A Pet Health Assessment and Intervention Method Based on Personalized Feature Templates

Abstract

This invention discloses a pet health assessment and intervention method based on personalized feature templates, comprising the following steps: establishing a basic parameter set based on the pet's individual information; collecting multimodal data of the pet through wearable monitoring devices and environmental sensing devices; generating a personalized feature template for the pet based on the collected historical data; performing cross-indicator consistency correction on the collected real-time data using the personalized feature template, and matching the data with the personalized feature template to generate a health assessment report; and generating a personalized intervention plan based on the health assessment report, matching the personalized feature template with a pet health intervention knowledge graph. This invention, based on personalized feature templates and context-adaptive networks, achieves accurate pet health assessment and intelligent intervention, possessing advantages such as high accuracy, strong adaptability, and traceability.

Description

Technical Field

[0001] This invention relates to the field of pet health, and more particularly to a method for pet health assessment and intervention based on personalized feature templates. Background Technology

[0002] In recent years, with the development of smart wearable devices and artificial intelligence technology, animal health monitoring has gradually evolved from manual observation to automation and data-driven approaches. Existing pet health management systems typically collect physiological and behavioral data such as body temperature, heart rate, activity level, and sleep patterns from pets using wearable devices, and then combine this data with mobile terminals or cloud platforms for health assessment and alerts. These systems often rely on fixed thresholds or population statistical models for evaluation, identifying health abnormalities by determining whether monitored indicators exceed standard ranges. For example, some existing technologies utilize multi-sensor fusion models to monitor pet vital signs, triggering alerts when heart rate or body temperature exceeds preset ranges, or generating simple trend reports based on historical data. However, these methods have several limitations in practical applications.

[0003] First, existing technologies generally employ uniform health assessment standards, lacking sufficient consideration of individual differences among pets. Pets of different breeds, ages, sexes, and living environments exhibit significant differences in physiological characteristics and behavioral patterns. A single threshold cannot accurately reflect the normal fluctuation range of an individual, leading to a high rate of false positives or false negatives. Second, traditional health assessment models often rely on static parameters and single-dimensional indicators, failing to effectively integrate the correlations between multimodal data such as physiological, behavioral, and environmental data. They also lack the ability to analyze time-series trends and contextual situations, making it difficult to identify shifts in health status caused by environmental changes or behavioral rhythms.

[0004] Furthermore, existing pet health analysis systems primarily rely on real-time monitoring and single-assessment, lacking a personalized modeling mechanism based on historical data. While some studies have attempted to introduce machine learning or deep learning methods to improve prediction accuracy, these models generally depend on fixed training data, making it difficult to adaptively adjust parameters over time and resulting in insufficient self-correction capabilities for newly collected data, leading to poor model generalization. Simultaneously, health interventions often rely on general recommendations, failing to generate differentiated intervention plans based on the pet's individual characteristics, lifestyle, and circadian rhythms, and lacking targeted behavioral guidance or time-series planning mechanisms.

[0005] In summary, existing pet health assessment and intervention technologies still have significant shortcomings in areas such as personalized modeling, cross-modal data fusion, contextual recognition, and dynamic intervention, making it difficult to achieve a refined depiction of pet health status and long-term dynamic management.

[0006] Therefore, how to provide a pet health assessment and intervention method based on personalized feature templates is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a pet health assessment and intervention method based on personalized feature templates. This invention, based on personalized feature templates and context-adaptive networks, achieves accurate pet health assessment and intelligent intervention, and has the advantages of high accuracy, strong adaptability and traceability.

[0008] A pet health assessment and intervention method based on personalized feature templates according to an embodiment of the present invention includes the following steps:

[0009] A basic parameter set is established based on the individual information of the pets. Multimodal data of the pets is collected through wearable monitoring devices and environmental sensing devices. After data cleaning and feature extraction, a standard dataset is constructed, which includes real-time data and historical data.

[0010] Based on past data in the standard dataset, individualized modeling is performed on the pet to generate a personalized feature template for the pet. The personalized feature template includes a set of individualized baseline intervals for each feature, a set of behavioral rhythm templates, a set of dynamic thresholds, and a correlation matrix between features.

[0011] Personalized feature templates are used to perform cross-indicator consistency correction on real-time data in the standard dataset to obtain a set of corrected feature vectors. The set of corrected feature vectors is then matched and calculated with the personalized feature templates to generate an assessment result that includes health level, a list of risk items, and evidence trajectory, and a health assessment report is generated.

[0012] Based on the evaluation results, personalized intervention plans are generated by matching personalized feature templates with a pet health intervention knowledge graph and using a time-series planning algorithm.

[0013] Optionally, the generation of the standard dataset specifically includes:

[0014] A basic parameter set is established based on the individual information of the pet, including the pet's breed, age, sex, weight, and the group baseline value obtained by statistical calculation in the pet health database based on the pet's breed, age, sex, and weight, as well as the unique identifier assigned to it;

[0015] Complete the binding and channel registration of wearable monitoring devices and environmental sensing devices. Based on a unified timeline, collect raw time-series data of pets’ physiological, behavioral and environmental indicators by channel. Physiological indicators include body temperature, heart rate, blood oxygen saturation and respiratory rate. Behavioral indicators include cadence, activity level and sleep status. Environmental indicators include ambient temperature, ambient humidity and ambient light intensity, forming a raw multimodal dataset.

[0016] The original multimodal dataset is time-aligned and resampled, and the data quality index is calculated using the sliding window method. Missing data is filled, noisy data is smoothed, and outliers are removed to generate a cleaned multimodal dataset.

[0017] Based on the population baseline values ​​in the basic parameter set, the cleaned multimodal dataset is Z-score standardized to construct a standard dataset, which includes both current and historical data.

[0018] Optionally, the generation of the personalized feature template specifically includes:

[0019] Historical feature sequence sets are obtained by dividing historical data from a standard dataset into windows and aligning them with channels according to a unified timeline.

[0020] Calculate individualized baseline intervals for physiological indicators in the historical feature sequence set, calculate the mean and standard deviation of each physiological indicator in the time series, determine the upper and lower bound intervals of the indicators, and summarize the upper and lower bound intervals of all physiological indicators to form an individualized baseline interval set.

[0021] Rhythm analysis is performed on behavioral indicators in the historical feature sequence set. The period length is determined by autocorrelation analysis. The mean, variance and trend of the behavioral indicators corresponding to each time phase are calculated within the period. The rhythm curves of each behavioral indicator are obtained by combining them in the order of time phases, which serve as a set of behavioral rhythm templates.

[0022] Based on environmental and behavioral indicators in the historical feature sequence set, time segments are discretized and classified according to different numerical ranges to obtain context category identifiers. The mean and standard deviation of physiological and behavioral indicators are calculated for each context category to determine the feature threshold range under each context category. The feature threshold ranges generated by the context category corresponding to each time moment are summarized to form a dynamic threshold set.

[0023] Based on physiological, behavioral, and environmental indicators in the historical feature sequence set, the linear correlation between any two indicators in the time series is calculated. The correlation coefficient is determined by the ratio of covariance to standard deviation. A square matrix with features as the dimension is constructed. The diagonal elements of the matrix represent the autocorrelation of features, and the off-diagonal elements represent the linear correlation between different features. The square matrix is ​​statistically significant to retain stable associated feature pairs, forming a feature correlation matrix.

[0024] The individualized baseline interval set, behavioral rhythm template set, dynamic threshold set, and feature correlation matrix are integrated to form a preliminary personalized feature template;

[0025] The initial personalized feature template is calibrated using the validation set in the past data. The coverage rate and threshold trigger rate of the feature interval are calculated. The interval coefficient and threshold coefficient in the initial personalized feature template are optimized in a way that minimizes the coverage error and false alarm error, so as to obtain the calibrated personalized feature template.

[0026] Optionally, the generation of the evaluation results specifically includes:

[0027] Real-time data from the standard dataset is divided into windows and aligned with channels according to a unified timeline to obtain a real-time feature sequence set. Based on environmental and behavioral indicators in the real-time feature sequence set, time segments are discretized and classified according to different numerical ranges to obtain context category identifiers. Standardized feature values ​​are calculated based on the dynamic threshold set in the personalized feature template to obtain a context-standardized feature set.

[0028] Cross-feature consistency correction is performed based on the feature correlation matrix in the personalized feature template. A context-adaptive feature coupling network is constructed, which includes three sub-layers: a physiological feature layer, a behavioral feature layer, and an environmental feature layer. Context modulation coefficients are applied to each edge weight according to the context category identifier of the context-standardized feature set to form a context feature map. Based on the phase and stability parameters of the behavioral rhythm template set, the weights of the behavioral feature layer and its bridging edges to the physiological feature layer are adjusted through a phase gating module. Using the out-of-bounds indication of the dynamic threshold set, the threshold feedback module suppresses or enhances the external edges of nodes that have exceeded the bounds, realizing local feedback correction of feature anomalies. The node confidence is calculated based on the node's data quality index. The confidence modulation module reduces the impact of low-confidence nodes on the network. Combining the historical edge weight adjustment and consistency correction statistics recorded in the state memory unit, the current edge weight is time-smoothed and updated. The context-standardized feature set is input into the network to calculate the weighted consistency estimate of each node and its neighboring nodes. The confidence-weighted node self-value is then fused to generate a cross-feature consistency correction feature set.

[0029] The cross-feature consistency correction feature set is matched with the individualized baseline interval set, and the baseline deviation of each feature is calculated. When the feature value is within the baseline interval, the deviation is zero. When the value exceeds the interval, the deviation is calculated according to the ratio of the excess part to the interval width. The deviation is accumulated using a sliding time window to form a time-series-based baseline deviation set that reflects short-term fluctuation trends.

[0030] The behavioral rhythm template set in the personalized feature template is used to perform rhythm phase mapping on the cross-feature consistency correction feature set, mapping the behavioral feature at the current moment to its rhythm period phase. The rhythm deviation is calculated based on the average curve and phase variance of the behavioral rhythm template set, and a rhythm deviation set is generated by combining the rhythm stability parameter.

[0031] Based on the dynamic threshold set in the personalized feature template, the threshold range of each feature in the current context category is adaptively adjusted. When the feature value is within the upper and lower threshold range, the out-of-bounds degree is zero. When the value exceeds the threshold range, the out-of-bounds degree is calculated according to the ratio of the deviation distance to the width of the threshold range, and a threshold out-of-bounds indication sequence is generated to obtain the dynamic threshold out-of-bounds set.

[0032] The baseline deviation set, rhythm deviation set, and dynamic threshold overrun set are fused into multiple features. The comprehensive deviation of each feature is obtained by weighted calculation. The weights are dynamically adjusted according to the feature correlation matrix in the personalized feature template to generate the comprehensive deviation set.

[0033] A risk item list is constructed based on the comprehensive deviation set and the dynamic threshold over-limit set. When the comprehensive deviation of a feature exceeds the preset risk threshold or the threshold is over-limited, the feature is added to the risk item list. The average comprehensive deviation of all features and the consistency index between features are calculated to generate a comprehensive health score. The comprehensive health score is divided into health levels according to the preset grading threshold.

[0034] The comprehensive deviation of each feature, threshold overrun indication, rhythm phase, context category, and time index are integrated into an evidence trajectory. The output includes the health level, a list of risk items, and the evidence trajectory, generating a corresponding health assessment report.

[0035] Optionally, the generation of the personalized intervention plan specifically includes:

[0036] Based on the health level, risk item list and evidence trajectory in the assessment results, a multi-label classification algorithm is used to identify specific health problems that require intervention. By analyzing the correlation between risk items, the primary and secondary health problems are determined, and a list of health problems to be addressed is formed in order of priority.

[0037] Based on the list of health problems to be addressed, a pre-built pet health intervention knowledge graph is queried. The pet health intervention knowledge graph includes disease symptoms, intervention measures, medication and nutrition, behavior adjustment and their relationships. Intervention measures that match the current health problem are retrieved from the knowledge graph to generate a set of candidate intervention measures.

[0038] Based on the individualized baseline interval and behavioral rhythm template in the personalized feature template, a multi-objective optimization algorithm is used to screen intervention measures. An optimization model is established that considers three objectives: health improvement effect, behavioral rhythm matching degree, and implementation feasibility. The Pareto optimal solution set is solved to obtain a set of personalized intervention measures that conform to individual characteristics.

[0039] Based on the behavioral rhythm template in the personalized feature template, the set of personalized intervention measures is reasonably allocated to time windows that do not conflict with the pet's physiological rhythm, thereby generating a personalized intervention plan.

[0040] The beneficial effects of this invention are:

[0041] This invention overcomes the shortcomings of existing technologies, such as fixed assessment standards, static models, difficulty in reflecting individual differences, and lack of targeted intervention measures, by introducing a pet health assessment and intervention method based on personalized feature templates. By constructing personalized feature templates that include an individualized baseline interval set, a behavioral rhythm template set, a dynamic threshold set, and a feature correlation matrix, it achieves comprehensive modeling of pet physiological, behavioral, and environmental characteristics, transforming health assessment from a uniform threshold judgment to a dynamic analysis based on individual biases. This invention can automatically generate differentiated templates based on the pet's breed, age, sex, weight, and lifestyle habits, allowing health assessment standards to adaptively adjust to individual changes, significantly improving the accuracy and sensitivity of the assessment.

[0042] Meanwhile, this invention utilizes a context-adaptive feature coupling network to achieve cross-index consistency correction of multimodal features. Based on the integration of physiological indicators, behavioral characteristics, and environmental factors, it employs phase gating, threshold feedback, confidence modulation, and state memory mechanisms to self-learn and update the dynamic correlations between features. This effectively eliminates interference from data acquisition errors and environmental noise, ensuring the stability and reliability of health assessment results. During the assessment process, the system not only generates health levels but also produces structured results containing a list of risk items and evidence trajectories, making the health status interpretable and traceable, and providing a quantitative basis for subsequent intervention optimization.

[0043] In the intervention phase, this invention, based on a pet health intervention knowledge graph, establishes a correlation between health risks and intervention measures related to behavior, nutrition, and environment. Through multi-label classification and multi-objective optimization algorithms, it generates a set of personalized intervention measures tailored to individual characteristics. Furthermore, it incorporates behavioral rhythm templates for time window allocation, achieving dynamic matching between intervention content and the pet's physiological rhythms. Thus, this invention forms an intelligent closed-loop system centered on personalized feature templates at both the assessment and intervention ends. This system continuously learns and optimizes the pet health model, achieving precise monitoring, intelligent assessment, and rhythmic intervention, significantly improving the scientific rigor and effectiveness of pet health management. Attached Figure Description

[0044] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0045] Figure 1 This is a flowchart of a pet health assessment and intervention method based on personalized feature templates proposed in this invention;

[0046] Figure 2 This is a schematic diagram of a personalized feature template for a pet health assessment and intervention method based on personalized feature templates proposed in this invention. Detailed Implementation

[0047] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0048] refer to Figure 1-2 A pet health assessment and intervention method based on personalized feature templates includes the following steps:

[0049] A basic parameter set is established based on the individual information of the pets. Multimodal data of the pets is collected through wearable monitoring devices and environmental sensing devices. After data cleaning and feature extraction, a standard dataset is constructed, which includes real-time data and historical data.

[0050] Based on past data in the standard dataset, individualized modeling is performed on the pet to generate a personalized feature template for the pet. The personalized feature template includes a set of individualized baseline intervals for each feature, a set of behavioral rhythm templates, a set of dynamic thresholds, and a correlation matrix between features.

[0051] Personalized feature templates are used to perform cross-indicator consistency correction on real-time data in the standard dataset to obtain a set of corrected feature vectors. The set of corrected feature vectors is then matched and calculated with the personalized feature templates to generate an assessment result that includes health level, a list of risk items, and evidence trajectory, and a health assessment report is generated.

[0052] Based on the evaluation results, personalized intervention plans are generated by matching personalized feature templates with a pet health intervention knowledge graph and using a time-series planning algorithm.

[0053] In this embodiment, the generation of the standard dataset specifically includes:

[0054] A basic parameter set is established based on the individual information of the pet, including the pet's breed, age, sex, weight, and the group baseline value obtained by statistical calculation in the pet health database based on the pet's breed, age, sex, and weight, as well as the unique identifier assigned to it;

[0055] Complete the binding and channel registration of wearable monitoring devices and environmental sensing devices. Based on a unified timeline, collect raw time-series data of pets’ physiological, behavioral and environmental indicators by channel. Physiological indicators include body temperature, heart rate, blood oxygen saturation and respiratory rate. Behavioral indicators include cadence, activity level and sleep status. Environmental indicators include ambient temperature, ambient humidity and ambient light intensity, forming a raw multimodal dataset.

[0056] The original multimodal dataset is time-aligned and resampled, and the data quality index is calculated using the sliding window method. Missing data is filled, noisy data is smoothed, and outliers are removed to generate a cleaned multimodal dataset.

[0057] Based on the population baseline values ​​in the basic parameter set, the cleaned multimodal dataset is Z-score standardized to construct a standard dataset, which includes both current and historical data.

[0058] In this embodiment, the generation of the personalized feature template specifically includes:

[0059] Historical feature sequence sets are obtained by dividing historical data from a standard dataset into windows and aligning them with channels according to a unified timeline.

[0060] Calculate individualized baseline intervals for physiological indicators in the historical feature sequence set, calculate the mean and standard deviation of each physiological indicator in the time series, determine the upper and lower bound intervals of the indicators, and summarize the upper and lower bound intervals of all physiological indicators to form an individualized baseline interval set.

[0061] Rhythm analysis is performed on behavioral indicators in the historical feature sequence set. The period length is determined by autocorrelation analysis. The mean, variance and trend of the behavioral indicators corresponding to each time phase are calculated within the period. The rhythm curves of each behavioral indicator are obtained by combining them in the order of time phases, which serve as a set of behavioral rhythm templates.

[0062] Based on environmental and behavioral indicators in the historical feature sequence set, time segments are discretized and classified according to different numerical ranges to obtain context category identifiers. The mean and standard deviation of physiological and behavioral indicators are calculated for each context category to determine the feature threshold range under each context category. The feature threshold ranges generated by the context category corresponding to each time moment are summarized to form a dynamic threshold set.

[0063] Based on physiological, behavioral, and environmental indicators in the historical feature sequence set, the linear correlation between any two indicators in the time series is calculated. The correlation coefficient is determined by the ratio of covariance to standard deviation. A square matrix with features as the dimension is constructed. The diagonal elements of the matrix represent the autocorrelation of features, and the off-diagonal elements represent the linear correlation between different features. The square matrix is ​​statistically significant to retain stable associated feature pairs, forming a feature correlation matrix.

[0064] The individualized baseline interval set, behavioral rhythm template set, dynamic threshold set, and feature correlation matrix are integrated to form a preliminary personalized feature template;

[0065] The initial personalized feature template is calibrated using the validation set in the past data. The coverage rate and threshold trigger rate of the feature interval are calculated. The interval coefficient and threshold coefficient in the initial personalized feature template are optimized in a way that minimizes the coverage error and false alarm error, so as to obtain the calibrated personalized feature template.

[0066] In this embodiment, the generation of the evaluation result specifically includes:

[0067] Real-time data from the standard dataset is divided into windows and aligned with channels according to a unified timeline to obtain a real-time feature sequence set. Based on environmental and behavioral indicators in the real-time feature sequence set, time segments are discretized and classified according to different numerical ranges to obtain context category identifiers. Standardized feature values ​​are calculated based on the dynamic threshold set in the personalized feature template to obtain a context-standardized feature set.

[0068] Cross-feature consistency correction is performed based on the feature correlation matrix in the personalized feature template. A context-adaptive feature coupling network is constructed, which includes three sub-layers: a physiological feature layer, a behavioral feature layer, and an environmental feature layer. Context modulation coefficients are applied to each edge weight according to the context category identifier of the context-standardized feature set to form a context feature map. Based on the phase and stability parameters of the behavioral rhythm template set, the weights of the behavioral feature layer and its bridging edges to the physiological feature layer are adjusted through a phase gating module. Using the out-of-bounds indication of the dynamic threshold set, the threshold feedback module suppresses or enhances the external edges of nodes that have exceeded the bounds, realizing local feedback correction of feature anomalies. The node confidence is calculated based on the node's data quality index. The confidence modulation module reduces the impact of low-confidence nodes on the network. Combining the historical edge weight adjustment and consistency correction statistics recorded in the state memory unit, the current edge weight is time-smoothed and updated. The context-standardized feature set is input into the network to calculate the weighted consistency estimate of each node and its neighboring nodes. The confidence-weighted node self-value is then fused to generate a cross-feature consistency correction feature set.

[0069] The cross-feature consistency correction feature set is matched with the individualized baseline interval set, and the baseline deviation of each feature is calculated. When the feature value is within the baseline interval, the deviation is zero. When the value exceeds the interval, the deviation is calculated according to the ratio of the excess part to the interval width. The deviation is accumulated using a sliding time window to form a time-series-based baseline deviation set that reflects short-term fluctuation trends.

[0070] The behavioral rhythm template set in the personalized feature template is used to perform rhythm phase mapping on the cross-feature consistency correction feature set, mapping the behavioral feature at the current moment to its rhythm period phase. The rhythm deviation is calculated based on the average curve and phase variance of the behavioral rhythm template set, and a rhythm deviation set is generated by combining the rhythm stability parameter.

[0071] Based on the dynamic threshold set in the personalized feature template, the threshold range of each feature in the current context category is adaptively adjusted. When the feature value is within the upper and lower threshold range, the out-of-bounds degree is zero. When the value exceeds the threshold range, the out-of-bounds degree is calculated according to the ratio of the deviation distance to the width of the threshold range, and a threshold out-of-bounds indication sequence is generated to obtain the dynamic threshold out-of-bounds set.

[0072] The baseline deviation set, rhythm deviation set, and dynamic threshold overrun set are fused into multiple features. The comprehensive deviation of each feature is obtained by weighted calculation. The weights are dynamically adjusted according to the feature correlation matrix in the personalized feature template to generate the comprehensive deviation set.

[0073] A risk item list is constructed based on the comprehensive deviation set and the dynamic threshold over-limit set. When the comprehensive deviation of a feature exceeds the preset risk threshold or the threshold is over-limited, the feature is added to the risk item list. The average comprehensive deviation of all features and the consistency index between features are calculated to generate a comprehensive health score. The comprehensive health score is divided into health levels according to the preset grading threshold.

[0074] The comprehensive deviation of each feature, threshold overrun indication, rhythm phase, context category, and time index are integrated into an evidence trajectory. The output includes the health level, a list of risk items, and the evidence trajectory, generating a corresponding health assessment report.

[0075] In this embodiment, the generation of the personalized intervention plan specifically includes:

[0076] Based on the health level, risk item list and evidence trajectory in the assessment results, a multi-label classification algorithm is used to identify specific health problems that require intervention. By analyzing the correlation between risk items, the primary and secondary health problems are determined, and a list of health problems to be addressed is formed in order of priority.

[0077] Based on the list of health problems to be addressed, a pre-built pet health intervention knowledge graph is queried. The pet health intervention knowledge graph includes disease symptoms, intervention measures, medication and nutrition, behavior adjustment and their relationships. Intervention measures that match the current health problem are retrieved from the knowledge graph to generate a set of candidate intervention measures.

[0078] Based on the individualized baseline interval and behavioral rhythm template in the personalized feature template, a multi-objective optimization algorithm is used to screen intervention measures. An optimization model is established that considers three objectives: health improvement effect, behavioral rhythm matching degree, and implementation feasibility. The Pareto optimal solution set is solved to obtain a set of personalized intervention measures that conform to individual characteristics.

[0079] Based on the behavioral rhythm template in the personalized feature template, the set of personalized intervention measures is reasonably allocated to time windows that do not conflict with the pet's physiological rhythm, thereby generating a personalized intervention plan.

[0080] Example 1:

[0081] To verify the feasibility of this invention in practice, it was applied to a long-term health monitoring scenario at a smart pet health management center. The center houses and fosters a large number of pets, primarily cats and dogs, with relatively stable living environments but significant individual differences. Previous health monitoring methods relied on manual inspections and equipment alarms, primarily judging indicators such as body temperature, heart rate, and activity level based on fixed thresholds. This approach ignores differences in breed, size, age, sex, and lifestyle habits among pets, easily leading to assessment errors. For example, low activity levels in older dogs are often misjudged as health abnormalities, while high heart rates in high-metabolic cats are often seen as danger signals, resulting in frequent false alarms, high costs of manual intervention, and low assessment accuracy. This invention addresses and validates these problems through a pet health assessment and intervention method based on personalized feature templates.

[0082] During implementation, pets wore wearable monitoring collars, and environmental sensing devices were deployed in their living spaces. The wearable collars collected real-time physiological and behavioral data such as body temperature, heart rate, respiratory rate, cadence, activity level, and sleep status, while the environmental sensing devices recorded external conditions such as temperature, humidity, and light intensity. All collected data was wirelessly transmitted and then synchronized along a unified timeline to resolve errors caused by delays and drift in data from different sensors. After data collection, a sliding window method was used to clean and quality control the multimodal data, removing outliers and imputing and smoothing missing data to create a standard dataset containing both real-time and historical data. Based on this, a population baseline was constructed according to the pets' basic parameters, such as breed, sex, weight, and age, and individualized reference ranges were generated, providing a data foundation for subsequent personalized modeling.

[0083] As data accumulates, the method of this invention is used to create individualized models for each pet. By analyzing historical data, individualized baseline interval sets are established for various physiological indicators. Time series analysis is used to identify behavioral patterns and construct behavioral rhythm templates. For example, the behavioral rhythm template of a medium-sized dog shows a clear diurnal activity pattern, while a domestic cat exhibits stronger nocturnal activity. The establishment of behavioral rhythm templates enables the assessment mechanism to identify rhythmic behavioral fluctuations, avoiding misjudging normal periodic changes as abnormalities. Simultaneously, based on a combination of environmental and behavioral indicators, time segments are divided into multiple contextual categories, such as "high temperature, high activity" and "low temperature, resting." Based on the statistical characteristics of physiological and behavioral indicators in each contextual category, a dynamic threshold set is determined, allowing the assessment criteria for pets under different environmental and behavioral states to be flexibly adjusted, forming a dynamic assessment baseline that changes with the context.

[0084] During the health assessment phase, personalized feature templates are used to perform cross-indicator consistency correction on real-time data. Specifically, based on the feature correlation matrix, physiological, behavioral, and environmental indicators are constructed into a context-adaptive feature coupling network. A phase gating module, combined with phase information from the behavioral rhythm template, adjusts the weights of rhythm-related features in real time. A threshold feedback module monitors feature out-of-bounds states; when an indicator deviates, the deviation is dynamically corrected based on the coupling relationship of related features. Combined with a confidence modulation mechanism, the influence of low-confidence features is attenuated based on the quality indicators of multimodal data, thereby improving overall robustness and assessment reliability. After cross-feature consistency correction, the resulting corrected feature set more accurately reflects the pet's overall health status, reducing misjudgments caused by single abnormalities.

[0085] Building upon this foundation, by matching individualized baseline interval sets with behavioral rhythm templates, feature deviation and rhythm deviation are calculated. These are then combined with a dynamic threshold overrun set for comprehensive analysis, outputting a health level, a list of risk items, and an evidence trajectory. The health level represents the overall health condition, the risk item list displays features exhibiting significant deviations, and the evidence trajectory records the evolution of deviations for each feature across different times, contexts, and rhythmic states. Through these results, managers can clearly understand the basis and temporal correlation of each health abnormality, thereby enabling evidence-based decision-making.

[0086] When pets exhibit persistent deviations or potential risks, personalized intervention plans are developed based on health assessment results. First, a health intervention knowledge graph is used to establish a matching relationship between risk items and interventions such as disease, nutrition, environmental regulation, and behavioral adjustment. Then, combining physiological and rhythmic information from individualized characteristic templates, a multi-objective optimization algorithm is used to select the set of interventions most suitable for the individual's characteristics. Finally, intervention times are rationally scheduled based on behavioral rhythm templates to ensure that intervention activities align with the pet's physiological rhythms. For example, for cats with low nighttime activity levels, feeding and playtime are adjusted; for dogs with elevated body temperatures in hot weather, cooling and rest time windows are planned. In this way, intervention measures not only conform to the pet's individual characteristics but also align with its daily routine, significantly improving the effectiveness and sustainability of health management.

[0087] In long-term operation, this method has been validated through continuous health monitoring and dynamic intervention on multiple pets. Practice shows that the personalized feature template can continuously self-calibrate, automatically adjusting feature intervals and threshold ranges over time, maintaining the flexibility and long-term accuracy of the assessment model. The context-adaptive feature coupling mechanism effectively reduces the bias between multimodal data, making health assessment results more stable and reliable. Compared with previous methods using fixed thresholds, the method of this invention shows significant advantages in identifying health abnormalities, reducing false alarms, and improving assessment interpretability. Furthermore, by combining knowledge graphs and behavioral rhythm templates to generate personalized intervention programs, a closed-loop process from health monitoring to intelligent intervention is achieved, making pet health management more scientific, precise, and sustainable.

[0088] In summary, the implementation of this invention demonstrates that the pet health assessment and intervention method based on personalized feature templates can effectively overcome the problems of rigid assessment standards, lack of reflection of individual differences, insufficient contextual adaptability, and lack of targeted intervention measures in existing technologies. Through in-depth modeling and dynamic contextual analysis of individual pet characteristics, this invention achieves optimization of the entire process from data collection and health assessment to intelligent intervention, providing a novel method for pet health management that is highly accurate, highly adaptable, and highly interpretable.

[0089] Table 1. Performance comparison between personalized feature template-based pet health assessment and intervention methods and traditional methods.

[0090] Health assessment accuracy 78.2% 84.6% 82.1% 94.3% 11.5 False alarm rate 15.6% 10.8% 12.3% 4.7% 56.5 Health risk identification rate 70.3% 81.4% 78.9% 93.2% 14.5 Cross-modal data consistency score 72.4% 80.1% 77.2% 92.5% 15.5 Assess stability (variance coefficient of results over different time periods) 0.21 0.17 0.19 0.08 57.1 Intervention execution matching degree 68.7% 75.3% 73.9% 91.8% 20.8

[0091] As shown in Table 1, the pet health assessment and intervention method based on personalized feature templates proposed in this invention significantly outperforms existing technologies in all indicators. Regarding health assessment accuracy, this method achieves 94.3%, an improvement of approximately 11.5 percentage points compared to the traditional fixed threshold method. This is mainly attributed to the accurate modeling of individual pet differences by the personalized feature templates. The false positive rate decreased from 15.6% in the traditional method to 4.7%, a reduction of over 56%, effectively solving the misjudgment problem caused by static thresholds. Furthermore, this method improves the health risk identification rate and cross-modal data consistency score by approximately 14% and 15%, respectively, indicating that the context-adaptive feature coupling network plays a crucial role in multimodal data fusion and cross-index correction. In terms of assessment stability, the result fluctuation coefficient decreased from 0.21 to 0.08, indicating that the system's assessment results are more stable across different time periods and possess higher robustness. The intervention execution matching rate improved to 91.8%, 16 percentage points higher than machine learning prediction methods. This is attributed to the personalized intervention plan generated by combining behavioral rhythm templates and multi-objective optimization algorithms, which can dynamically match the timing and content of interventions based on the pet's rhythmic characteristics and health status. In summary, the method of this invention achieves a closed-loop improvement from health monitoring to intervention execution through personalized modeling, cross-modal correction, and intelligent intervention optimization, significantly improving the accuracy, stability, and practicality of the assessment.

[0092] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for pet health assessment and intervention based on personalized feature templates, characterized in that, Includes the following steps: A basic parameter set is established based on the individual information of the pets. Multimodal data of the pets is collected through wearable monitoring devices and environmental sensing devices. After data cleaning and feature extraction, a standard dataset is constructed, which includes real-time data and historical data. Based on past data in the standard dataset, individualized modeling is performed on the pet to generate a personalized feature template for the pet. The personalized feature template includes a set of individualized baseline intervals for each feature, a set of behavioral rhythm templates, a set of dynamic thresholds, and a correlation matrix between features. Personalized feature templates are used to perform cross-indicator consistency correction on real-time data in the standard dataset to obtain a set of corrected feature vectors. The set of corrected feature vectors is then matched and calculated with the personalized feature templates to generate an assessment result that includes health level, a list of risk items, and evidence trajectory, and a health assessment report is generated. Based on the evaluation results, personalized intervention plans are generated by matching personalized feature templates with a pet health intervention knowledge graph and using a time-series planning algorithm.

2. The pet health assessment and intervention method based on individualized feature templates according to claim 1, characterized in that, The generation of the standard dataset specifically includes: A basic parameter set is established based on the individual information of the pet, including the pet's breed, age, sex, weight, and the group baseline value obtained by statistical calculation in the pet health database based on the pet's breed, age, sex, and weight, as well as the unique identifier assigned to it; Complete the binding and channel registration of wearable monitoring devices and environmental sensing devices. Based on a unified timeline, collect raw time-series data of pets’ physiological, behavioral and environmental indicators by channel. Physiological indicators include body temperature, heart rate, blood oxygen saturation and respiratory rate. Behavioral indicators include cadence, activity level and sleep status. Environmental indicators include ambient temperature, ambient humidity and ambient light intensity, forming a raw multimodal dataset. The original multimodal dataset is time-aligned and resampled, and the data quality index is calculated using the sliding window method. Missing data is filled, noisy data is smoothed, and outliers are removed to generate a cleaned multimodal dataset. Based on the population baseline values ​​in the basic parameter set, the cleaned multimodal dataset is Z-score standardized to construct a standard dataset, which includes both current and historical data.

3. The pet health assessment and intervention method based on individualized feature templates of claim 1, wherein, The generation of the personalized feature template specifically includes: Historical feature sequence sets are obtained by dividing historical data from a standard dataset into windows and aligning them with channels according to a unified timeline. Calculate individualized baseline intervals for physiological indicators in the historical feature sequence set, calculate the mean and standard deviation of each physiological indicator in the time series, determine the upper and lower bound intervals of the indicators, and summarize the upper and lower bound intervals of all physiological indicators to form an individualized baseline interval set. Rhythm analysis is performed on behavioral indicators in the historical feature sequence set. The period length is determined by autocorrelation analysis. The mean, variance and trend of the behavioral indicators corresponding to each time phase are calculated within the period. The rhythm curves of each behavioral indicator are obtained by combining them in the order of time phases, which serve as a set of behavioral rhythm templates. Based on environmental and behavioral indicators in the historical feature sequence set, time segments are discretized and classified according to different numerical ranges to obtain context category identifiers. The mean and standard deviation of physiological and behavioral indicators are calculated for each context category to determine the feature threshold range under each context category. The feature threshold ranges generated by the context category corresponding to each time moment are summarized to form a dynamic threshold set. Based on physiological, behavioral, and environmental indicators in the historical feature sequence set, the linear correlation between any two indicators in the time series is calculated. The correlation coefficient is determined by the ratio of covariance to standard deviation. A square matrix with features as the dimension is constructed. The diagonal elements of the matrix represent the autocorrelation of features, and the off-diagonal elements represent the linear correlation between different features. The square matrix is ​​statistically significant to retain stable associated feature pairs, forming a feature correlation matrix. The individualized baseline interval set, behavioral rhythm template set, dynamic threshold set, and feature correlation matrix are integrated to form a preliminary personalized feature template; The initial personalized feature template is calibrated using the validation set in the past data. The coverage rate and threshold trigger rate of the feature interval are calculated. The interval coefficient and threshold coefficient in the initial personalized feature template are optimized in a way that minimizes the coverage error and false alarm error, so as to obtain the calibrated personalized feature template.

4. The pet health assessment and intervention method based on individualized feature templates of claim 1, wherein, The generation of the evaluation results specifically includes: Real-time data from the standard dataset is divided into windows and aligned with channels according to a unified timeline to obtain a real-time feature sequence set. Based on environmental and behavioral indicators in the real-time feature sequence set, time segments are discretized and classified according to different numerical ranges to obtain context category identifiers. Standardized feature values ​​are calculated based on the dynamic threshold set in the personalized feature template to obtain a context-standardized feature set. Cross-feature consistency correction is performed based on the feature correlation matrix in the personalized feature template. A context-adaptive feature coupling network is constructed, which includes three sub-layers: a physiological feature layer, a behavioral feature layer, and an environmental feature layer. Context modulation coefficients are applied to each edge weight according to the context category identifier of the context-standardized feature set to form a context feature map. Based on the phase and stability parameters of the behavioral rhythm template set, the weights of the behavioral feature layer and its bridging edges to the physiological feature layer are adjusted through a phase gating module. Using the out-of-bounds indication of the dynamic threshold set, the threshold feedback module suppresses or enhances the external edges of nodes that have exceeded the bounds, realizing local feedback correction of feature anomalies. The node confidence is calculated based on the node's data quality index. The confidence modulation module reduces the impact of low-confidence nodes on the network. Combining the historical edge weight adjustment and consistency correction statistics recorded in the state memory unit, the current edge weight is time-smoothed and updated. The context-standardized feature set is input into the network to calculate the weighted consistency estimate of each node and its neighboring nodes. The confidence-weighted node self-value is then fused to generate a cross-feature consistency correction feature set. The cross-feature consistency correction feature set is matched with the individualized baseline interval set, and the baseline deviation of each feature is calculated. When the feature value is within the baseline interval, the deviation is zero. When the value exceeds the interval, the deviation is calculated according to the ratio of the excess part to the interval width. The deviation is accumulated using a sliding time window to form a time-series-based baseline deviation set that reflects short-term fluctuation trends. The behavioral rhythm template set in the personalized feature template is used to perform rhythm phase mapping on the cross-feature consistency correction feature set, mapping the behavioral feature at the current moment to its rhythm period phase. The rhythm deviation is calculated based on the average curve and phase variance of the behavioral rhythm template set, and a rhythm deviation set is generated by combining the rhythm stability parameter. Based on the dynamic threshold set in the personalized feature template, the threshold range of each feature in the current context category is adaptively adjusted. When the feature value is within the upper and lower threshold range, the out-of-bounds degree is zero. When the value exceeds the threshold range, the out-of-bounds degree is calculated according to the ratio of the deviation distance to the width of the threshold range, and a threshold out-of-bounds indication sequence is generated to obtain the dynamic threshold out-of-bounds set. The baseline deviation set, rhythm deviation set, and dynamic threshold overrun set are fused into multiple features. The comprehensive deviation of each feature is obtained by weighted calculation. The weights are dynamically adjusted according to the feature correlation matrix in the personalized feature template to generate the comprehensive deviation set. A risk item list is constructed based on the comprehensive deviation set and the dynamic threshold over-limit set. When the comprehensive deviation of a feature exceeds the preset risk threshold or the threshold is over-limited, the feature is added to the risk item list. The average comprehensive deviation of all features and the consistency index between features are calculated to generate a comprehensive health score. The comprehensive health score is divided into health levels according to the preset grading threshold. The comprehensive deviation of each feature, threshold overrun indication, rhythm phase, context category, and time index are integrated into an evidence trajectory. The output includes the health level, a list of risk items, and the evidence trajectory, generating a corresponding health assessment report.

5. The pet health assessment and intervention method based on individualized feature templates of claim 1, wherein, The generation of the personalized intervention plan specifically includes: Based on the health level, risk item list and evidence trajectory in the assessment results, a multi-label classification algorithm is used to identify specific health problems that require intervention. By analyzing the correlation between risk items, the primary and secondary health problems are determined, and a list of health problems to be addressed is formed in order of priority. Based on the list of health problems to be addressed, a pre-built pet health intervention knowledge graph is queried. The pet health intervention knowledge graph includes disease symptoms, intervention measures, medication and nutrition, behavior adjustment and their relationships. Intervention measures that match the current health problem are retrieved from the knowledge graph to generate a set of candidate intervention measures. Based on the individualized baseline interval and behavioral rhythm template in the personalized feature template, a multi-objective optimization algorithm is used to screen intervention measures. An optimization model is established that considers three objectives: health improvement effect, behavioral rhythm matching degree, and implementation feasibility. The Pareto optimal solution set is solved to obtain a set of personalized intervention measures that conform to individual characteristics. Based on the behavior rhythm template in the personalized feature template, the personalized intervention measure set is reasonably distributed to a time window that does not conflict in time and conforms to the physiological rhythm of the pet, and a personalized intervention scheme is generated.

Images