suspicious and weakly positive reactions in patch tests. On the other hand, although the "Technical Guidelines for Human Skin Patch Testing" (public consultation) recommends that, among 30 subjects, the number of cases with Grade 1 adverse skin reactions should not exceed 5 (excluding 5), or the number of cases with Grade 2 adverse skin reactions should not exceed 2 (for deodorant patch tests, the number of cases with Grade 1 adverse skin reactions should not exceed 10, and the number of cases with Grade 2 adverse skin reactions should not exceed 5), or there should be no more than one case of Grade 3 or higher adverse skin reaction, as consumers' demands for the gentleness and safety of cosmetics gradually increase, especially for special consumers such as those with sensitive skin, the above patch test evaluation criteria are insufficient to meet consumers' safety requirements and make it difficult to effectively assess the safety risks of cosmetics to target users.

[0004] Therefore, there is an urgent need to propose a classification method for human patch test results of cosmetics suitable for sensitive skin. This method involves collecting adverse reaction feedback from cosmetics that claim to be suitable for sensitive skin and have already been marketed, combining the patch test results to build a model, and selecting the optimal patch test classification standard. This will enable the safety assessment of cosmetics for people with sensitive skin and effectively avoid safety risks after cosmetics are launched on the market. Summary of the Invention

[0005] This invention proposes a method for classifying the results of human patch tests for cosmetics suitable for sensitive skin, in order to solve the technical problem that the current patch test evaluation standards are insufficient to meet consumers' safety requirements and make it difficult to effectively assess the safety risks of cosmetics to target users.

[0006] To achieve the above objectives, in a first aspect, the present invention proposes a method for classifying the results of cosmetic patch tests on sensitive skin, comprising:

[0007] Step 1: Collect review data of marketed cosmetics that claim to be suitable for sensitive skin, perform deduplication and noise reduction processing on the review data, and incorporate the deduplication and noise reduction processing review data into the review dataset corresponding to the cosmetic.

[0008] The comment dataset was analyzed using a trained RoBERTa-large language model to identify adverse reaction comments and calculate adverse reaction scores and adverse reaction rates.

[0009] Step 2: Perform a patch test on the cosmetic product described in Step 1, record the adverse reaction level and number of adverse reactions in the patch test results, and set the classification line;

[0010] Step 3: Extract the product characteristics of the cosmetic;

[0011] Step 4: Set the adverse reaction rate as the dependent variable, the product characteristics as the secondary independent variable, and the patch test results under different classification lines as the primary independent variable to establish a generalized linear model;

[0012] Step 5: Test the likelihood ratio chi-square and significance of each generalized linear model. If the generalized linear model is significant, compare the Bayesian information criterion of all significant generalized linear models laterally, select the model with the best fit, and take its corresponding fractal line as the best fractal for the cosmetic patch test.

[0013] Furthermore, in step 1, the cosmetic's market information and applicable skin type information can be screened and verified by combining it with information filed with the National Medical Products Administration. The cosmetic's review data comes from internal data legally obtained by our company.

[0014] Furthermore, in step 1, the deduplication of the comment data can be done manually or using a classic hash algorithm or a database deduplication method. The classic hash algorithm preprocesses each comment (including removing spaces, punctuation, and standardizing case) and then calculates its MD5, SHA-1, or CRC32 hash value. If two comments have the same hash value, they are considered duplicates, and only one is retained. If the comment data is stored in a database such as MySQL or MongoDB, the database's built-in deduplication functions can be used directly. For example, MySQL uses the DISTINCT keyword to filter unique comments; MongoDB uses the $group operation (grouping by comment field) in the aggregate pipeline for deduplication.

[0015] Furthermore, in step 1, the noise reduction processing of the comment data can be done manually, or it can be efficiently and automatically processed using text processing libraries built into the Python programming language (such as the re regular expression library, the jieba word segmentation library, the string string processing library, etc.).

[0016] Noise reduction using Python text processing libraries includes: using the `re` regular expression library and the `jieba` word segmentation library to identify URLs / contact information, removing purely advertising or marketing content. For example, comment data may contain information primarily composed of one or more keywords such as "link," "personal shopper," "flash sale," "add WeChat," "money-back guarantee," "add verified account," "WeChat," "private message," "buy one get one free," "clearance sale," "special offer," "promotion," "discount," "free shipping," "stockpiling," and "customer service." Length filtering is performed using the `MIN_LENGTH` and `MAX_LENGTH` functions, invalid character filtering using the `VALID_CHAR_PATTERN` function, and null value filtering using the `EMPTY_KEYWORDS` function. Meaningless content is removed, such as comments shorter than 5 characters, comments longer than 150 characters, comments consisting solely of symbols or emoticons, garbled characters, and repeated characters, and comments with semantically empty values. The CORE_TOPIC_KEYWORDS function is used for core keyword matching to remove irrelevant content. For example, by determining whether the core keywords cover "usage behavior" and / or "adverse reactions", irrelevant comments such as casual chat and logistics complaints are removed.

[0017] Furthermore, in step 1, the RoBERTa-large language model is trained using the following steps:

[0018] Step 1-1: Establish adverse reaction term sets for three categories: sensitization, irritation, and acne-causing. Based on these, sample the review dataset for each cosmetic product and perform manual annotation. The manual annotation is performed in parallel by multiple annotators.

[0019] The sampling process needs to ensure that the sample size of the comment data involved in the three categories of adverse reaction terms (sensitization, irritation, and acne-causing) is balanced.

[0020] Step 1-2: Calculate the Kappa coefficient of annotation consistency among multiple annotators. If the Kappa coefficient is not less than 0.8, the annotation is considered qualified. If the Kappa coefficient is greater than 0.8, repeat step 1-1 until the annotation is qualified. Sample the comment dataset with qualified annotations to form divergence samples.

[0021] Steps 1-3: Organize the comment data from the divergent samples into a .csv format text file; the .csv format text file consists of comment text and tags;

[0022] Specifically, the label can be set to 0, 1, 2, 3, where 0 is a non-adverse reaction category label, 1 is a sensitizing category label, 2 is an irritant category label, and 3 is an acne-causing category label;

[0023] Steps 1-4: Split the .csv format text file into a training set, a validation set, and a test set;

[0024] The training set is used for parameter learning of the RoBERTa-large language model, the validation set is used for parameter tuning during the training process of the RoBERTa-large language model, and the test set is used for the final performance evaluation of the RoBERTa-large language model.

[0025] Specifically, the .csv format text file is split into a training set, a validation set, and a test set in a 7:2:1 ratio;

[0026] Steps 1-5: Load the Tokenizer into the RoBERTa-large language model to be trained, and define text preprocessing functions based on the Tokenizer;

[0027] Specifically, the text preprocessing function defined based on the Tokenizer includes: extracting text tokens, adding [CLS] / [SEP] special symbols, converting to input_ids, generating attention_mask, and truncating / padding according to the set maximum sequence length (e.g., 64 tokens), to ensure that the output format of the text preprocessing function meets the input requirements of the RoBERTa-large model;

[0028] Specifically, the tokenizer corresponding to the RoBERTa-large model to be trained can be loaded using the Python programming language, for example, by combining the Hugging Face transformers library;

[0029] Steps 1-6: Load the .csv format file of the training set, convert it into a Dataset object, then apply the text preprocessing function in batches to the comment text processing, and generate new fields containing input_ids and attention_mask;

[0030] Specifically, the `load_dataset` function of the datasets library is used to load the .csv format file of the training set and convert it into a standard Dataset object from the datasets library, which facilitates batch processing of the comment text;

[0031] Call the map method of the Dataset object to apply the preprocessing functions defined in steps 1-5 in batches to the text fields of the dataset, such as the "comment text" field, and generate new fields containing input_ids and attention_mask;

[0032] By using methods such as set_format(type="torch"), the processed data, such as input_ids, attention_mask and label fields, are converted into PyTorch tensor format to obtain a training dataset that perfectly matches the input format required by the RoBERTa-large language model to be trained.

[0033] Steps 1-7: Load the RoBERTa-large language model to be trained and add a classification head; select an optimizer and set the initial learning rate and weight decay; use a linear learning rate scheduler and warm up; select accuracy and macro-average F1 score as evaluation metrics to perform model training and validation.

[0034] Specifically, the Adam W optimizer was selected, and the initial learning rate was set to 5e. -5 The weights decay to 0.01; a linear learning rate scheduler is used, with the first 10% of steps used for warm-up.

[0035] Steps 1-8: Write the training loop and the number of loop rounds involved, traverse each training round, and perform training and model parameter updates in sequence to obtain the pre-trained RoBERTa-large language model;

[0036] Steps 1-9: Validate and evaluate the performance of the pre-trained RoBERTa-large language model based on the validation set, and save the pre-trained RoBERTa-large language model with the best performance based on the validation set.

[0037] Steps 1-10: Load the best-performing pre-trained RoBERTa-large language model, test it on the test set, and check the model's recognition results for adverse reaction comments. If the recognition results reach the expected accuracy, then the best-performing pre-trained RoBERTa-large language model is used as the trained RoBERTa-large language model.

[0038] Furthermore, in step 1-1, the adverse reaction terms for the sensitization category include: sensitization, erythema, itching, burning sensation, rash, and edema;

[0039] Adverse reaction terms in the irritation category include: irritation, stinging, pain, and spiciness;

[0040] Adverse reactions related to acne include: acne, pimples, and breakouts.

[0041] Furthermore, in steps 1-8, if the number of iterations is too low, it is insufficient for the model to fully learn the data features. For example, traversing only once may only remember the patterns of some samples, i.e., "underfitting"; if the number of iterations is too high, the model may over-memorize details in the training set (including noisy data), resulting in a decline in performance on unseen validation / test sets, i.e., "overfitting".

[0042] Furthermore, in step 1, the adverse reaction score for a single adverse reaction review is calculated using the following formula:

[0043] ;

[0044] in, DR The adverse reaction score is the score for a single adverse reaction review. Base This is a base score based on adverse reaction characteristics. Degree This is the degree grading coefficient. To prevent zero constant, The number of items involved is based on the dimensions of the affected area and / or the medical scenario. For dimension weights, To correct the dimensionality coefficient; where, The value is 0.85.

[0045] Furthermore, the basic scores based on adverse reaction characteristics for adverse reaction entries in the sensitization category are as follows:

[0046] Sensitization: 1.0 point; Erythema: 1.0 point; Itching: 1.0 point; Burning sensation: 1.1 point; Rash: 1.2 point; Edema: 1.5 point.

[0047] The basic scores based on adverse reaction characteristics for adverse reaction entries in the stimulus category are as follows:

[0048] Irritation is rated 1.0, stinging is rated 1.0, spiciness is rated 1.0, and pain is rated 1.2.

[0049] The basic scores based on adverse reaction characteristics for adverse reaction entries in the acne-causing category are as follows:

[0050] Acne-causing factor is 1.0, blackheads are 1.0, and breakouts are 1.2.

[0051] Furthermore, the degree grading coefficient is determined based on the degree level and semantic features:

[0052] L1 represents mild, vague symptoms, with a corresponding severity grading coefficient of 0.8;

[0053] L2 represents the exact symptoms, with a corresponding severity grading coefficient of 1.0;

[0054] L3 indicates severe symptoms, with a severity grading coefficient of 1.2.

[0055] Furthermore, the dimensional weights and corrected dimensionality coefficients Involving the dimensions of body part area and / or medical treatment scenario:

[0056] If adverse reaction comments mention a large area, the corresponding dimension weight... and corrected dimensionality coefficients They are 0.4 and 0.8 respectively;

[0057] If adverse reaction comments mention medical treatment or diagnosis, their corresponding dimension weighting and corrected dimensionality coefficients They are 0.6 and 1.5 respectively;

[0058] If the comments do not mention the above content, the dimensionality coefficient will be adjusted. It is 0.

[0059] Furthermore, in step 1, the adverse reaction rate is calculated using the following formula:

[0060] Adverse reaction rate due to sensitization: ;

[0061] Adverse reaction rate to stimuli: ;

[0062] Acne-causing adverse reaction rate: ;

[0063] in, N The total number of reviews for the cosmetic product. n The total number of adverse reaction reviews for the aforementioned cosmetic product. For adverse reaction reviews categorized as sensitization, the adverse reaction score is... The adverse reaction score is used to evaluate the adverse reaction comments in the stimulus category. Adverse reaction score for acne-causing adverse reaction review.

[0064] Furthermore, in step 3, the product characteristics of the cosmetic include usage method, claimed efficacy, and basic ingredients;

[0065] The product features include usage methods such as rinse-off and leave-on types; the claimed effects include whitening and spot removal, sun protection, acne treatment, nourishing, repairing, anti-wrinkle, firming, soothing, oil control, exfoliation, moisturizing, suitable for sensitive skin, tear-free, and gentle and non-irritating; the basic ingredients in the product features refer to whether the cosmetic contains potentially risky ingredients in the full ingredient list filed with the National Medical Products Administration.

[0066] Furthermore, the risky ingredients include alcohol, vitamin A and its derivatives, acidic components, sodium lauryl sulfate / sodium laureth sulfate surfactants, silicone oil, polyethylene glycol-containing components, paraben preservatives, mineral oil, pigments, and fragrances.

[0067] Furthermore, in step 4, the generalized linear model is a gamma-log linkage model.

[0068] Compared with the prior art, the present invention has at least the following beneficial effects:

[0069] 1. The present invention provides a method for classifying the results of cosmetic patch tests on human skin suitable for sensitive skin, which can better match the classification and interpretation of cosmetic patch test results suitable for sensitive skin populations, and solve the limitations of patch test results in the application of cosmetic industry.

[0070] 2. The present invention provides a method for classifying the results of cosmetic patch tests on human skin suitable for sensitive skin. It uses adverse reaction reviews from e-commerce platforms, which are more suitable for people with sensitive skin, as the reference for patch result classification instead of conventional adverse reaction monitoring systems. This method has higher granularity and improves accuracy through reasonable classification and multi-dimensional correction coefficients, which helps patch tests provide safety references under higher safety requirements.

[0071] 3. The present invention provides a method for classifying the results of human patch tests for cosmetics suitable for sensitive skin. The data comes from e-commerce platforms, with a large amount of data and fast return speed, which is convenient for collection and rapid feedback, and can realize the dynamicization of patch test safety standards. Attached Figure Description

[0072] Figure 1 This is a schematic diagram illustrating the operation process of a method for classifying the results of a cosmetic patch test on sensitive skin, according to the present invention. Detailed Implementation

[0073] The various aspects of the present invention will be further described in detail below.

[0074] Unless otherwise defined or stated, all technical and scientific terms used herein have the same meaning as are familiar to a user skilled in the art. Furthermore, any methods and materials similar to or equivalent to those described herein may be applied to the methods of this invention.

[0075] Firstly, this invention proposes a method for classifying the results of human patch tests for cosmetics suitable for sensitive skin, in order to solve the technical problem that the current patch test evaluation standards are insufficient to meet consumers' safety requirements and make it difficult to effectively assess the safety risks of cosmetics to target users.

[0076] See appendix Figure 1 A method for classifying the results of cosmetic patch tests on human skin suitable for sensitive skin, comprising:

[0077] Step 1: Collect review data of marketed cosmetics that claim to be suitable for sensitive skin, perform deduplication and noise reduction processing on the review data, and incorporate the deduplication and noise reduction processing review data into the review dataset corresponding to the cosmetic.

[0078] The comment dataset was analyzed using a trained RoBERTa-large language model to identify adverse reaction comments and calculate adverse reaction scores and adverse reaction rates.

[0079] Step 2: Perform a patch test on the cosmetic product described in Step 1, record the adverse reaction level and number of adverse reactions in the patch test results, and set the classification line;

[0080] Step 3: Extract the product characteristics of the cosmetic;

[0081] Step 4: Set the adverse reaction rate as the dependent variable, the product characteristics as the secondary independent variable, and the patch test results under different classification lines as the primary independent variable to establish a generalized linear model;

[0082] Step 5: Examine the likelihood ratio chi-square and significance of each generalized linear model. If the generalized linear model is significant, then horizontally compare the Bayesian information criteria of all significant generalized linear models, select the model with the best goodness of fit, and use the corresponding segmentation line as the best segmentation for the cosmetics corresponding to the patch test.

[0083] Example 1 Adverse Reaction Comment Recognition

[0084] To clearly illustrate the process of analyzing the comment dataset using the trained RoBERTa-large language model, the process of identifying adverse reaction comments is demonstrated here with a few example comments from the comment dataset:

[0085] 1. Prepare a representative set of example comments: This set should cover all categories and edge cases involved in the model design objectives, specifically including: adverse reaction comments of clear sensitization categories, adverse reaction comments of clear irritation categories, adverse reaction comments of clear comedogenic categories, clear non-adverse reaction comments, vaguely expressed suspicious comments, and non-string type data, so as to ensure that the test samples can comprehensively cover the actual scenarios that the model may encounter. See Table 1;

[0086] Table 1. Representative Set of Example Comments

[0087]

[0088] 2. Preprocessing and feature transformation before using the trained RoBERTa-large language model: Load the Tokenizer tokenizer and perform tokenization on each valid example comment. For example, for the comment "After using this essence for 3 days, my cheeks started to turn red, got a lot of small rashes, and were extremely itchy. It should be an allergy", the Tokenizer tokenizer splits it into the token sequence: "用", "了", "这", "款", "精", "华", "液", "3", "天", ",", "脸", "颊", "开", "始", "发", "红", ",", "起", "了", "很", "多", "小", "疹", "子", ",", "特", "别", "痒", ",", "应", "该", "是", "过", "敏", "了".

[0089] The token sequence is mapped to the corresponding integer ID based on the pre-trained vocabulary learned by the trained RoBERTa-large language model, forming the input_ids sequence. At the same time, the input_ids sequence is truncated or padded according to the set maximum sequence length (e.g., 64 tokens) (if less than 64 bits, the ID corresponding to [PAD] is used for padding), and an attention_mask with the same length as the input_ids sequence is generated (the effective token position is 1, and the padding position is 0). For example, if the length of the adverse reaction comment of the aforementioned explicit sensitization category after word segmentation is 35, then the first 35 bits of the attention_mask are 1, and the last 29 bits are 0.

[0090] Finally, the processed input_ids sequence and attention_mask are converted into PyTorch tensor format and moved to GPU or CPU;

[0091] 3. Start the trained RoBERTa-large language model for inference and calculate the confidence score: Input the PyTorch tensor into the trained RoBERTa-large language model, perform forward propagation with gradient calculation disabled, and obtain the unnormalized raw score sequence output by the model. For example, for the aforementioned adverse reaction comments with clear sensitization categories, the trained RoBERTa-large language model outputs an unnormalized raw score sequence of [-2.1, 3.5, -1.8, -0.5], where the numbers in the sequence represent adverse reactions, sensitization categories, irritation categories, and acne-causing categories from left to right.

[0092] The unnormalized original score sequence is then normalized using the softmax function, transforming it into a probability distribution by calculating the probability percentage of each category. For example, for the aforementioned adverse reaction reviews with a clear sensitization category, the trained RoBERTa-large language model infers and calculates that the probability of its sensitization category is 0.9872, while the probabilities of other categories are all below 0.05. The category with the highest probability is selected as the predicted label for the review (e.g., "adverse reaction review with sensitization category"), and this highest probability value is recorded as the confidence level of the prediction (e.g., 0.9872).

[0093] 4. Filter the results and label the status based on the confidence threshold. Compare the predicted confidence of each comment with the preset threshold (e.g., 0.7). For example, for "After using this serum for 3 days, my cheeks started to turn red, and I developed many small rashes that were very itchy. It must be an allergic reaction," if the confidence is ≥0.7, the status is labeled as "Adverse reaction comment in the sensitization category, normal," indicating that the model has a high degree of certainty about the prediction result. If the confidence is <0.7, the status is labeled as "Non-adverse reaction, suspicious," requiring manual intervention for verification. For "Non-adverse reaction," the status is labeled as "Non-adverse reaction, normal." For invalid data (e.g., empty strings, non-strings), directly label the prediction type as "Invalid data" and the status as "Skip," thus distinguishing inference results with different levels of reliability. See Table 2.

[0094] Table 2. Model inference and confidence calculation results for a representative set of example comments.

[0095]

[0096] It should be noted that the effective volume of each cosmetic review should be greater than or equal to 500. Determining the effective volume threshold is a crucial step in cosmetic consumer review analysis. An appropriate threshold ensures the representativeness and reliability of the analysis results; typically, when analyzing issues such as allergies and adverse reactions, an effective volume greater than 500 reviews guarantees statistical significance.

[0097] It should be noted that in step 2, the patch test is a closed patch test, and the method is the closed patch test for human skin in the "Cosmetic Safety Technical Specifications (2015 Edition)". The closed patch test is a commonly used method for testing the safety of cosmetics. It assesses the sensitization or irritation of a substance by attaching a patch containing the test substance to the skin, fixing it in place, and observing the skin reaction. The main mechanism of the patch test is type IV hypersensitivity, and the allergens used are usually small molecule compounds, i.e., organic or inorganic substances with a relative molecular mass < 500,000.

[0098] It should be noted that in step 4, the generalized linear model is a gamma-log join model. The gamma-log join model is a type of generalized linear model (GLM) that consists of a stochastic component, a systematic component, and a join component. Its model formula is as follows:

[0099] Linear predictor:

[0100] in, For model parameters, These are explanatory variables.

[0101] Forecast Mean With linear predictors The relationship is: .

[0102] This join function ensures that the predicted mean is always non-negative, making it suitable for handling data where the response variable is non-negative and may exhibit a right-skewed distribution (e.g., the severity or frequency of adverse reactions to cosmetics). On the one hand, a linear predictor can introduce multiple explanatory variables to capture complex variable relationships.

[0103] It should be noted that the Omnibus test needs to be performed on all established gamma-log join models to calculate the likelihood ratio chi-square and significance. The likelihood ratio chi-square is mainly used to compare the goodness of fit of the models and test whether the model is significantly better than the null model (i.e., the model containing only the intercept term). If the p-value is less than the preset significance level (0.05), it indicates that the patch test results contribute statistically significant to the model. Finally, the gamma-log join models that passed the Omnibus test are compared horizontally with the Bayesian Information Criterion (BIC). BIC is an indicator that measures the goodness of fit and complexity of a model. The smaller the BIC value, the better the balance between goodness of fit and simplicity. Through BIC comparison, the model with the best goodness of fit and the simplest structure is selected as the basis for establishing the classification criteria for patch test results.

[0104] It should be noted that the usage methods in the product features include rinse-off and leave-on types; the claimed effects in the product features include whitening and spot removal, sun protection, acne treatment, nourishing, repairing, anti-wrinkle, firming, soothing, oil control, exfoliation, moisturizing, suitable for sensitive skin, tear-free, and gentle (non-irritating); the basic ingredients in the product features refer to whether the cosmetic contains potentially risky ingredients in the full ingredient list filed with the National Medical Products Administration.

[0105] It should be noted that the risky ingredients include alcohol, vitamin A and its derivatives, acidic components, sodium lauryl sulfate / sodium laureth sulfate surfactants, silicone oil, polyethylene glycol-containing components, paraben preservatives, mineral oil, pigments, and fragrances. The 2015 edition of the *Cosmetic Safety Technical Specifications* states that alcohol at high concentrations may irritate the skin, causing irritation, burning, and redness, especially in people with sensitive skin; Vitamin A and its derivatives may decompose under sunlight, producing phototoxic reactions, leading to redness, peeling, and in high concentrations, may cause skin irritation, dryness, or flaking; the 2015 edition of the *Cosmetic Safety Technical Specifications* limits the concentration of acidic ingredients, as they may excessively exfoliate the stratum corneum, damaging the skin barrier function and causing stinging, redness, or burning sensations, especially in sensitive skin; the 2015 edition of the *Cosmetic Safety Technical Specifications* limits the concentration of sodium lauryl sulfate / sodium laureth sulfate, as it may damage the skin barrier, leading to dryness, irritation, or sensitivity; and the EU Cosmetic Regulation (EC No.) Cosmetic ingredients (EC No. 1223 / 2009) restrict the use of certain silicone oils, as some silicone oils may form a clogging film on the skin surface, leading to clogged pores and causing acne or pimples; the Cosmetic Ingredient Review (CIR) states that polyethylene glycol may irritate damaged skin, enhance the skin permeability of other ingredients, and increase potential risks; the EU Cosmetic Regulations (EC No. 1223 / 2009) restrict the concentration of parabens, as parabens may cause skin irritation or allergic reactions; mineral oils may form a clogging film on the skin surface, leading to clogged pores and causing acne or pimples; fragrances and pigments are common allergens in cosmetics and may cause contact dermatitis or allergic reactions.

[0106] Example 2: Calculation of Adverse Reaction Rate

[0107] In this embodiment, one cosmetic product was selected, and mainstream e-commerce platforms (such as Taobao and JD.com) were collected. After deduplication and noise reduction, a total of 39,869 valid reviews were obtained. The trained RoBERTa-large language model was used to analyze the review data and identify adverse reaction reviews.

[0108] Step 1: Identify negative reaction comments in the review data and extract negative reaction features;

[0109] Step 2: Determine the severity of the adverse reaction for each individual adverse reaction review;

[0110] Step 3: Identify the scope and context of adverse reactions in a single adverse reaction review;

[0111] Step 4: Calculate the adverse reaction score for a single adverse reaction review:

[0112] ;

[0113] in, DR The adverse reaction score is the score for a single adverse reaction review. Base This is a base score based on adverse reaction characteristics. Degree This is the degree grading coefficient. To prevent zero constant, The number of items involved is based on the dimensions of the affected area and / or the medical scenario. For dimension weights, To correct the dimensionality coefficient; where, The value is 0.85;

[0114] Step 5: Calculate the adverse reaction rate of the cosmetic product. The adverse reaction rate is calculated using the following formula:

[0115] Adverse reaction rate due to sensitization: ;

[0116] Adverse reaction rate to stimuli: ;

[0117] Acne-causing adverse reaction rate: ;

[0118] in, N The total number of reviews for the cosmetic product. n The total number of adverse reaction reviews for the aforementioned cosmetic product. For adverse reaction reviews categorized as sensitization, the adverse reaction score is... The adverse reaction score is used to evaluate the adverse reaction comments in the stimulus category. Adverse reaction score for acne-causing adverse reaction review.

[0119] In comparison, the formula for calculating adverse reaction rate in the prior art is as follows:

[0120] Adverse reaction rate of sensitization ;

[0121] Adverse reaction rate of stimulation ;

[0122] Acne-causing adverse reaction rate ;

[0123] in, N The total number of reviews for the cosmetic product. The total number of adverse reaction comments for cosmetics sensitizing categories. The total number of adverse reaction reviews for cosmetics categorized as irritant. This represents the total number of adverse reaction reviews for cosmetic products categorized as acne-causing.

[0124] Analysis results show that 72 out of 39,869 valid reviews were marked as adverse reaction reviews. These were verified by manual review and were not incorrectly identified. The results obtained from the adverse reaction rate calculation method of this invention (Method 1) and the prior art's adverse reaction rate calculation method (Method 2) are shown in Tables 3-4. It can be seen that compared to the adverse reaction rate calculation results in Method 2, Method 1 shows a higher sensitization adverse reaction rate and a lower irritation and acne-causing adverse reaction rate. This indicates that the cosmetic product has a more severe degree of sensitization-related adverse reactions, while the irritation and acne-causing adverse reactions are milder. Therefore, this demonstrates that Method 1 has a reasonable calculation logic and appropriate numerical settings, effectively distinguishing between adverse reaction reviews of different degrees without affecting the overall trend.

[0125] Table 3. Results obtained from the adverse reaction rate calculation method of the present invention (Method 1) and the adverse reaction score calculation method of the prior art (Method 2).

[0126]

[0127] Table 4. Results obtained from the adverse reaction rate calculation method of the present invention (Method 1) and the adverse reaction rate calculation method of the prior art (Method 2).

[0128]

[0129] Example 3: Construction of Patch Result Typing

[0130] Step 1: Cosmetic Selection

[0131] This embodiment selected 64 cosmetic products and collected review data from mainstream e-commerce platforms. After deduplication and noise reduction, a total of 1,789,068 valid reviews were obtained. The review data was analyzed using a trained RoBERTa-large language model, and 8,793 reviews were marked as mentioning adverse reactions. The sensitization rate, irritation rate, and acne-causing rate of each cosmetic product were calculated.

[0132] Step 2: Extract product features

[0133] Categorized by usage method: 64 cosmetics, including 11 rinse-off cosmetics and 53 leave-on cosmetics; Categorized by efficacy claims: 53 cosmetics claim "moisturizing", 60 cosmetics claim "soothing", 38 cosmetics claim "repairing", 4 cosmetics claim "anti-wrinkle", 4 cosmetics claim "acne treatment", 7 cosmetics claim "spot removal and whitening", 7 cosmetics claim "oil control", and 5 cosmetics claim "firming". All cosmetics do not contain potentially risky ingredients and claim to be suitable for sensitive skin.

[0134] Step 3: Setting the parting line

[0135] In human skin patch testing, 1, 2, and 3 suspected reactions were set as the dividing lines. If the number of reactions in the cosmetic patch test reached or exceeded the dividing line, it was judged as a failure (recorded as 1); if it was below the dividing line, it was judged as a pass (recorded as 0). The results of human skin patch testing of 64 cosmetics showed that 42 cosmetics had 0 suspected reactions, 8 cosmetics had 1 suspected reaction, 7 cosmetics had 2 suspected reactions, and 7 cosmetics had 3 or more suspected reactions.

[0136] Step 4: Model Building

[0137] ① Using the sensitization adverse reaction rate as the dependent variable, the method of use and efficacy claims as co-independent variables, and the patch test results under different fractal lines as the principal independent variables, three independent generalized linear models (GLMs) were established. The model type was a gamma-log linkage model: Model 1: a generalized linear model based on 1 case of suspected reaction at the fractal line; Model 2: a generalized linear model based on 2 cases of suspected reaction at the fractal line; Model 3: a generalized linear model based on 3 cases of suspected reaction at the fractal line.

[0138] ② Using the adverse reaction rate as the dependent variable, the method of use and efficacy claims as co-independent variables, and the patch test results under different fractal lines as the principal independent variables, three independent generalized linear models (GLMs) were established. The model type was a gamma-log connection model: Model 1: a generalized linear model based on 1 case of suspected reaction at the fractal line; Model 2: a generalized linear model based on 2 cases of suspected reaction at the fractal line; Model 3: a generalized linear model based on 3 cases of suspected reaction at the fractal line.

[0139] ③ Using the acne-causing adverse reaction rate as the dependent variable, the method of use and efficacy claims as co-independent variables, and the patch test results under different fractal lines as the principal independent variables, three independent generalized linear models (GLMs) were established. The model type was a gamma-log linkage model: Model 1: a generalized linear model based on 1 case of suspected reaction at the fractal line; Model 2: a generalized linear model based on 2 cases of suspected reaction at the fractal line; Model 3: a generalized linear model based on 3 cases of suspected reaction at the fractal line.

[0140] Step 5: Compare model goodness

[0141] Omnibus tests were performed on all established gamma-log join models, calculating the likelihood ratio, chi-square, and significance to test whether the model was significantly better than the null model. If the p-value was less than the preset significance level (0.05), it indicated that the patch test results contributed statistically to the model. The Bayesian Information Criterion (BIC) was compared across all gamma-log join models that passed the Omnibus test. A smaller BIC value indicated a better model fit, which served as the basis for establishing the classification criteria for patch test results.

[0142] The effects of Bayesian Information Criterion (BIC) and patch test results on the adverse reaction rates of different models were compared, and the results are shown in Table 5.

[0143] Table 5. Comparison of model BIC values ​​and the significance of patch test results

[0144]

[0145] The results show that, for adverse sensitization reactions, Model 2 has a significant impact on the patch test results, while Models 1 and 3 do not have a significant impact, indicating that Model 2 has the best fit. In other words, when two or more suspected reactions are found in patch tests of cosmetics with the same product characteristics, the risk of adverse sensitization reactions after the cosmetic is marketed increases significantly.

[0146] Regarding adverse irritation reactions, patch test results had a significant impact on the rate of adverse irritation reactions in both Model 1 and Model 2, but no significant impact in Model 3. The BIC value of Model 2 was lower than that of Model 1 and Model 3, indicating that Model 2 had the best fit. This means that when two or more suspected reactions are found in patch test results for cosmetics with the same product characteristics, the risk of adverse irritation reactions after the cosmetic is marketed is significantly increased.

[0147] Regarding acne-causing adverse reactions, the patch test results in all three models did not have a significant impact on the acne-causing adverse reaction rate. Therefore, under the current product characteristics, there is no need to establish a classification line based on user feedback on acne-causing adverse reactions.

[0148] Based on this invention, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspects set forth herein can be used to implement the device and / or practice the method. Furthermore, this device and / or practice the method can be implemented using other structures and / or functionalities besides one or more of the aspects set forth herein.

[0149] It should be noted that the above embodiments can be freely combined as needed. The above description is only a preferred embodiment of the present invention. It should be pointed out that for users of ordinary skills in this field, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0150] All documents mentioned in this invention are incorporated herein by reference as if each document were individually incorporated by reference. Furthermore, it should be understood that after reading the foregoing description of this invention, users skilled in the art can make various alterations or modifications to this invention, and these equivalent forms also fall within the scope defined by the appended claims.

Claims

1. A method for classifying the results of cosmetic patch tests on sensitive skin, characterized in that, Includes the following steps: Step 1: Collect review data of marketed cosmetics that claim to be suitable for sensitive skin, perform deduplication and noise reduction processing on the review data, and incorporate the deduplication and noise reduction processing review data into the review dataset corresponding to the cosmetic. The comment dataset was analyzed using a trained RoBERTa-large language model to identify adverse reaction comments and calculate adverse reaction scores and adverse reaction rates. Step 2: Perform a patch test on the cosmetic product described in Step 1, record the adverse reaction level and number of adverse reactions in the patch test results, and set the classification line; Step 3: Extract the product characteristics of the cosmetic; Step 4: Set the adverse reaction rate as the dependent variable, the product characteristics as the secondary independent variable, and the patch test results under different classification lines as the primary independent variable to establish a generalized linear model; Step 5: Test the likelihood ratio chi-square and significance of each generalized linear model. If the generalized linear model is significant, compare the Bayesian information criterion of all significant generalized linear models laterally, select the model with the best fit, and use its corresponding fractal line as the best classification for the cosmetic patch test; where, In step 1, the adverse reaction score for a single adverse reaction review is calculated using the following formula: ； in, DR The adverse reaction score is the score for a single adverse reaction review. Base This is a base score based on adverse reaction characteristics. Degree This is the degree grading coefficient. To prevent zero constant, The number of items involved is based on the dimensions of the affected area and / or the medical scenario. For dimension weights, To correct the dimensionality coefficient; where, The value is 0.85; In step 1, the adverse reaction rate is calculated using the following formula: Adverse reaction rate due to sensitization: ; Adverse reaction rate to stimuli: ; Acne-causing adverse reaction rate: ; in, N The total number of reviews for the cosmetic product. n This represents the total number of adverse reaction reviews for the aforementioned cosmetic product. For adverse reaction reviews categorized as sensitization, the adverse reaction score is... The adverse reaction score is used to evaluate the adverse reaction comments in the stimulus category. Adverse reaction score for acne-causing adverse reaction review.

2. The method for classifying the results of cosmetic patch tests for sensitive skin as described in claim 1, characterized in that, In step 1, the RoBERTa-large language model is trained using the following steps: Step 1-1: Establish adverse reaction term sets for three categories: sensitization, irritation, and acne-causing. Based on these, sample the review dataset for each cosmetic product and perform manual annotation. The manual annotation is performed in parallel by multiple annotators. Step 1-2: Calculate the Kappa coefficient of annotation consistency among multiple annotators. If the Kappa coefficient is not less than 0.8, the annotation is considered qualified. If the Kappa coefficient is greater than 0.8, repeat step 1-1 until the annotation is qualified. Sample the comment dataset with qualified annotations to form divergence samples. Steps 1-3: Organize the comment data from the divergent samples into a .csv format text file; The .csv format text file consists of comment text and tags; Steps 1-4: Split the .csv format text file into a training set, a validation set, and a test set; Steps 1-5: Load the Tokenizer into the RoBERTa-large language model to be trained, and define text preprocessing functions based on the Tokenizer; Steps 1-6: Load the .csv format file of the training set, convert it into a Dataset object, then apply the text preprocessing function in batches to the comment text processing, and generate new fields containing input_ids and attention_mask; Steps 1-7: Load the RoBERTa-large language model to be trained and add a classification head; select an optimizer and set the initial learning rate and weight decay; Use a linear learning rate scheduler and warm up; select accuracy and macro-average F1 score as evaluation metrics to perform model training and validation; Steps 1-8: Write the training loop and the number of loop rounds involved, traverse each training round, and perform training and model parameter updates in sequence to obtain the pre-trained RoBERTa-large language model; Steps 1-9: Validate and evaluate the performance of the pre-trained RoBERTa-large language model based on the validation set, and save the pre-trained RoBERTa-large language model with the best performance based on the validation set. Steps 1-10: Load the best-performing pre-trained RoBERTa-large language model, test it on the test set, and check the model's recognition results for adverse reaction comments. If the recognition results reach the expected accuracy, then the best-performing pre-trained RoBERTa-large language model is used as the trained RoBERTa-large language model.

3. The method for classifying the results of cosmetic patch tests for sensitive skin as described in claim 2, characterized in that, In step 1-1, the adverse reaction terms for the sensitization category include: sensitization, erythema, itching, burning sensation, rash, and edema; Adverse reaction terms in the irritation category include: irritation, stinging, pain, and spiciness; Adverse reactions related to acne include: acne, pimples, and breakouts.

4. The method for classifying the results of cosmetic patch tests for sensitive skin as described in claim 3, characterized in that, The basic scores based on adverse reaction characteristics for adverse reaction entries in the sensitization category are as follows: Sensitization: 1.0 point; Erythema: 1.0 point; Itching: 1.0 point; Burning sensation: 1.1 point; Rash: 1.2 point; Edema: 1.5 point. The basic scores based on adverse reaction characteristics for adverse reaction entries in the stimulus category are as follows: Irritation is rated 1.0, stinging is rated 1.0, spiciness is rated 1.0, and pain is rated 1.

2. The basic scores based on adverse reaction characteristics for adverse reaction entries in the acne-causing category are as follows: Acne-causing factor is 1.0, blackheads are 1.0, and breakouts are 1.

2.

5. The method for classifying the results of cosmetic patch tests for sensitive skin as described in claim 4, characterized in that, The degree grading coefficient is determined based on degree level and semantic features: L1 represents mild, vague symptoms, with a corresponding severity grading coefficient of 0.8; L2 represents the exact symptoms, with a corresponding severity grading coefficient of 1.0; L3 indicates severe symptoms, with a severity grading coefficient of 1.

2.

6. The method for classifying the results of cosmetic patch tests for sensitive skin as described in claim 5, characterized in that, The dimensional weights and corrected dimensionality coefficients Involving the dimensions of body part area and / or medical treatment scenario: If adverse reaction comments mention a large area, the corresponding dimension weight... and corrected dimensionality coefficients They are 0.4 and 0.8 respectively; If adverse reaction comments mention medical treatment or diagnosis, their corresponding dimension weighting and corrected dimensionality coefficients They are 0.6 and 1.5 respectively; If the comment mentions other content, the dimensionality coefficient will be adjusted. It is 0.

7. The method for classifying the results of cosmetic patch tests for sensitive skin as described in claim 6, characterized in that, In step 3, the product characteristics of the cosmetic include the method of use, claimed efficacy, and basic ingredients.

8. The method for classifying the results of cosmetic patch tests for sensitive skin as described in claim 7, characterized in that, In step 4, the generalized linear model is a gamma-log linkage model.

Images