A method and system for interpretable assessment of scalp health status based on multi-modal data

Abstract

This invention discloses an interpretable assessment method and system for scalp health based on multimodal data, belonging to the fields of intelligent health monitoring and artificial intelligence technology. The method includes: acquiring multimodal data such as scalp image features, hair loss counts, hair resilience, washing and conditioning behaviors, physiological indicators, and product ingredients; mapping these data to a unified latent space using a feature encoder to establish a personal baseline feature vector; calculating the deviation between the real-time feature vector and the baseline vector; using an interpretability analysis model to identify key influencing factors and output a contribution ranking; and generating personalized intervention suggestions. The entire personal analysis process is completed locally, with the cloud only receiving de-identified structured features for group insight aggregation. This invention achieves interpretable, reliable, and privacy-secure scalp health assessment, balancing personal services with industry data value.

Description

Technical Field

[0001] This invention relates to the fields of intelligent health monitoring and artificial intelligence technology, specifically to an interpretable assessment method and system for scalp health status based on multimodal data. Background Technology

[0002] Scalp health is influenced by a combination of factors, including the condition of hair follicles, sebum secretion, and inflammation, as well as the user's hair care habits, sleep quality, stress levels, and the ingredients in the shampoos and conditioners used. Data from a single dimension cannot fully reflect changes in scalp health.

[0003] In existing technologies, scalp health monitoring is mostly limited to displaying a single indicator or simple threshold alarms. For example, it calculates hair follicle density through image analysis or provides the daily amount of hair loss through hair loss statistics. These solutions have three main shortcomings: First, the data dimension is singular, and it cannot handle the complex correlations between multimodal heterogeneous data; second, it lacks personalized baselines, usually using group mean or fixed thresholds for judgment, ignoring individual differences; third, the analysis results are uninterpretable, and users can only see that "the condition has worsened," but cannot know whether it is caused by insufficient sleep, improper washing and care habits, or unsuitable products.

[0004] Furthermore, the application of existing interpretability analysis technology in the field of health monitoring is mostly limited to academic research or reports for professionals. There is no technical solution that can transform interpretability analysis results into health advice that ordinary users can understand and act on, and achieve commercial value output at the group level while protecting personal privacy.

[0005] Therefore, there is an urgent need for a technical solution that can integrate and analyze multimodal scalp health data, establish personalized baselines, calculate deviations, output interpretable attribution results and personalized recommendations, and simultaneously output privacy protection and commercial value through localization processing and desensitized group insights. Summary of the Invention

[0006] The purpose of this invention is to provide an interpretable assessment method and system for scalp health status based on multimodal data, in order to solve the problems of single data dimension, lack of personalized baseline, uninterpretable attribution, and difficulty in balancing privacy and commercial value in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: An interpretable assessment method for scalp health status based on multimodal data includes the following steps: S1: Obtain multimodal scalp health data of the user within a set period. The multimodal scalp health data includes at least scalp image feature data, hair loss count data, hair toughness data, washing and care behavior feature data, physiological indicator data, and washing and care product ingredient feature data. S2: Input the multimodal scalp health data into a pre-constructed feature encoder, and use the feature encoder to map the data of different modalities to the same latent space, and fuse them to generate a baseline feature vector representing the user's personal scalp health status. S3: In subsequent monitoring cycles, generate real-time feature vectors in the same manner as S2, and calculate the deviation between the real-time feature vectors and the reference feature vectors; S4: Using at least one interpretable analysis model, perform feature contribution analysis on the deviation, identify the key influencing factors that cause the change in the deviation, and output the contribution ranking of each key influencing factor; S5: Based on the ranking of the key influencing factors and their contributions, automatically generate targeted scalp health intervention suggestions and output them to the user.

[0008] Furthermore, the methods for acquiring the multimodal scalp health data mentioned in S1 include: acquiring scalp image feature data and hair loss count data through a smart combing terminal; acquiring washing and care behavior feature data and physiological indicator data through a wearable terminal; and acquiring washing and care product ingredient feature data after user authorization.

[0009] Furthermore, the washing and care behavior feature data includes a quantitative evaluation value for rubbing coverage. The quantitative evaluation value for rubbing coverage is generated by mapping the inertial measurement unit signal of the wearable terminal to the head coordinate system, calculating the spatial distribution of signal energy in each part of the head within the rubbing action segment, and outputting the quantitative evaluation value for coverage of each part based on the spatial distribution.

[0010] Furthermore, the hair toughness data is obtained by: collecting the voltage waveform signal output by the flexible strain sensor on the combing terminal, performing pulse detection on the waveform signal, extracting the peak amplitude and rising slope of each pulse, and statistically analyzing the pulse characteristics during one combing process to generate a hair toughness assessment value.

[0011] Furthermore, the feature encoder described in S2 is a multimodal Transformer encoder or an autoencoder, used to map scalp image features, mechanical features, behavioral features, physiological features and component features to a latent space of a unified dimension.

[0012] Furthermore, the deviation calculation method described in S3 includes any one or more combinations of cosine similarity, Euclidean distance, and Mahalanobis distance.

[0013] Furthermore, the interpretability analysis model described in S4 includes any one or more combinations of SHAP, LIME, Integrated Gradients, and attention weight analysis. This interpretability analysis model is used to quantify the marginal contribution of each modal input feature to the deviation and output a ranking of contributions.

[0014] Furthermore, it also includes a dynamic threshold warning step: automatically setting a dynamic threshold for deviation based on the fluctuation range of the user's baseline feature vector within the initial period; when the deviation between the real-time feature vector and the baseline feature vector exceeds the dynamic threshold, issuing a warning to the user.

[0015] Furthermore, it also includes multiple rounds of in-depth explanation and interaction steps: when users ask follow-up questions about the output intervention suggestions, based on the identified key influencing factors and their contribution ranking, targeted in-depth explanation content is generated and fed back to the user in natural language.

[0016] Furthermore, all steps from S1 to S5 are completed locally on the user's terminal. The cloud server does not access the raw multimodal scalp health data, but only receives structured feature data that has undergone anonymization.

[0017] Furthermore, it also includes the steps of generating and outputting de-identified group insights: after obtaining user authorization, the user terminal uploads the de-identified structured health feature data to the cloud server; the cloud server aggregates the de-identified feature data according to at least one dimension of population, product, region, and time to generate a group insight report; the group insight report only contains aggregated statistical conclusions and does not contain any identifiable or inferable data of any individual user.

[0018] Furthermore, it also includes multi-level data comparison and trust presentation steps: result-level indicators use long-term comparison windows, process-level indicators use short-term comparison windows, behavior-level indicators are used to show improvements in user behavior, and stability-level indicators are used to output "stable and without anomalies" interpretation results when the data is stable over a long period of time.

[0019] This invention also provides an interpretable assessment system for scalp health status based on multimodal data, comprising: a data acquisition module, a feature encoding module, a deviation calculation module, an interpretable attribution module, and a suggestion generation module. The feature encoding module is used to fuse multimodal data to generate a baseline feature vector and a real-time feature vector; the deviation calculation module is used to calculate the deviation between the real-time feature vector and the baseline feature vector; the interpretable attribution module is used to identify key influencing factors using an interpretability analysis model and output a contribution ranking; the suggestion generation module is used to generate personalized scalp health intervention suggestions based on the key influencing factors. All modules of the system are deployed locally on the user terminal, and the cloud only receives anonymized structured feature data. Beneficial effects

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. Multidimensional data fusion: For the first time, six types of heterogeneous data, including scalp image features, hair loss count, hair toughness, washing and care behavior, physiological indicators and product ingredients, are uniformly mapped to the latent space through a feature encoder to form a true multimodal health profile, which solves the problem of single data dimension in existing technologies; 2. Personalized baseline: By establishing an individual baseline feature vector, comparisons are made "with oneself" rather than "with the group," avoiding misjudgments caused by individual differences; 3. Explainable Attribution: By outputting contribution rankings through an interpretability analysis model, users not only know "what happened," but also "why it happened," significantly improving trust. 4. Localized AI protects privacy: The entire process from S1 to S5 is completed locally on the user's terminal, and the raw data does not leave the device, thus eliminating the risk of privacy leakage from the architecture. 5. The commercial value of anonymized group insights: By "selling insights, not data" in a way that protects user privacy, we provide personal care product companies with group-level insight reports such as product efficacy evaluation and user segmentation matching, thereby enabling the compliant monetization of data assets; 6. Multi-layered trust presentation: By comparing the results layer, process layer, behavior layer and stability layer, the problem of "short-term data not changing leading to user distrust" is solved, allowing users to continuously see the value of their behavior. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention.

[0022] Figure 2 This is an overall flowchart of the method of the present invention.

[0023] Figure 3 Flowchart for multimodal feature encoding and baseline establishment.

[0024] Figure 4 This is a flowchart for interpretable attribution analysis.

[0025] Figure 5 Output a logic diagram for desensitized groups.

[0026] Figure 6 This diagram illustrates the comparison and trust between multiple layers of data.

[0027] Figure 7 A logical diagram for representing the boundary of abnormal scalp trends.

[0028] Figure 8 This is a flowchart for determining low-confidence attribution and human-machine collaboration.

[0029] The meanings of the numbers in the attached figures are as follows: Figure 1 In China: 100-Smart combing terminal, 200-Smart ring, 300-Smart bracelet, 400-Smart pendant, 500-Shampoo ingredient data, 600-Cloud server, 700-Output layer; Figure 3 In the middle: 800 - individual baseline feature vectors; Figure 7 In the middle: 810 - recognition layer, 820 - analysis layer, 830 - expression layer. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. However, it should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0031] The "multimodal scalp health data" described in this invention refers to heterogeneous data related to scalp and hair health collected from different data sources, including but not limited to: scalp image features, hair loss counts, hair resilience, washing and conditioning behavior characteristics, physiological indicators, and washing and conditioning product ingredient characteristics. The specific data types and combinations thereof do not constitute a limitation on the scope of protection of this invention.

[0032] In this invention, "user terminal" refers to any personal computing device with data processing and local computing capabilities, including but not limited to smartphones, tablets, and personal computers. The user terminal is responsible for completing all personal-level analysis and calculations locally. Example 1: System Composition

[0033] This embodiment provides an interpretable assessment system for scalp health status based on multimodal data, deployed locally on a user terminal. The system includes: a data acquisition module, a feature encoding module, a deviation calculation module, an interpretable attribution module, and a suggestion generation module.

[0034] The data acquisition module receives data from the smart styling terminal and the wearable terminal, and preprocesses it locally on the user terminal. The smart styling terminal transmits scalp condition feature maps, hair loss counts, and hair resilience assessment values ​​to the user terminal via Bluetooth; the wearable terminal transmits shampooing behavior feature data and physiological indicator data to the user terminal via Bluetooth; and the shampoo ingredient feature data, after user authorization, is obtained by the user terminal from the e-commerce platform or local ingredient database via the application programming interface. Example 2: Multimodal Feature Encoding and Personal Baseline Establishment

[0035] The first week of user system use serves as a personal baseline establishment period. The feature encoding module receives all preprocessed multimodal data from this first week.

[0036] The scalp condition feature map is encoded into a fixed-length visual feature vector through a convolutional neural network; hair loss count is scalar data; hair resilience assessment value is scalar data; washing and conditioning behavior feature parameters (rubbing coverage, rubbing duration, rinsing duration) are structured vectors; physiological indicators (deep sleep duration, light sleep duration, heart rate variability) are structured vectors; and shampoo ingredient features are ingredient vectors.

[0037] The feature encoding module employs a multimodal Transformer encoder to map the aforementioned heterogeneous data to a unified-dimensional latent space, fusing them to generate a fixed-dimensional personal baseline feature vector. This baseline feature vector represents the user's individual scalp health status.

[0038] For example, the feature encoding module maps image feature vectors, scalar data, behavioral vectors, physiological vectors, and component vectors to the same dimension through different linear projections, and then performs cross-modal fusion through a self-attention mechanism, finally outputting a fixed-dimensional baseline feature vector.

[0039] After entering the continuous monitoring period, the system generates real-time feature vectors daily or every few days in the same manner as in Example 2. Example 3: Deviation Calculation and Dynamic Threshold Early Warning

[0040] The deviation calculation module calculates the cosine similarity between the real-time feature vector and the individual baseline feature vector. The higher the similarity, the closer the current state is to the individual's normal state; the lower the similarity, the greater the deviation.

[0041] The system automatically sets a dynamic threshold based on the fluctuation range of the baseline feature vector within the initial period. For example, if the cosine similarity between the daily real-time feature vector and the baseline feature vector fluctuates between 0.92 and 0.97 within the initial period, the dynamic threshold can be set to 0.90. When the similarity falls below 0.90 on a certain day, the system issues a warning to the user, indicating a significant deviation in scalp health. Example 4: Explainable Attribution Analysis

[0042] When the system detects that the deviation exceeds the dynamic threshold, the attribution module can initiate analysis.

[0043] The interpretable attribution module utilizes SHAP as an example implementation of the interpretability analysis model. It trains a lightweight gradient boosting tree model using multimodal input features as the feature set and deviation as the target variable. The SHAP library is then used to calculate the marginal contribution of each input feature to the deviation prediction result, and the contribution ranking is output.

[0044] For example, SHAP analysis results might show that "reduced deep sleep duration this week" contributes 0.6, "recently used shampoo contains SLS" contributes 0.3, and "decreased shampooing coverage" contributes 0.1. The system then determines and ranks the key influencing factors based on this information.

[0045] It is important to emphasize that SHAP is only one optional implementation of the interpretability analysis model. This invention protects the complete methodological steps of "performing feature contribution analysis on deviation and outputting contribution ranking", not the specific interpretability tool itself. Example 5: Generation of Personalized Intervention Recommendations

[0046] The suggestion generation module automatically generates personalized scalp health intervention suggestions based on the ranking of key influencing factors and their contributions output by the explainable attribution module.

[0047] The suggestion generation module includes a pre-set library of intervention suggestion templates, with each template corresponding to a category of key influencing factors. For example, the suggestion template for "insufficient deep sleep" is: "It is recommended to go to bed 30 minutes earlier, aiming for at least 6 hours of deep sleep." The suggestion template for "unsuitable shampoo ingredients" is: "It is recommended to switch to an amino acid surfactant shampoo to reduce SLS irritation."

[0048] The system sorts factors by contribution, selects the top 1 to 3 key factors, calls the corresponding templates, and generates combined intervention suggestions. These suggestions are then output to the user's terminal interface in natural language. Example 6: Multi-round Deep Interpretation Interaction

[0049] After receiving the intervention suggestion, users can ask follow-up questions through their user terminals. For example, a user might ask, "Why has hair loss decreased?"

[0050] The system identifies the intent to ask follow-up questions, retrieves stored data on key influencing factors and their contribution ranking, and generates targeted, in-depth explanations. For example: "Based on data from the last four weeks, your reduced hair loss is mainly related to two factors. First, your deep sleep time increased by 15%, which greatly helps scalp repair. Second, the ingredients in your recently switched shampoo are more suitable for your scalp type. If you continue with these two points, the hair loss trend will improve further."

[0051] The content is explained in detail and fed back to the user in natural language, and can be further combined with a voice assistant for broadcasting. Example 7: Localized AI Deployment

[0052] All personal-level analysis steps in this system—S1 Data Acquisition and Preprocessing, S2 Feature Encoding and Baseline Establishment, S3 Deviation Calculation, S4 Explainable Attribution, and S5 Recommendation Generation—are deployed locally on the user's terminal.

[0053] The user terminal performs the above analysis using local computing resources, without uploading the original multimodal scalp health data to the cloud. The original scalp image data is discarded immediately after semantic segmentation is completed locally on the intelligent scalp management terminal; only the anonymized feature maps are uploaded to the user terminal. The analysis process on the user terminal does not result in any outflow of raw data.

[0054] The cloud server only receives anonymized, structured health profile data for aggregate analysis at the group level. The cloud does not handle, store, or process any individual user's raw images, hair loss details, or behavioral sequence data.

[0055] This "device-edge-cloud" three-tier architecture ensures that personal privacy is protected at the architectural level, rather than relying on users' trust in the platform. Example 8: Generation and Output of Insights into Desensitized Groups

[0056] After obtaining explicit authorization from the user, the user terminal uploads the anonymized structured health feature data to the cloud server. The anonymized structured data only includes: scalp condition category statistics, hair loss trend statistics, hair resilience score, washing and conditioning behavior statistics, physiological indicator statistics, and shampoo ingredient feature vectors. It does not contain any original images, personal identification information, or fine-grained behavioral sequences that could be used to deduce personal identity.

[0057] A group insight analysis module is deployed on a cloud server. This module aggregates and analyzes anonymized feature data according to dimensions such as audience, product, region, and time, and generates group insight reports.

[0058] For example, for a certain oil-control shampoo, the cloud-based aggregated analysis output the following group insights: "Among users with oily scalps, the product achieved an 8-week improvement rate of 62% in scalp oil, and an improvement rate of 18% among users with dry scalps; the improvement rate was the highest among female users aged 25-35, reaching 68%."

[0059] The group insight report contains only aggregated statistical conclusions and does not include any identifiable or inferential data about individual users. Brands should obtain these insights through API interfaces or periodic reports for product development and user education, and must not use them for reverse identification or targeted user harassment. Example 9: Multi-level Data Comparison and Trust Presentation

[0060] To alleviate the problem of "user distrust due to lack of short-term data changes", this system adopts a multi-level data comparison logic.

[0061] The results layer metrics (hair volume, hair follicle density) use a long-term comparison window, presenting trends in units of 8 weeks or longer by default. The system clearly states when a user first uses it: "Hair is a slow variable; changes in hair volume usually take more than 8 weeks to show an initial trend."

[0062] Process-level indicators (hair loss, scalp oil distribution, hair resilience, and coverage during rubbing) use short-term comparison windows, presenting changes on a daily or weekly basis. Users can see the fluctuations in these indicators in a short period and receive immediate feedback.

[0063] Behavioral metrics (frequency of combing, quality of hair care, and improvement in deep sleep duration) record changes in users' lifestyles in real time. Even if a user's hair volume has not changed, the system will display behavioral improvements such as "You combed your hair 3 more times this week than last week, and the coverage of rubbing the right side of your scalp has increased by 15%".

[0064] The stability layer indicator is used to interpret data that has been stable over a long period. When neither the outcome indicator nor the process indicator shows significant anomalies, the system outputs the conclusion "scalp condition is stable, no abnormalities, and within a good range," preventing users from misinterpreting a "flat line" as "no change."

[0065] Through the above multi-level comparisons, users can feel the value of the system even when their hair volume has not changed, thereby establishing long-term trust. Example 10: Privacy and Security

[0066] This system strictly adheres to the following privacy protection principles: 1. Original images never leave the device: The original scalp images are discarded immediately after semantic segmentation is completed locally on the intelligent scalp terminal, and never leave the device; 2. Fully localized personal analytics: Multimodal fusion, baseline establishment, deviation calculation, interpretable attribution, and recommendation generation are all completed locally on the user's terminal; 3. Only anonymized data is received in the cloud: The cloud server only receives anonymized structured feature data and does not access any original personal data; 4. Insights output cannot be reversed: The group insight reports obtained by B-end brands only contain aggregated statistical conclusions and do not contain any identifiable or reversible data for individual users; 5. Complete user control: Users can withdraw their authorization for group analysis at any time, and their personal data will immediately stop participating in group aggregation after withdrawal. Example 11: Trend Analysis and Boundary Expression of Abnormal Scalp Manifestations

[0067] Once the recognition layer confirms abnormalities such as localized hair loss, diffuse hair loss, changes in hairline, or areas of trauma on the scalp, the attribution module can output the trend, magnitude, and ranking of possible key influencing factors of the abnormal area.

[0068] The identification layer is completed collaboratively by multiple devices. The smart grooming terminal uses an optical scanning module to identify visual features such as smooth areas of the scalp, missing hair follicle openings, scalp scars, or redness and swelling; the smart ring and smart bracelet identify possible abnormalities in the user's grooming behavior; and the smart pendant and smart bracelet identify whether the user is in a recovery period or under stress through physiological signals.

[0069] After the anomaly is confirmed at the identification layer, the attribution module can quantify the magnitude and trend of changes in the abnormal area. All outputs use a "trend description" approach. For example, the system outputs: "The hair follicle density in this area has decreased by 18% in the last four weeks, a rate of decrease higher than the average level of the surrounding area." The system does not use disease diagnostic labels such as "alopecia areata" or "androgenetic alopecia."

[0070] This principle of boundary-based expression ensures that this system always functions as a health management aid. The system only describes phenomena and trends, does not name diseases, and does not replace professional medical diagnosis. Whether to seek medical attention and how to treat them are decisions made independently by the user. Example 12: Long-term application of desensitized personal care behavioral characteristic data (reservation)

[0071] Provided user authorization is obtained and anonymization is completed, the anonymized personal care behavior data, in addition to being used to generate group insight reports, can also be used to train behavior recognition models and health trend prediction models for personal care scenarios. This application belongs to a long-term data asset application scenario, does not change the current core positioning of focusing on personal health management, and does not constitute the necessary technical features of the claims. Example 13: Human-Machine Collaboration for Low-Confidence Attribution and Candidate Conclusions

[0072] When the confidence level of the explainable attribution module or the scenario determination module is lower than the preset threshold, the system does not output a single definite conclusion. Instead, it combines multi-dimensional auxiliary information (time, location, historical behavior, wearing status, etc.) to generate multiple candidate judgments and outputs candidate options to the user. The final intervention suggestion or in-depth explanation is then output after the user confirms the information.

[0073] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural modifications made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for interpretable assessment of scalp health status based on multimodal data, characterized in that, Includes the following steps: S1: Obtain multimodal scalp health data of the user within a set period. The multimodal scalp health data includes at least scalp image feature data, hair loss count data, hair toughness data, washing and care behavior feature data, physiological indicator data, and washing and care product ingredient feature data. S2: Input the multimodal scalp health data into a pre-constructed feature encoder, and use the feature encoder to map the data of different modalities to the same latent space, and fuse them to generate a baseline feature vector representing the user's personal scalp health status. S3: In subsequent monitoring cycles, generate real-time feature vectors in the same manner as S2, and calculate the deviation between the real-time feature vectors and the reference feature vectors; S4: Using at least one interpretable analysis model, perform feature contribution analysis on the deviation, identify the key influencing factors that cause the change in the deviation, and output the contribution ranking of each key influencing factor; S5: Based on the ranking of the key influencing factors and their contributions, automatically generate targeted scalp health intervention suggestions and output them to the user.

2. The method according to claim 1, characterized in that, The methods for obtaining multimodal scalp health data mentioned in S1 include: The system acquires scalp image feature data and hair loss count data through a smart combing terminal; acquires washing and care behavior feature data and physiological indicator data through a wearable terminal; and acquires washing and care product ingredient feature data after user authorization.

3. The method according to claim 2, characterized in that, The washing and care behavior characteristic data includes a quantitative evaluation value for rubbing and covering, which is generated in the following way: The inertial measurement unit signal of the wearable terminal is mapped to the head coordinate system, the spatial distribution of signal energy in each part of the head within the rubbing motion segment is calculated, and the coverage quantification evaluation value of each part is output based on the spatial distribution.

4. The method according to claim 2, characterized in that, The hair toughness data was obtained through the following methods: The voltage waveform signal output by the flexible strain sensor on the combing terminal is collected, pulse detection is performed on the waveform signal, the peak amplitude and rising slope of each pulse are extracted, and the pulse characteristics of one combing process are statistically analyzed to generate a hair toughness assessment value.

5. The method according to claim 1, characterized in that, The feature encoder described in S2 is a multimodal Transformer encoder or an autoencoder, used to map scalp image features, mechanical features, behavioral features, physiological features and component features to a latent space of a unified dimension.

6. The method according to claim 1, characterized in that, The deviation calculation method described in S3 includes any one or more combinations of cosine similarity, Euclidean distance, and Mahalanobis distance.

7. The method according to claim 1, characterized in that, The interpretability analysis model described in S4 includes any one or more combinations of SHAP, LIME, Integrated Gradients, and attention weight analysis.

8. The method according to claim 1, characterized in that, It also includes a dynamic threshold warning step: Based on the fluctuation range of the user's baseline feature vector within the initial period, a dynamic threshold for deviation is automatically set; when the deviation between the real-time feature vector and the baseline feature vector exceeds the dynamic threshold, an early warning is issued to the user.

9. The method according to claim 1, characterized in that, It also includes multiple rounds of in-depth explanation and interaction steps: When users ask follow-up questions about the intervention suggestions, the system generates targeted, in-depth explanations based on the identified key influencing factors and their ranking by contribution, and provides these explanations to the user in natural language.

10. The method according to claim 1, characterized in that, All steps from S1 to S5 are completed locally on the user's terminal, and the cloud server does not access the original multimodal scalp health data.

11. The method according to claim 10, characterized in that, It also includes the steps for generating and outputting insights into desensitized groups: After obtaining user authorization, the user terminal uploads the de-identified structured health feature data to the cloud server; The cloud server aggregates anonymized feature data according to at least one dimension of people, products, regions, and time to generate a group insight report; the group insight report only contains aggregated statistical conclusions and does not contain any identifiable or inferable data of any individual user.

12. A scalp health status interpretable assessment system based on multimodal data, used to implement the method as described in any one of claims 1 to 11, characterized in that, include: The data acquisition module is used to acquire users' multimodal scalp health data; The feature encoding module is used to fuse multimodal data to generate baseline feature vectors and real-time feature vectors; The deviation calculation module is used to calculate the deviation between the real-time feature vector and the baseline feature vector; The Explainable Attribution module is used to identify key influencing factors and output a ranking of their contributions using an interpretable analysis model. The suggestion generation module is used to generate personalized scalp health intervention suggestions based on key influencing factors.

Images