Method, device and medium for generating body abnormal state information based on ai model

Abstract

This invention discloses a method for generating abnormal bodily state information based on an AI model. The method collects continuous user action data using a wearable device, identifies candidate drinking actions using a first AI model, and determines the actual drinking action using a second AI model based on the context sequence. It analyzes drinking behavior trends based on actual drinking actions over a continuous time period and estimates fluid intake based on action characteristics. The drinking behavior trend, intake, and user profile information are input into a third AI model to generate abnormal bodily state labels and confidence levels. When the frequency and confidence level of the label within the statistical period meet preset conditions, an abnormal bodily state alert is automatically generated and output. This invention achieves reliable conversion from behavioral data to personalized abnormal state alerts through multi-level AI models for seamless monitoring and quantitative analysis of daily behavior, providing a feasible technical means for early health observation.

Description

Technical Field

[0001] This invention relates to the field of intelligent health monitoring technology, and in particular to a method, device and medium for generating abnormal bodily state information based on an AI model. Background Technology

[0002] With the increasing awareness of health among the general public, the demand for preventive monitoring of the correlation between daily behaviors, especially drinking habits, and chronic diseases (such as kidney disease) is growing. Currently, there is a lack of effective, routine, and non-invasive monitoring and assessment methods for specific occupational groups (such as taxi drivers and office workers) that may experience abnormal physical conditions due to prolonged fixed postures and irregular drinking habits.

[0003] In existing technologies, directly monitoring an individual's water intake presents significant challenges. Camera-based visual recognition methods are limited by privacy concerns, equipment deployment costs, and coverage limitations, hindering widespread adoption. While wearable devices with high penetration rates (such as smart bracelets) have built-in sensors that can collect human motion data, relying solely on single, instantaneous motion recognition (such as raising a hand) to determine drinking behavior results in low accuracy, a high false positive rate, and an inability to reliably extrapolate actual intake, thus failing to support meaningful assessments of abnormal bodily states.

[0004] Furthermore, existing health promotion or general reminder methods cannot provide quantitative assessments and personalized feedback based on individual behavioral patterns. Whether users follow health advice and their long-term behavioral trends lack objective and continuous monitoring data. Therefore, a technological solution is needed that can utilize existing widely available devices to indirectly and trend-wise assess specific abnormal physical states through continuous behavioral analysis and provide effective alerts. Summary of the Invention

[0005] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows: According to a first aspect of the present invention, a method for generating abnormal bodily state information based on an AI model is provided, the method comprising the following steps: The user's continuous motion data is collected through sensors in wearable devices.

[0006] Based on the continuous action data, candidate drinking actions are identified using the first AI model.

[0007] Based on the context sequence of the candidate drinking actions, the second AI model determines the actual drinking action.

[0008] Based on the real drinking actions identified within a continuous preset time period, the user's drinking behavior trend is determined.

[0009] Based on the characteristic parameters of the actual drinking action, the user's fluid intake is estimated.

[0010] The drinking behavior trend, the liquid intake, and the user's preset profile information are input into the third AI model to generate and output abnormal body state labels and confidence levels.

[0011] When the frequency of the abnormal physical state label in a preset statistical period reaches a preset threshold and the confidence level of the abnormal physical state label meets the preset confidence enhancement condition, an abnormal physical state prompt message is generated and output.

[0012] According to a second aspect of the present invention, an electronic device is provided, including a processor and a memory; the processor executes the steps of the method described in the first aspect of the present invention by invoking a program or instructions stored in the memory.

[0013] According to a third aspect of the present invention, a computer-readable storage medium is provided that stores a program or instructions that cause a computer to perform the steps of the method described in the first aspect of the present invention.

[0014] The present invention has at least the following beneficial effects: (1) Breakthrough in the technical bottleneck of non-invasive precision monitoring: By utilizing the widespread sensors of consumer-grade wearable devices and through the collaboration of multi-level AI models, high-precision, non-intrusive long-term monitoring of subtle, daily health behaviors such as drinking water has been achieved, solving the long-standing problem of the lack of objective and continuous quantitative methods in this field.

[0015] (2) A reliable transformation from behavioral data to abnormal physical state has been achieved: Through the progressive analysis framework of action recognition-trend analysis-multi-source fusion assessment, the original action signal is transformed into a stable personal behavior pattern, and combined with physiological characteristics, personal profile and environmental context, dynamic and personalized quantitative assessment of abnormal physical state is carried out, which greatly improves the scientificity and specificity of abnormal judgment.

[0016] (3) It provides a scientific and intervention-friendly closed loop for preventive health management: It introduces a triggering mechanism based on long-term confidence trends to ensure that health alerts originate from stable and reliable abnormal state patterns, rather than occasional fluctuations. Finally, it outputs graded alert information that is tailored to specific scenarios, forming a complete preventive health management solution from early abnormality identification to effective behavioral intervention, which has significant practical value.

[0017] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a method for generating abnormal bodily state information based on an AI model, as provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. A process can be terminated when its operation is complete, but it may also have additional steps not included in the figures. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0023] This invention provides a method for generating abnormal bodily state information based on an AI model, such as... Figure 1 As shown, the method includes the following steps: The S100 collects continuous motion data from the user through sensors in a wearable device.

[0024] In this invention, continuous motion data refers to raw time-series signals continuously collected from a wearable device worn by the user at a preset sampling frequency, capable of characterizing the user's hand and wrist movements and related physiological states. Depending on the device's sensor configuration, the specific composition and acquisition method of the continuous motion data have the following two embodiments: Example 1: Single-modal motion data based on inertial sensors This embodiment addresses the most basic implementation conditions, and its core is to acquire pure kinematic data of the user's hand.

[0025] In this embodiment, the wearable device is a common consumer-grade smart wearable device that integrates an inertial measurement unit, such as a smart bracelet or smartwatch. This wearable device is a general-purpose device worn by users daily to record exercise or health data, requiring no special modification or additional hardware.

[0026] In this embodiment, continuous motion data specifically refers to, and only refers to, the sequence of inertial signals generated by the inertial measurement unit. The device continuously samples at a preset frequency (e.g., 20Hz to 200Hz) to obtain raw data containing the following two sets of sequences: Triaxial acceleration sequence: reflects the changes in motion acceleration of the equipment in each linear direction.

[0027] Three-axis angular velocity sequence: reflects the changes in the rotational angular velocity of the equipment around each axis.

[0028] These two sets of sequences together form the raw data foundation for analyzing users' spatial movement patterns such as "raising hands," "looking up," and "putting down."

[0029] Example 2: Multimodal motion data based on inertial and physiological sensors This embodiment is an enhancement of Embodiment 1, aiming to provide a richer and more discriminative data foundation for subsequent analysis by fusing multi-dimensional sensor signals.

[0030] In this embodiment, the wearable device is a smart wearable device that integrates an inertial measurement unit and at least one physiological signal sensor, typically represented by mainstream smart bracelets or smartwatches. Typical physiological signal sensors include: optical heart rate sensors or blood volume pulse sensors, which are standard in most smart bracelets or watches and are used to monitor heart rate and pulse waves; bioelectrical impedance sensors, which are becoming increasingly common in mid-to-high-end devices and are used to measure body fat or respiratory rate; electrodermal activity sensors, found in some devices focused on health monitoring; and bone conduction microphones or myosophyll microphones, which, under strict adherence to privacy guidelines such as local processing, can be used to collect surface vibrations or internal acoustic signals close to the skin.

[0031] In this embodiment, continuous motion data is expanded to synchronously acquired multimodal time series. The wearable device acquires the following two types of data streams in parallel in a time-synchronized or aligned manner: Inertial signal sequence: a triaxial acceleration and triaxial angular velocity sequence that is exactly the same as that in Example 1.

[0032] Physiological signal sequence: a synchronized data stream from at least one of the physiological sensors. Depending on the specific configuration of the device, this sequence may include one or more of the following signals: Optical volumetric pulse wave signal: acquired by an optical heart rate sensor, containing rhythmic information such as heart rate and respiratory rate.

[0033] Bioelectrical impedance signal: Collected by a bioelectrical impedance sensor, it reflects changes in thoracic respiratory impedance.

[0034] Myosophytic or body surface vibration signals: acquired by a high-sensitivity accelerometer or myosophytic microphone, which may contain characteristic waveforms of events such as swallowing.

[0035] S200, based on the continuous action data, candidate drinking actions are identified through the first AI model.

[0036] This step aims to initially identify potential drinking motion segments, or candidate drinking motions, from the continuous motion data collected by S100. Regardless of whether the continuous motion data contains physiological signals, this step is based on the inertial signal sequence within it.

[0037] This step specifically includes: S201, the inertial signal sequence (i.e., the triaxial acceleration and triaxial angular velocity sequence) is input into the first AI model, and the positioning results of one or more candidate motion segments are output. Each positioning result includes the start and end positions of the candidate motion segment in the inertial signal sequence and a preliminary confidence level.

[0038] The first AI model, after training, is able to identify, in real time, action segments with the typical spatiotemporal characteristic of "holding a container, raising it, and bringing it close to the mouth" from complex, continuous daily action flows. Its output localization result is a preliminary judgment of this type of characteristic.

[0039] S202, based on the output positioning results, each candidate action segment is formally marked as a candidate drinking action, and the start and end timestamps corresponding to the candidate action segment are extracted from the start and end position information of the candidate action segment as the standardized output of the candidate drinking action.

[0040] In this invention, the first AI model can be implemented using any model architecture suitable for time series detection or classification, and its function is to perform rapid and preliminary screening from the original signal to candidate segments. For example, the model can be selected from, but is not limited to: one-dimensional convolutional neural network (1D-CNN), temporal convolutional network (TCN), or long short-term memory network (LSTM).

[0041] S300, based on the context sequence of the candidate drinking actions, the second AI model determines the actual drinking action.

[0042] This step receives the candidate drinking actions and their start and end timestamps output from step S200, and verifies them using different methods depending on the data type collected by step S100: Method 1: Verification based on inertial context (when the data contains only inertial signals) To reduce false positives (such as misinterpreting scratching one's head or answering a phone call as drinking water), a time-context-based verification mechanism is introduced. Specifically: (1) Taking the time of occurrence of the candidate drinking action as the center, extract the inertial signal sequence numbers within the first number of time windows preceding the center and the second number of time windows following the center to form an extended data window. The first number and the second number can be set according to the actual application scenario, and their values ​​should be such that the extracted context time window can cover the preceding and following actions of a complete drinking behavior. As an exemplary and non-limiting implementation, the first number and the second number can be independently selected from 1 to 3, wherein the duration of each time window is 2 to 10 seconds.

[0043] (2) Input the inertial signal sequence within the extended data window into a pre-trained motion primitive recognition model. The motion primitive recognition model is used for time series segmentation and classification of the inertial signal sequence. It can be implemented using any one or more of the following model architectures after sufficient training: Temporal convolutional networks: Due to their causal dilated convolutional structure, they can efficiently capture long-term dependencies and are very suitable for fine temporal-point-level classification of continuous action signals.

[0044] Recurrent neural networks based on long short-term memory networks or gated recurrent units: are good at processing sequential data and can model the evolution of actions over time.

[0045] Transformer encoders utilize a self-attention mechanism to effectively measure the correlation between signals at different time points within a sequence, thereby accurately defining action boundaries.

[0046] A hybrid model combining convolutional neural networks and recurrent neural networks: first, convolutional layers are used to extract local spatiotemporal features, and then recurrent layers are used to model temporal relationships.

[0047] The output of this action primitive recognition model is an action category label at each sampling time point or short time slice, thereby converting a continuous signal stream into a discrete sequence of basic action labels with semantic meaning.

[0048] That is, the action primitive recognition model segments and classifies the entire signal into a series of discrete basic action label sequences arranged in chronological order (e.g., "hand still" -> "forearm movement" -> "holding object close to mouth" -> "hand lowers"). In this output label sequence, the specific label that corresponds exactly in time to the action segment located by the first AI model (e.g., "holding object close to mouth") is regarded as the "candidate drinking action" to be verified. The entire label sequence is the context action sequence.

[0049] (3) Input the context action sequence into the second AI model (a model specifically for verification). The second AI model evaluates the degree of consistency between the overall pattern of the sequence and the typical behavior pattern before and after drinking water, and outputs the probability that the current candidate drinking action is a real drinking action.

[0050] If this probability exceeds a preset confidence threshold, the candidate drinking action is ultimately marked as a real drinking action. The specific value of the preset confidence threshold can be determined by balancing the precision and recall of the algorithm on the validation set. As an example, this threshold can be set between 0.7 and 0.95. Those skilled in the art will understand that this threshold can be adjusted according to different tolerances for false positives and false negatives.

[0051] In this invention, the second AI model can adopt an architecture of recurrent neural network (RNN), Transformer encoder or time series feature extraction followed by fully connected classifier, to model the context action sequence composed of discrete action labels and evaluate its overall rationality.

[0052] Method 2: Verification based on multimodal feature fusion (when the data contains inertial and physiological signals) When continuous motion data includes synchronized inertial signal sequences and physiological signal sequences, the identification process, in addition to Method 1, adds cross-validation from physiological signals, resulting in higher confidence. The verification process steps are as follows: (1) Inertial context analysis: Perform the same context verification steps as in method one on the inertial signal sequence corresponding to the candidate drinking action to obtain the probability value based on the authenticity of the action pattern, which is used as the first confidence level P. motion .

[0053] (2) Physiological feature analysis: Within the time period corresponding to the candidate drinking action segment and within a preset physiological verification window (e.g., 1-3 seconds after the action ends), the synchronously acquired physiological signal sequence is analyzed to detect whether there are physiological features strongly correlated with the drinking behavior: Swallowing event feature detection: Analyze vibration signals from myosophy microphones or high-sensitivity accelerometers to identify the presence of unique short-term muscle vibration waveforms or acoustic pulses corresponding to swallowing actions.

[0054] Respiratory regulation feature detection: Analyze pulse wave signals (which can derive respiratory rate) from optical heart rate sensors or respiratory signals from bioimpedance sensors to identify whether a brief non-pathological apnea or deep expiratory pattern occurs after the action ends.

[0055] If at least one of the above strongly correlated physiological features (such as swallowing event features) is detected, then the second confidence level P is increased. physio Set it to the first preset value; otherwise, if no relevant features are detected, then P will be set to the first preset value. physio Set to the second preset value. The first preset value is greater than the second preset value.

[0056] (3) Decision fusion: The final decision is made by combining the results of the inertia analysis and the physiological characteristic analysis. The first confidence level and the second confidence level are fused to obtain the comprehensive confidence level, and the comprehensive confidence level is used to determine whether it is a real drinking action.

[0057] In this invention, the fusion of the first confidence level and the second confidence level can be achieved in any of the following ways: Using AI decision-making models: P physio and P motion As input features, these features are fed into a trained decision AI model, which directly outputs a comprehensive confidence score or a classification decision. The decision AI model can be any model suitable for processing low-dimensional feature inputs and performing binary classification, such as, but not limited to, multilayer perceptrons, support vector machines, gradient boosting trees, or logistic regression classifiers.

[0058] The fusion is performed using pre-defined fusion rules: definite mathematical rules are employed. For example, a weighted average rule can be used: Overall confidence level = α × P motion +β×P physio , where α and β are preset weighting coefficients, and α+β=1.

[0059] The calculated overall confidence level is compared with a preset final judgment threshold. If the overall confidence level exceeds the threshold, the current candidate drinking action is finally marked as a high-confidence real drinking action; otherwise, it is excluded.

[0060] In this invention, the specific values ​​of the first preset value and the second preset value, as well as the preset final judgment threshold, are all configurable parameters. In an exemplary embodiment, the first preset value can be set in the range of 0.7 to 0.95 (e.g., 0.85), the second preset value can be set in the range of 0.05 to 0.3 (e.g., 0.15), and the final judgment threshold can be set in the range of 0.65 to 0.9 (e.g., 0.75). Those skilled in the art will understand that these parameters can be adaptively adjusted according to different requirements for recognition accuracy and recall in actual applications.

[0061] S400, based on the real drinking actions identified within a continuous preset time period, determines the user's drinking behavior trend.

[0062] This step aims to determine the stability and regularity of user drinking behavior through statistical analysis based on the actual drinking actions verified by the S300 over multiple consecutive time periods. The preset time period is the basic time unit for statistical analysis, and its duration can be set according to monitoring needs, such as 1 hour, half a day (e.g., morning / afternoon), or 1 day. Shorter time periods (e.g., 1 hour) can reflect more refined intraday patterns, while longer time periods (e.g., 1 day) are helpful for observing diurnal variations.

[0063] Furthermore, the S400 specifically includes: S401, Count the frequency of actual drinking actions in each preset time period within N consecutive preset time periods.

[0064] Using the preset time period as the basic unit, the cumulative number of occurrences of the actual drinking action in each of the N consecutive such time periods is counted, resulting in a frequency sequence [F1, F2, ..., Fi, ..., FN] arranged in chronological order. Here, Fi represents the frequency of the actual drinking action in the i-th preset time period, i is a value from 1 to N, and N is the size of the initial observation window. Its value should ensure that it can initially reflect the behavioral pattern. As an exemplary and non-limiting implementation, N can be set between 5 and 14 (for example, when the preset time period is days, N=7 represents the initial observation of one week's data).

[0065] S402, calculate the fluctuation value of the frequency over time.

[0066] Calculate the variance (or standard deviation) of the frequency sequence [F1, F2, ..., Fi, ..., FN] above, as a quantitative indicator to measure the volatility of drinking behavior, i.e., the volatility value. The larger the variance, the greater the fluctuation in the number of times of drinking within each time period, and the more irregular the behavior; the smaller the variance, the more stable the number of times of drinking.

[0067] S403, when the fluctuation value is less than the preset stability threshold, the statistical characteristics of the frequency of drinking actions within the N consecutive preset time periods are determined as the drinking behavior trend.

[0068] The calculated variance is compared with a preset stability threshold. This stability threshold is a critical parameter used to determine whether a behavior has formed a stable trend; its specific value can be set based on a large amount of user data or experience. As an example, the threshold can be set to the order of variance corresponding to 10% to 25% of the frequency average.

[0069] If the variance is less than the stability threshold, the user's drinking behavior is considered to have stabilized over the past N time periods. In this case, the statistical characteristics of the frequency sequence (such as its arithmetic mean or the slope of the trend line obtained through linear regression) can be used to determine the user's drinking behavior trend. For example, if the average number of times a user drinks water per day is determined to be 8 times, then "average 8 times / day" is the core trend feature.

[0070] Furthermore, the method also includes: S404, when the fluctuation value is greater than or equal to the preset stability threshold, the observation period is extended to M preset time periods, and the fluctuation value is recalculated and judged based on the frequency of drinking actions within the M preset time periods, until the fluctuation value is less than the preset stability threshold, where M > N.

[0071] If the evaluation finds that the variance is greater than or equal to the stability threshold, it indicates that within the current N observation windows, user drinking behavior fluctuates significantly and has not yet shown a stable pattern. In this case, an adaptive extension mechanism is automatically triggered. Extend the observation window from N time periods to M time periods. Recalculate the variance based on frequency data [F1, F2, ..., Fj, ..., FN] over a longer time span and perform stability assessment again. Fj represents the frequency of actual drinking actions within the j-th preset time period, with j ranging from 1 to M. The value of M should be significantly greater than N to include more historical data to smooth short-term fluctuations. For example, M can be set to 1.5 to 3 times N (e.g., if N=7, then M can be 10 to 21). This process can be iterated until the calculated variance is less than the stability threshold, thereby determining the trend based on the final window data that meets the conditions; or after reaching the maximum extension limit (e.g., continuous observation for 28 days), directly use the window data to determine the trend and mark it as a trend with high volatility.

[0072] S500: Estimate the user's liquid intake based on the characteristic parameters of the actual drinking action.

[0073] This step aims to analyze and quantify the characteristics of real drinking actions verified by the S300, thereby estimating the user's actual fluid intake. The process consists of two core components: single-event estimation and time-based summary evaluation.

[0074] Furthermore, the S500 specifically includes: S501, based on the duration and amplitude characteristics of the actual drinking action, query the pre-established mapping model that associates action characteristics with liquid intake to obtain the estimated intake of a single drinking event.

[0075] For each event marked as a real drinking action, its key feature parameters are extracted, mainly including: Duration: The time elapsed from the start of an action (e.g., the hand begins to rise continuously) to the end (e.g., the hand returns to stillness after being lowered).

[0076] Amplitude characteristics of movement: The amplitude, speed or force of the "drinking" movement is quantified by analyzing inertial signals (such as the peak value of triaxial acceleration, the integral area or the angle change in a specific direction).

[0077] By inputting one or more of the aforementioned feature parameters into a pre-established mapping model, the estimated intake corresponding to the drinking water event can be queried or calculated. The establishment and implementation of the mapping model includes, but is not limited to, the following three implementation methods: Implementation Method 1: Continuous Estimation Model Based on Regression Principle: The relationship between characteristic parameters and intake is modeled as a continuous mathematical function.

[0078] Model building: During the training phase, user drinking action data is collected in a controlled environment (using a water cup of known capacity), and the actual amount consumed is recorded simultaneously. The data is fitted using machine learning algorithms (such as linear regression and support vector regression) to establish a direct mathematical relationship (i.e., regression equation) between action features (such as duration and amplitude integral) and actual intake.

[0079] Application: In the application phase, the characteristic parameters of the new action are substituted into the regression equation to directly calculate a continuous intake estimate.

[0080] Implementation Method 2: Discrete Estimation Model Based on Classification and Table Lookup Principle: First, categorize the actions into a limited number of typical patterns, and then assign a fixed intake baseline value to each pattern.

[0081] Model building: Classifier training: Based on the calibrated data, a classification model (such as a decision tree, Naive Bayes, or a lightweight neural network) is trained to classify drinking actions into several preset drinking categories based on action features, such as "sip", "normal swallowing", or "big gulp".

[0082] Lookup table construction: By statistically analyzing the actual drinking volume of all calibrated data under the same drinking water category, the average value is calculated, and a mapping relationship table between drinking water category and average intake is established.

[0083] Application: For a new action, firstly, based on the feature parameters of the new action, the category to which the new action belongs is determined by a classification model. Then, the lookup table is queried to output the corresponding average intake as an estimate.

[0084] Implementation Method 3: Hybrid Estimation Model Based on Category and Duration Principle: By comprehensively considering the type of movement (which determines the base level) and the specific duration (which determines fine-tuning within the level), a more precise estimate can be made.

[0085] Model building: Similarly, in Method 2, a classification model and a baseline lookup table are established to determine the baseline intake corresponding to each drinking water category.

[0086] By calibrating the data, we analyze the pattern of intake variation with duration within each category, and determine a duration coefficient k for each category. This duration coefficient is a scaling factor that represents the rate of change in intake relative to the baseline amount per unit duration (e.g., per second). For example, it can be determined by the slope of a linear regression of the data for each category.

[0087] Application: Intake is estimated using the formula = Vbase + k × T or Vbase × (1 + k × T). Where Vbase is the baseline intake for the category obtained from a table, k is the duration coefficient corresponding to that category, and T is the actual duration of the action. For example, if Vbase = 50ml and k = 5ml / second for the "normal swallowing" category, then the estimated intake for an action lasting 3 seconds is 50 + 5 × 3 = 65ml.

[0088] S502, based on the estimated intake of the single drinking event, an assessment result is obtained to characterize the user's liquid intake within a preset time period.

[0089] After obtaining the estimated intake from a single drinking event, the data is aggregated and analyzed over a preset time period (which can be consistent with the definition in step S400, such as 1 hour or 1 day) to obtain an assessment result characterizing the user's fluid intake within that time period. This result can be output in one or more of the following formats: Cumulative total intake: The total intake for this period is estimated by summing up all the estimated intakes from each individual intake within the same time period.

[0090] Average single intake: Calculate the average of the estimated intake per single time period, reflecting typical drinking habits.

[0091] Intake frequency and distribution: Combine action frequency to analyze the distribution of drinking behavior over time (e.g., concentrated in the morning or evenly distributed).

[0092] Intake compliance rate: The cumulative total intake is compared with the daily recommended water intake calculated based on user profiles (such as weight and activity level) to obtain the percentage of compliance.

[0093] S600, the drinking behavior trend, the liquid intake and the user's preset profile information are input into the third AI model to generate and output abnormal body state labels and confidence levels.

[0094] This step receives the behavioral quantification results from the preceding steps and performs comprehensive analysis using a third AI model. The aim is to generate a state assessment related to a specific abnormal physical condition (taking kidney load as an example). Specifically, the third AI model determines the basal metabolic load level based on the preset profile information and, combined with the drinking behavior trend and fluid intake, outputs a label for the abnormal physical condition.

[0095] The following examples illustrate the progression in data completeness and evaluation accuracy.

[0096] First Implementation Example: Evaluation Based on Core Behaviors and Static Profiles This embodiment utilizes the most basic and necessary data to complete the status assessment. Input data includes: 1. Core drinking behavior data: Drinking behavior trends: Output from S400, representing statistical characteristics of user regularity (such as average daily drinking frequency and trend stability).

[0097] Liquid intake assessment results: Output from S500, representing quantitative indicators of the user's actual intake (such as average daily total intake, achievement rate).

[0098] 2. User's preset profile information: This includes, but is not limited to, gender, age, occupation type, and average daily working hours. This information is used to determine personalized baseline metabolic needs and serves as a personalized baseline for assessing physical load or abnormal conditions. The user's preset profile information can be obtained through various means, such as active registration and entry by the user in associated applications, synchronization from other devices or health accounts authorized by the user, or inference based on long-term behavioral data by the relevant processing logic in the method of this invention after user confirmation. All information acquisition and processing comply with privacy protection policies and are carried out with the user's informed consent.

[0099] In this embodiment, S600 specifically includes: S601, Personalized Baseline Calculation: The model first determines the user's personalized hydration requirement baseline based on user profile information. This can be achieved through built-in rules, such as applying the formula recommended by medical research: Basal requirement (ml) = weight (kg) × (30-40 ml), and then fine-tuning according to age and gender. Simultaneously, the basal load characteristic is marked as "sedentary occupation" based on occupation type (e.g., "driver," "programmer").

[0100] S602, Deviation Calculation: Calculates the quantitative deviation between the fluid intake assessment results and the personalized water requirement baseline. If the input is the average daily total intake, then Deviation = (Actual intake - Requirement baseline) / Requirement baseline; if the input is the achievement rate, then Deviation = 1 - Achievement rate.

[0101] S603, Trend Integration Analysis: This analyzes the deviation by combining the stability of drinking behavior trends (i.e., the stability assessment results output in S400) and generating corresponding assessment results. For example: If the trend is stable (small variance) and the deviation is consistently negative, it is determined to be "long-term habitual water deficiency".

[0102] If the trend is unstable (large variance) and the deviation is negative, it may be judged as "occasional insufficient water intake".

[0103] S604 combines the magnitude and trend of long-term deviations, maps them to discrete labels of abnormal bodily states, and outputs the confidence level of the label (e.g., 0.0-1.0).

[0104] Long-term deviation refers to the average degree or sustained deviation of a user's fluid intake assessment (from S500) from their personalized hydration baseline over multiple consecutive preset time periods (e.g., the past 7 days, 14 days). It is not a random value on a single day, but rather a statistical characteristic (e.g., mean, median) within a stable trend window determined through the S400 step. It is used to measure the severity of the deficiency. For example, a long-term deviation of -10% indicates "mild deficiency," and -30% indicates "severe deficiency." Assuming the user's daily intake compliance rates over the past 7 days were 65%, 70%, 68%, 40%, 67%, 66%, and 69%, and the average compliance rate after removing the random value (40%) was approximately 67%, then the long-term deviation would be approximately 1 - 0.67 = 0.33 (i.e., a 33% deficiency).

[0105] Trend characteristics refer to the pattern of deviation over time within the observation window. This directly stems from the stability analysis results of S400 on "drinking behavior trends." It mainly includes two categories: 1. Stable trend: Small fluctuations in deviation across time periods (variance less than the threshold) indicate a fixed behavioral pattern.

[0106] 2. Fluctuation trend: Large fluctuations in deviation across different time periods (variance greater than or equal to the threshold) indicate irregular behavior.

[0107] Trend characteristics are used to distinguish between "habitual" and "occasional." Combined with long-term deviation, it can be determined whether it is "long-term habitual insufficiency" (stable trend + negative deviation) or "occasional insufficiency" (fluctuating trend + negative deviation). Assume that the user's daily intake compliance rate over the past 7 days was 65%, 70%, 68%, 40%, 67%, 66%, and 69%. If the data series (after removing 40%) shows relatively small fluctuations, and the variance calculated is less than the stability threshold, it is determined to be a "stable trend."

[0108] Among them, the abnormal physical state label can be, for example: Label A (Baseline State): "Normal Kidney Load". This corresponds to deviations within an acceptable range (e.g., within ±10%) and regular behavior.

[0109] Label B (Attention Status): "Slightly increased renal load." Corresponds to a slight deviation or intermittent negative deviation (e.g., deviation -20% to -30%).

[0110] Label C (Significant Abnormal State): "High Renal Load State". Corresponds to a long-term significantly negative deviation (e.g., deviation less than -30%).

[0111] For example, given a "stable trend" plus a "33% long-term deviation", the third AI model is likely to determine it as "long-term habitual water deficiency", and output a significant abnormal state label (such as "high kidney load").

[0112] Confidence level is calculated based on data continuity (such as the number of effective monitoring days) and trend stability. For example: Confidence level = smoothing function (number of effective monitoring days) × (1 - normalized trend variance). The more stable the trend and the more continuous the data, the higher the confidence level.

[0113] Second Implementation Example: Dynamic and Refined Assessment Incorporating Environmental and Contextual Factors This embodiment, based on the first embodiment, introduces available dynamic external data and contextual information, and uses a model to perform real-time, contextualized calibration of the core assessment, thereby generating personalized status assessment results that better fit the user's actual situation.

[0114] In this embodiment, the input data, based on all the input data from the first embodiment (core drinking behavior data, basic user profile information), is supplemented with the following two types of correction factors: Environmental correction factor: Ambient temperature data (the highest or average temperature of the day obtained through device location services or network API).

[0115] Behavioral context correction factor: Low-autonomy behavior time periods in the user's schedule information (such as continuous event blocks like "in a meeting" or "driving" identified through secure calls to the system calendar interface). In this embodiment, the specific execution flow of S600 is as follows: S610, Initial Assessment: First, as described in the first embodiment, the third AI model performs personalized baseline calculation, deviation analysis and trend integration based on the initial drinking behavior data and user profile, and outputs a basic status label (such as "normal kidney load") and its corresponding basic confidence level.

[0116] S620, Feature Correction: The third AI model dynamically adjusts the key quantitative features used for evaluation based on the newly added correction factor. a. Environmental Demand Adjustment: When the ambient temperature exceeds a preset threshold (e.g., 28°C), the personalized daily water demand baseline is dynamically adjusted upwards based on a preset temperature-water demand relationship (e.g., for every 5°C above the standard temperature (e.g., 22°C), the daily water demand baseline increases by 8%). This relationship can be represented by the following formula: Corrected Baseline = Original Baseline × [1 + k × (Current Temperature - Standard Temperature)], where k is the temperature influence coefficient, representing the percentage increase in water demand for every 1°C above the standard temperature. The value of the coefficient k can be determined by referencing physiological consensus or experimental data; as an example, its value can be selected within the range of 0.03 to 0.05 (i.e., 3% / °C to 5% / °C). Based on this corrected baseline, the environmentally corrected fluid intake deviation is recalculated. This deviation is typically greater than the initial value, and the environmentally corrected fluid intake deviation more closely reflects the actual physiological needs of the day, objectively reflecting the actual degree of dehydration under high temperatures.

[0117] b. Contextual Impact Correction: The third AI model identifies periods of low autonomy behavior and contextualizes the assessment logic. Specifically, when calculating the overall daily behavioral deviation or state assessment value, the negative weight of "not drinking water" records occurring during these objectively constrained periods is reduced (e.g., the weighting coefficient is reduced from 1.0 to 0.2). This produces context-weighted behavioral assessment features.

[0118] S630, Comprehensive Reassessment and Output Generation: The third AI model uses the features corrected by S620 (deviation after environment correction, and behavioral evaluation features after context weighting) combined with user profiles to conduct a new round of comprehensive analysis and generate the final output: Final abnormal physical state label: Decisions are made based on the revised features, and updated state labels are output. For example, the initial label "normal kidney load" may be updated to "potentially increased kidney load" due to environmental revision showing that actual intake is severely insufficient; or the original label may be maintained but the contextual factors may be noted after the influence of involuntary dehydration is excluded due to situational revision.

[0119] Final Confidence Level: The final output confidence level is a positive gain on top of the baseline confidence level. This gain stems from the introduction of objective external data (temperature, schedule), which enhances the corroboration and interpretability of the assessment conclusions. The gain value can be preset based on the data quality of the correction factor, for example: Final Confidence Level = min(1.0, Baseline Confidence Level + c), where c is the gain value preset based on the reliability of the correction factor (e.g., 0.1), and min() is the minimum value operation. This means that conclusions derived through dynamic refinement of the assessment have higher reliability.

[0120] This embodiment introduces environmental and contextual factors through a standardized data interface and defines clear feature correction rules and reassessment processes. This not only significantly improves the personalization and contextual accuracy of physical abnormality assessment, but also enhances the credibility of the results due to a more complete chain of evidence, achieving a technological upgrade from basic judgment to precise insight.

[0121] In this invention, the third AI model can employ any machine learning model suitable for multi-feature fusion and classification. As an exemplary, and not limiting, implementation such a model could be a gradient boosting decision tree model, a random forest, or a deep neural network classifier. This model receives the fused feature vectors and outputs discrete state level labels and their confidence scores.

[0122] S700, when the frequency of the abnormal physical state label in a preset statistical period reaches a preset threshold and the confidence level of the abnormal physical state label meets the preset confidence enhancement condition, an abnormal physical state prompt message is generated and output.

[0123] This step is the core interface between the invention and the user for health intervention. It does not react instantly to a single status tag, but rather intelligently decides whether to trigger a health alert and, if so, what level, based on long-term, high-confidence status patterns. This aims to provide effective reminders while avoiding excessive disruption to the user.

[0124] In this invention, a specific abnormal physical condition label (such as "increased potential kidney burden") appears at a frequency that reaches or exceeds a preset threshold within a continuous preset statistical period. Furthermore, when the confidence level of the output associated with this condition label meets a preset confidence enhancement condition within the statistical period, an abnormal physical condition prompt message is generated and output.

[0125] The preset statistical period is the length of the time window for frequency statistics. To effectively capture persistent abnormal states, this period is typically significantly longer than the basic analysis unit (preset time period) in S400. Example: It can be set to 14 days or 30 days. The preset threshold is the critical value for determining whether a state pattern occurs "frequently." For example, within a 14-day statistical period, if a significant abnormal state label appears for more than 10 days (i.e., frequency > 70%), it is determined that the threshold has been reached.

[0126] In this invention, the preset confidence enhancement condition is set to ensure that the alert for abnormal physical states is based on a definite judgment with gradually increasing confidence, rather than an uncertain result with fluctuations. Specifically, the confidence enhancement condition means that, within a preset statistical period, the trend index calculated from the confidence sequence corresponding to the abnormal physical state label is greater than or equal to zero. As an exemplary implementation, the trend index can be obtained by calculating the slope of the linear regression of the confidence sequence; when the slope is greater than or equal to zero, it is determined that the confidence enhancement condition is met.

[0127] Furthermore, the S700 may specifically include: S710 selects the corresponding prompt level from a predefined multi-level prompt library as the target prompt level based on the level of the abnormal physical state label, the intensity of the state pattern, and the confidence level.

[0128] Among them, the level of the abnormal physical state label is the severity of the abnormal state directly represented by the label output by S600 (e.g., "baseline state", "state of concern", "significant abnormal state").

[0129] The intensity of a state pattern is characterized by the following two quantifiable technical parameters, used to determine the persistence and stability of abnormal bodily states: Number of consecutive occurrences: The preset statistical number of consecutive occurrences of this status label.

[0130] Overall frequency of occurrence: The percentage of days that the status label appears within the preset statistical period (e.g., 24 out of 30 days, with a frequency of 80%).

[0131] Confidence level: Characterized by the following two quantifiable technical parameters, used to determine the reliability and deterministic changes in state judgment: Average confidence level: The arithmetic mean of all confidence level values ​​for this status label within the statistical period.

[0132] Confidence trend: Based on the confidence sequence within the statistical period, the slope calculated by linear regression is used to quantify the direction and magnitude of the change in confidence over time.

[0133] As an exemplary, and not restrictive, hierarchical strategy, the multi-level suggestion library can be defined as follows: Level 1 (Attention Level): The trigger condition is that the overall frequency of the attention status label in the first statistical period exceeds the first threshold (e.g., 50%), and the average confidence level is higher than the basic threshold, e.g., 0.65.

[0134] Level 2 (Recommended): The trigger condition is that any of the following conditions are met: (1) Significantly abnormal state labels appear, and their single confidence level is greater than the first confidence level, for example, 0.7.

[0135] (2) The number of consecutive occurrences of the status label is greater than or equal to the first preset number of periods, for example, 2, and the overall occurrence frequency of the most recent period is greater than the first preset occurrence frequency threshold, for example, 60%, the average confidence level is greater than the second confidence level, for example, 0.75, and the slope of the confidence trend is greater than or equal to 0 (indicating stability or increase).

[0136] Level 3 (Action Level): The triggering conditions are that the number of consecutive occurrences of the significant abnormal status label is greater than or equal to the second preset number of periods, such as 3, and the overall occurrence frequency of the most recent period is greater than the second preset occurrence frequency threshold, such as 80%, the average confidence level is greater than the third confidence level, such as 0.8, and the slope of the confidence trend is greater than 0 (indicating a continuous increase).

[0137] S720 generates and outputs corresponding bodily abnormality alerts based on the determined target alert level. The information at different levels increases progressively in terms of detail, specificity of suggestions, and strength of action guidance. The prompts corresponding to Level 1 (cognitive alerts) mainly include descriptions of the abnormal state and general health advice. The aim is to make users aware of potential state patterns.

[0138] Example: "System analysis has revealed a recent trend of insufficient water intake (22 out of the past 30 days have been marked as abnormal). Please remember to drink water regularly." The prompts for Level 2 (quantitative analysis and behavioral guidance) are based on Level 1, adding personalized quantitative data (such as specific intake gaps) and actionable behavioral improvement suggestions.

[0139] Example: "Data analysis shows that your recent daily fluid intake is approximately 1600 ml, which is about 400 ml lower than your personal recommendation. We suggest you try using a timer reminder and make sure to drink water during work breaks." The corresponding Level 3 alert (enhanced warning and professional referral): Building on Level 2, it emphasizes the persistence, severity, and potential health impact of the abnormal condition, and explicitly recommends seeking professional medical evaluation.

[0140] Example: "The system has continuously monitored that you have a long-term, significant lack of water intake, accompanied by a sedentary lifestyle, which may increase the metabolic burden on organs such as the kidneys. Given that this significant abnormal pattern has become stable, it is strongly recommended that you consult a doctor or nutritionist for a targeted health check-up." When generating prompts, the content can be further contextualized by incorporating the user's pre-defined profile information (especially occupation type) to provide more targeted and actionable suggestions. This process is achieved through a content adaptation rule base, which defines the mapping relationship between different user attributes and suggestion expressions.

[0141] Example: For users whose occupation type is "driver", the prompt can include wording that fits their work scenario, such as "Please take advantage of breaks to drink water in time"; for users whose occupation type is "office worker", it can be suggested that "Place a visible water cup on your desk to remind yourself to drink water regularly".

[0142] When displaying abnormal health alerts, a simple feedback interface can be provided simultaneously, such as buttons with options like "Understood" or "This alert is inaccurate." User-submitted feedback data will be anonymized, recorded, and stored.

[0143] The feedback data can be used for subsequent model optimization, for example: When a certain number of users mark the prompts for the same status label as "inaccurate", a review of the historical data and judgment logic corresponding to that label can be triggered.

[0144] A large amount of positive feedback (“understood”) can be used to verify and reinforce the effectiveness of the current anomaly detection threshold and alert strategy.

[0145] Subject to privacy and ethical considerations, anonymized feedback data can be used as incremental learning data to fine-tune relevant AI models (such as third-party AI models) so that their decisions are more in line with the actual perceptions of the user group.

[0146] In summary, the technical solution provided by this invention achieves significant technological progress by constructing a complete technical closed loop from seamless data collection to intelligent anomaly alerts. Its technical effects are specifically reflected in the following aspects: 1. Accurate behavior recognition and assessment of abnormal physical states based on low-precision universal sensors were achieved. Traditional health monitoring either relies on specialized medical equipment (high cost and highly invasive) or on user-reported data (inaccurate data and low compliance). This invention creatively utilizes inertial sensors widely found in consumer wearable devices (such as wristbands) and, through multi-level AI model collaboration (initial screening by the first AI model and contextual verification by the second AI model), successfully identifies the specific and subtle health-related behavior of drinking water from continuous and complex daily action data streams with high precision. This overcomes the technical bottleneck of low accuracy in single-action recognition, making it possible to conduct seamless, long-term, and refined health behavior monitoring using the most widely available and user-friendly devices.

[0147] 2. A progressive analytical model of "short-term actions → long-term trends → personalized status assessment" has been established, providing a highly scientific early warning system. This invention does not react to isolated events, but rather constructs a multi-layered analysis system: from single action recognition (S200), to behavioral pattern trend analysis (S400), and then to a comprehensive state assessment integrating physiological characteristics (S500) and personal profiles (S600). This progressive analysis framework of "signal -> event -> pattern -> state" can effectively filter out sporadic noise and capture truly health-related long-term unhealthy habits. In particular, by using confidence-enhancing conditions (such as trend indicators ≥ 0) to trigger alerts, it ensures that the warnings are based on increasingly certain and stable abnormal state patterns, rather than instantaneous fluctuations, greatly improving the scientific rigor and seriousness of the alert information and reducing the interference of false alarms on users.

[0148] 3. Provide highly personalized health insights through multimodal information fusion and context awareness. This invention goes beyond simple behavior monitoring. In a preferred embodiment, by fusing inertial data with physiological signals (such as swallowing characteristics), physiological verification of the action is achieved, significantly improving the specificity of the judgment. Furthermore, by integrating user profiles (age, occupation), environmental data (temperature), and even daily schedule information, the third AI model can perform dynamic and contextualized assessments of abnormal states. For example, it can distinguish between increased water demand due to high temperatures and insufficient water intake due to personal habits, thereby outputting more personalized labels and prompts for abnormal physical states that are more in line with the user's actual life situation (such as providing different hydration suggestions for drivers and office workers), achieving a leap from general alerts to personalized health management.

[0149] 4. A preventive health management program that can be optimized in a closed loop has been developed. This invention constructs a complete "monitoring-analysis-evaluation-intervention-feedback" cycle. The output of tiered physical abnormality alerts (from cognitive reminders to medical advice recommendations) forms an effective intervention mechanism. Simultaneously, an optional feedback mechanism provides a data foundation for continuous model optimization. This approach shifts the focus of health management upstream, emphasizing the early identification and alerting of "potential abnormal behavioral patterns," aiming to promote user behavior change. It embodies the core concept of preventative health management and has significant practical value in alleviating the preventative pressures of diseases strongly correlated with long-term lifestyle habits, such as chronic kidney disease.

[0150] This invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in this invention.

[0151] It is understood that the steps in the methods of the above embodiments of the present invention can be implemented by software programs, hardware circuits, or a combination of both in electronic devices with data processing capabilities. The electronic devices include, but are not limited to: wearable devices worn by the user (such as smart bracelets or watches), the user's mobile terminal (such as smartphones or tablets), and remote cloud servers. The software programs are stored in corresponding memory for execution by a processor. In a typical implementation, the wearable device is primarily responsible for performing the data acquisition step (S100) and possible simple preprocessing; while the subsequent AI model recognition, analysis, evaluation, and decision-making steps (S200 to S700) can be completed by the processor in the mobile terminal or cloud server executing the computer program instructions in the memory. The final abnormal physical condition alert is presented via the user interface of the mobile terminal. The devices transmit data and interact with each other via wired or wireless communication, collaboratively completing the entire process of the method described in this invention.

[0152] It should be noted that, since this invention primarily relies on indirect analysis and inference based on sensor data collected by consumer-grade wearable devices, its estimation of individual drinking behavior, intake, and final assessment of abnormal physical states inevitably involves a certain degree of probability and estimation error. Therefore, the "abnormal physical state tags" and "abnormal physical state alerts" generated by this invention are essentially forward-looking health status alerts and behavioral improvement suggestions based on data trends and probabilistic models. Its core purpose is to help users identify potential unhealthy lifestyle trends through long-term, imperceptible monitoring and provide friendly warnings in the early stages of abnormal state accumulation, thereby prompting users to take proactive health management actions or seek professional medical consultation. The output of this invention is not, and should not be construed as, a legally binding medical diagnosis or an absolute disease judgment. Any specific medical decisions should be made by professional medical personnel in conjunction with a more comprehensive clinical examination. The inventive value of this invention lies in providing a low-cost, universally applicable, and effective preventative health management tool, rather than replacing professional medical diagnosis.

[0153] The method for generating abnormal bodily state information based on an AI model disclosed in this invention has clear feasibility and strong industrial applicability, as detailed below: I. Feasibility Description The hardware foundation is mature and readily available: The core data acquisition of this invention relies on the sensors of consumer-grade wearable devices (smart bracelets, watches). These devices have achieved large-scale popularization, and their built-in inertial sensors (three-axis accelerometers, three-axis gyroscopes) and physiological signal sensors (optical heart rate sensors, bioelectrical impedance sensors, etc.) are all existing mature hardware configurations. There is no need to design additional dedicated sensors or modify equipment, and there are no technical obstacles to hardware acquisition and deployment.

[0154] The AI ​​model architecture is feasible and controllable: The three-level AI model adopted in this application (the first AI model is used for candidate drinking action recognition, the second AI model is used for real drinking action verification, and the third AI model is used for abnormal state assessment) is all based on existing mature deep learning and machine learning architectures, including one-dimensional convolutional neural networks (1D-CNN), temporal convolutional networks (TCN), long short-term memory networks (LSTM), gradient boosting decision trees, etc. The action data and physiological signal data required for model training can be obtained through controlled environment calibration and user authorization collection. The model parameters can be flexibly adjusted according to the actual scenario, and the training and deployment process is clear and executable.

[0155] The technical steps are operable and reproducible: Each technical step disclosed in this invention (data acquisition, action recognition, trend analysis, intake estimation, anomaly assessment, and prompt generation) has clear execution logic and quantitative standards. For example, parameters such as the time window setting for action recognition (2-10 seconds), the variance threshold for fluctuation value calculation, and the confidence judgment threshold (0.7-0.95) are all configurable quantitative indicators. The processes such as context action sequence extraction, multimodal feature fusion, and hierarchical prompt triggering are all supported by clear algorithms and can be completely reproduced through computer program code. There is no technical impossibility.

[0156] Strong compatibility and adaptability to diverse scenarios: This application supports two data acquisition modes: "single-modal inertial signal" and "multi-modal inertial + physiological signal", which can be adapted to wearable devices with different accuracy levels (from entry-level wristbands to high-end health watches); at the same time, the acquisition methods of user profile parameters (gender, age, occupation, etc.) (user registration, authorization synchronization, behavior inference) are flexible and diverse, and time parameters such as preset time periods and statistical cycles can be adjusted according to monitoring needs, which can be compatible with diverse application scenarios such as personal health management and occupational health screening.

[0157] II. Practicality Notes Precisely addressing the pain points of existing technologies: This invention addresses the core pain points of existing drinking behavior monitoring, namely "privacy leakage (camera solutions), high false judgment rate (single action recognition), and lack of personalization (general reminders)". Through a technical solution of "non-intrusive data collection + multi-level AI verification + trend evaluation", it achieves high-precision quantification and personalized abnormal warning of daily drinking behavior, fills the technological gap of non-invasive and routine health monitoring, and meets the urgent need for preventive monitoring of chronic diseases in the context of improved public health awareness.

[0158] With a wide range of applications and strong applicability, this invention can be directly integrated into the firmware or accompanying APP of existing wearable devices without requiring users to purchase additional hardware. It is suitable for personal health management of ordinary consumers, especially for high-risk occupational groups such as taxi drivers and office workers who sit for long periods of time and have irregular drinking habits. At the same time, the abnormal body status tags and graded prompts generated by the invention can provide objective behavioral data support for health management platforms and medical institutions, and have dual application scenarios of consumer-grade applications and medical auxiliary applications.

[0159] User-friendly and easy to promote: This invention adopts a seamless monitoring mode throughout the process. Users only need to wear the wearable device normally to complete the data collection without manual recording or triggering operations, which greatly reduces the user compliance threshold. The hierarchical prompt mechanism and contextualized suggestions (combined with occupation and environment adjustments) not only ensure the effectiveness of risk warnings, but also avoid excessive interference, which is in line with users' daily usage habits and has the foundation for large-scale promotion.

[0160] Possesses closed-loop optimization and expansion potential: This invention constructs a complete closed loop of "monitoring-prompt-feedback-model optimization" through a user feedback interface, which can continuously optimize the accuracy of the AI ​​model based on user behavior habits and feedback data, improving its adaptability for long-term use; at the same time, the core technology framework (action recognition → trend analysis → multi-dimensional fusion evaluation) can be transferred to the monitoring of other daily health behaviors such as sedentary posture and eating frequency, and has the potential to be expanded into a one-stop health management tool, with significant industrial application value.

[0161] In summary, the technical solution of this invention is based on existing mature hardware and AI technology, with clear and feasible steps. It can effectively solve practical technical problems, meet diverse application needs, and has sufficient feasibility and high industrial applicability.

[0162] This invention also provides a computer-readable storage medium storing computer-executable instructions for performing the methods described in this invention.

[0163] It should be understood that the various processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein. The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for generating abnormal bodily state information based on an AI model, characterized in that, The method includes the following steps: Collect continuous motion data of users through sensors in wearable devices; Based on the continuous action data, candidate drinking actions are identified using the first AI model; Based on the context sequence of the candidate drinking actions, the actual drinking action is determined by the second AI model; Based on the real drinking actions identified within a continuous preset time period, the user's drinking behavior trend is determined. Based on the characteristic parameters of the actual drinking action, the user's fluid intake is estimated; The drinking behavior trend, the liquid intake, and the user's preset profile information are input into the third AI model to generate and output abnormal body state labels and confidence levels. When the frequency of the abnormal physical state label in a preset statistical period reaches a preset threshold and the confidence level of the abnormal physical state label meets the preset confidence enhancement condition, an abnormal physical state prompt message is generated and output.

2. The method according to claim 1, characterized in that, The step of determining the actual drinking action based on the context sequence of the candidate drinking actions using a second AI model specifically includes: Extract the candidate action types identified within the first number of time windows preceding and the second number of time windows following the occurrence time of the candidate drinking action to form the context action sequence; The context action sequence is input into the second AI model to obtain the probability that the candidate drinking action is judged as a real drinking action; candidate drinking actions with a probability exceeding a set threshold are marked as real drinking actions.

3. The method according to claim 1 or 2, characterized in that, The process of determining the user's drinking behavior trend based on real drinking actions identified within a continuous preset time period specifically includes: Statistically count the frequency of actual drinking actions in each of the N consecutive preset time periods; Calculate the fluctuation value of the frequency over time; When the fluctuation value is less than the preset stability threshold, the statistical characteristics of the frequency of drinking actions within the N consecutive preset time periods are determined as the drinking behavior trend.

4. The method according to claim 3, characterized in that, The method further includes: When the fluctuation value is greater than or equal to the preset stability threshold, the observation period is extended to M preset time periods. The fluctuation value is recalculated and judged based on the frequency of drinking actions within the M preset time periods until the fluctuation value is less than the preset stability threshold, where M > N.

5. The method according to claim 1, characterized in that, The estimation of a user's fluid intake based on characteristic parameters of actual drinking actions is specifically as follows: Based on the duration and amplitude characteristics of the actual drinking action, a pre-established mapping model that associates action characteristics with liquid intake is queried to obtain the estimated intake of a single drinking event. Based on the estimated intake from the single drinking event, an assessment result is obtained to characterize the user's fluid intake over a preset time period.

6. The method according to claim 1, characterized in that, The preset profile information includes at least one of the user's gender, age, occupation type, and average daily working hours; the third AI model determines the basal metabolic load level based on the preset profile information, and outputs the abnormal physical state label by combining the drinking behavior trend and the fluid intake.

7. The method according to claim 1, characterized in that, The preset confidence enhancement condition refers to the fact that, within the preset statistical period, the trend index calculated based on the confidence sequence corresponding to the abnormal physical state label is greater than or equal to zero.

8. The method according to claim 1, characterized in that, The wearable device's sensors include inertial sensors and at least one physiological signal sensor; The process of collecting continuous motion data of the user through the sensors of the wearable device includes collecting inertial signal sequences generated by the inertial sensor and physiological signal sequences generated by the physiological signal sensor. The determination of a genuine drinking action is also based on the analysis of the physiological signal sequence to identify features associated with swallowing events or specific breathing patterns.

9. An electronic device, characterized in that, Including processor and memory; The processor executes the steps of the method as described in any one of claims 1 to 8 by invoking programs or instructions stored in the memory.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a program or instructions that cause a computer to perform the steps of the method as described in any one of claims 1 to 8.

Images