Period information prediction method and device, electronic equipment, and storage medium

Abstract

Embodiments of the present application provide a cycle information prediction method and device, electronic equipment and storage medium. In the method, on the one hand, by obtaining multi-source input data, feature extraction and feature fusion are performed on the multi-source input data, the correlation between different types of input data can be established, and the prediction reliability of cycle information can be improved. On the other hand, by obtaining auxiliary information such as abnormal source information, auxiliary decision information and state transition information through feature fusion, a prediction adjustment factor is determined through the auxiliary information, and then the target prediction model is adaptively adjusted based on the prediction adjustment factor. The model can be adjusted accordingly in the complex scene where the input data is inaccurate or abnormal, accurately matches the current data state, abnormal type and cycle change trend, thereby reducing the cycle prediction error in the complex scene and further improving the prediction reliability of cycle information.

Description

Technical Field

[0001] This application relates to the field of health management technology, and more specifically, to a method and apparatus for predicting periodic information, an electronic device, and a storage medium. Background Technology

[0002] Periodic information prediction is a core technology in the field of time series data analysis, widely applied in scenarios such as menstrual cycle monitoring, time-series work condition prediction, and periodic event extrapolation. In related technologies, periodic information prediction typically relies on physiological data collected by smart wearable devices and manually entered menstrual information from users to construct a periodic prediction model. However, this approach usually assumes that users have stable and regular periods over a long period and relies excessively on manually entered menstrual information. When faced with irregular periods from users or / or anomalies in the manually entered menstrual information, it can easily lead to periodic prediction errors, resulting in poor reliability of periodic information prediction. Therefore, there is an urgent need for a periodic information prediction method that can improve the reliability of periodic information prediction. Summary of the Invention

[0003] Embodiments of this application provide a method, apparatus, electronic device, and storage medium for predicting periodic information, in order to solve the problem of poor reliability in predicting periodic information.

[0004] According to one aspect of the embodiments of this application, a method for predicting periodic information is provided, comprising: acquiring multi-source input data; wherein the multi-source input data includes multimodal physiological detection data and periodic correlation data containing historical periodic data; performing feature extraction and feature fusion on the multi-source input data to obtain a feature fusion representation; acquiring auxiliary information based on the feature fusion representation, the auxiliary information including at least one of anomaly source information, auxiliary decision information, and state transition information; determining a prediction adjustment factor based on the auxiliary information; adaptively adjusting a target prediction model based on the prediction adjustment factor, and performing periodic prediction based on the adaptively adjusted target prediction model to obtain a periodic prediction result.

[0005] In one embodiment of this application, the step of extracting and fusing features from the multi-source input data to obtain a feature fusion representation includes: encoding different types of data in the multi-source input data to obtain multiple sub-feature representations; and performing feature fusion processing on the multiple sub-feature representations to obtain the feature fusion representation.

[0006] In one embodiment of this application, determining the prediction adjustment factor based on the auxiliary information includes: acquiring score thresholds corresponding to the anomaly source information, the auxiliary decision information, and the state transition information, respectively; comparing the anomaly source information, the auxiliary decision information, and the state transition information with the corresponding preset score thresholds to obtain a comparison result; based on the comparison result, determining the representation weights of different types of data in the multi-source input data, determining the attention of the target prediction model to different types of data in the multi-source input data, and determining the confidence level and prediction interval width of the periodic prediction result; and determining the prediction adjustment factor from the representation weights, the attention, and the confidence level and prediction interval width.

[0007] In one embodiment of this application, the step of performing periodic prediction based on the adaptively adjusted target prediction model to obtain a periodic prediction result includes: inputting the feature fusion representation into the adaptively adjusted target prediction model to perform periodic prediction and obtain the periodic prediction result; or, using the auxiliary information as an additional input feature, concatenating it with the feature fusion representation to obtain a target fusion feature; and inputting the target fusion feature into the adaptively adjusted target prediction model to perform periodic prediction and obtain the periodic prediction result.

[0008] In one embodiment of this application, the process of acquiring the auxiliary information is performed by the multi-task auxiliary output layer of the target prediction model. The multi-task auxiliary output layer optimizes parameters in the following way: constructing a target time window; extracting data from the multi-source input data based on the target time window to obtain target data; acquiring prior features based on the target data and generating weakly supervised targets for corresponding tasks based on the prior features; using the prior features as model input and the weakly supervised targets as training targets to perform multi-task joint training on the target prediction model, so as to achieve parameter optimization of the multi-task auxiliary output layer.

[0009] In one embodiment of this application, the prior features include recording behavior rule sub-scores, and the step of obtaining prior features based on the target data includes: extracting recording behavior information of historical periodic data from the target data; calculating a gain influence factor and a decay influence factor to characterize the recording credibility of the historical periodic data based on the recording behavior information; and using the gain influence factor and the decay influence factor as the recording behavior rule sub-scores.

[0010] In one embodiment of this application, the prior features include data quality rule sub-scores, and the step of obtaining prior features based on the target data includes: extracting data quality information of multimodal physiological testing data from the target data; calculating a reliability impact factor based on the data quality information to characterize the data quality of the multimodal physiological testing data; and using the reliability impact factor as the data quality rule sub-scores.

[0011] In one embodiment of this application, the prior features include historical periodic baseline rule sub-scores. The step of obtaining prior features based on the target data includes: extracting periodic feature information of historical periodic data from the target data; calculating a deviation influence factor based on the periodic feature information, wherein the deviation influence factor is used to characterize the degree of deviation of the historical periodic data from historical patterns; and using the deviation influence factor as the historical periodic baseline rule sub-scores.

[0012] In one embodiment of this application, the prior features include state transition rule sub-scores. The step of obtaining prior features based on the target data includes: extracting periodic feature information, user background information, and recorded behavior information from the target data; calculating a probability influencing factor to characterize changes in the user's physiological state based on the periodic feature information, the user background information, and the recorded behavior information; and using the probability influencing factor as the state transition rule sub-scores.

[0013] In one embodiment of this application, the prior features include anomaly source rule sub-scores. Obtaining prior features based on the target data includes: calculating a record behavior rule sub-score, a data quality rule sub-score, and a historical period baseline rule sub-score based on the target data; calculating a classification evidence score to characterize the anomaly source type based on the record behavior rule sub-score, the data quality rule sub-score, and the historical period baseline rule sub-score; and using the classification evidence score as the anomaly source rule sub-score.

[0014] In one embodiment of this application, the weakly supervised objective of the corresponding task includes an anomaly source classification task objective, a decision support task objective, a state transition risk task objective, and a periodic prediction task objective. The prior features include record behavior rule sub-scores, data quality rule sub-scores, historical periodic baseline rule sub-scores, state transition rule sub-scores, anomaly source rule sub-scores, and the target data. Generating the weakly supervised objective of the corresponding task based on the prior features includes: calculating a classification probability label based on the anomaly source rule sub-scores, and using the classification probability label as the anomaly source classification task objective; and calculating the record behavior rule sub-scores... The following steps are performed: A record reliability score is obtained; a reliability score for the multimodal physiological testing data is obtained based on the data quality rule sub-score; a risk score for cycle deviation is obtained based on the historical cycle baseline rule sub-score; the record reliability score, the reliability score for the multimodal physiological testing data, and the risk score for cycle deviation are used as the target of the auxiliary decision-making task; a state transition risk score is obtained based on the state transition rule sub-score, and this state transition risk score is used as the target of the state transition risk task; and the cycle indication information in the target data is used as the target of the cycle prediction task.

[0015] According to one aspect of the embodiments of this application, a device for predicting periodic information is provided, comprising: an acquisition module, a feature fusion module, a first prediction module, an adjustment module, and a second prediction module; wherein, the acquisition module is configured to acquire multi-source input data; wherein, the multi-source input data includes multimodal physiological detection data and periodic correlation data containing historical periodic data; the feature fusion module is configured to perform feature extraction and feature fusion on the multi-source input data to obtain a feature fusion representation; the first prediction module is configured to acquire auxiliary information based on the feature fusion representation, the auxiliary information including at least one of anomaly source information, auxiliary decision information, and state transition information; the adjustment module is configured to determine a prediction adjustment factor based on the auxiliary information; and the second prediction module is configured to adaptively adjust a target prediction model based on the prediction adjustment factor, and perform periodic prediction based on the adaptively adjusted target prediction model to obtain a periodic prediction result.

[0016] According to one aspect of the present application, an electronic device is provided, including: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, cause the electronic device to implement the periodic information prediction method as described above.

[0017] According to one aspect of the embodiments of this application, a computer-readable storage medium is provided, on which computer-readable instructions are stored, which, when executed by a computer's processor, cause the computer to perform the periodic information prediction method as described above.

[0018] In the technical solutions provided by the embodiments of this application, on the one hand, by acquiring multi-source input data and performing feature extraction and feature fusion on the multi-source input data, it is possible to establish the correlation between different types of input data and form a feature fusion representation with global representation capabilities, which facilitates subsequent prediction to obtain more accurate auxiliary information and improves the prediction reliability of periodic information. On the other hand, by acquiring auxiliary information such as anomaly source information, auxiliary decision-making information, and state transition information through feature fusion representation, and using this auxiliary information to determine the prediction adjustment factor, and then adaptively adjusting the target prediction model based on the prediction adjustment factor, the model can make corresponding adjustments in complex scenarios such as inaccurate or abnormal input data, accurately match the current data state, anomaly type, and periodic change trend, thereby reducing the periodic prediction error in complex scenarios and further improving the prediction reliability of periodic information.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0021] Figure 1 This is a schematic diagram of an implementation environment related to the periodic information prediction method illustrated in an exemplary embodiment of this application; Figure 2 This is a flowchart illustrating a method for predicting periodic information in an exemplary embodiment of this application; Figure 3 yes Figure 2 The flowchart of step S240 in the illustrated embodiment, which describes a method for determining a predictive adjustment factor in an exemplary embodiment, is shown. Figure 4 This is a flowchart illustrating an exemplary embodiment of the present application of a method for weakly supervised training of a multi-task auxiliary output layer by constructing prior features; Figure 5 This is a flowchart illustrating a method for generating weakly supervised targets based on prior features, as shown in an exemplary embodiment of this application. Figure 6 This is a network architecture diagram of a target prediction model illustrated in an exemplary embodiment of this application; Figure 7 This is a schematic diagram illustrating the predictive application of periodic information, as shown in an exemplary embodiment of this application.

[0022] Figure 8 This is a schematic diagram illustrating the predictive application of periodic information, as shown in another exemplary embodiment of this application; Figure 9 This is a schematic diagram of the structure of a periodic information prediction device shown in an exemplary embodiment of this application; Figure 10 A schematic diagram of the structure of a computer system suitable for implementing the electronic device of the present application is shown. Detailed Implementation

[0023] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments identical to those described in this application. Rather, they are merely examples of apparatuses and methods identical to some aspects of this application as detailed in the appended claims.

[0024] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented as application programs, in one or more hardware modules or integrated circuits, or in different models and / or processor devices and / or microcontroller devices.

[0025] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0026] It should be noted that "multiple" as mentioned in this application refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0027] Before providing a further detailed description of the embodiments of this application, the nouns and terms used in the embodiments of this application are explained, and the nouns and terms used in the embodiments of this application shall be interpreted as follows: Periodic information refers to time-series data that changes regularly over a period of time. In some embodiments, periodic information in this application may refer to the complete physiological cycle of regular changes in bodily hormones, i.e., menstruation. For example, the cyclical phases of the physiological cycle include menstruation, the follicular phase, ovulation, and the luteal phase. In other embodiments, it may refer to device runtime periodic data, environmental time-series periodic monitoring data, and business cycle flow data, etc.

[0028] Multimodal physiological data refers to physiological data detected by smart wearable devices over a continuous period of time, such as skin temperature data, heart rate data, heart rate variability data, respiratory rate data, and sleep data.

[0029] Periodic correlation data includes historical periodic data, which refers to data such as the length of historical menstrual cycles and the start date of historical menstrual cycles.

[0030] In some embodiments, the cycle-related data also includes menstrual-related event records, symptom input data, and user background data.

[0031] The menstrual-related event record information may include the start date and end date of the menstrual period recorded by the user, as well as the corresponding business date. It may also include information such as the record creation time, record update time, number of record modifications, record source, historical activity level, and signs of concentrated supplementation after a long period of non-recording.

[0032] Symptom input data can include subjective symptoms such as abdominal pain, fatigue, mood-related symptoms, and bloating, as well as the time of their occurrence.

[0033] User background data may include basic attribute information such as age. In other embodiments, user background data may also include status information such as postpartum status, whether hormonal contraception has been discontinued, and whether there have been long-term changes in daily routine, either actively filled in by the user or legally obtained by the business system.

[0034] Ancillary information can be used not only as explanatory output but also as internal decision support information within the model, participating in the generation of periodic prediction results. In some embodiments, the ancillary information includes at least one of anomaly source information, decision support information, and state transition information.

[0035] Anomaly source information refers to the probability used to characterize the scenario to which the input data belongs. In some embodiments, scenarios include regular and predictable scenarios, recorded abnormal scenarios, sensor abnormal scenarios, and scenarios with high physiological variability.

[0036] Decision support information refers to auxiliary scores used to explain the basis of forecasts and participate in the adjustment of cyclical forecasts.

[0037] State transition information refers to the risk score of the current cycle deviating from historical patterns.

[0038] Prediction adjustment factors refer to data used to adaptively adjust the target prediction model. In some embodiments, prediction adjustment factors include one or more of gating weights, attention, and confidence.

[0039] A target prediction model is a deep learning model used to predict periodic information. In some embodiments, the target prediction model consists of a multimodal representation learning module, a multi-task auxiliary output layer, an adaptive prediction adjustment unit, and a periodic prediction head.

[0040] In related technologies, periodic information prediction typically relies on physiological data collected by smart wearable devices and manually entered menstrual information from users to construct a periodic prediction model. However, this approach usually assumes that users have stable and regular periods over a long period and relies excessively on manually entered menstrual information. When faced with irregular periods from users or / or anomalies in the manually entered menstrual information, it can easily lead to periodic prediction errors, resulting in poor reliability of periodic information prediction. Therefore, there is an urgent need for a periodic information prediction method that can improve the reliability of periodic information prediction.

[0041] Based on this, embodiments of this application provide a method, apparatus, electronic device, and storage medium for predicting periodic information, which can solve the problem of poor prediction performance of periodic information. The implementation environment of the periodic information prediction method provided in the embodiments of this application is described below.

[0042] This application provides a prediction scheme for periodic information, which can be applied to, for example... Figure 1 In the implementation environment shown, such as Figure 1 As shown, the prediction environment for periodic information provided in this application includes a terminal 110 and a server 120. The prediction scheme for periodic information provided in this application can be executed by the terminal 110 or the server 120, or it can be executed collaboratively by the terminal 110 and the server 120.

[0043] The terminal 110 and the server 120 can communicate via a network, which can be a wired network or a wireless network. Therefore, the terminal 110 and the server 120 can be directly or indirectly connected via wired or wireless communication. For example, the terminal 110 can be indirectly connected to the server 120 via a wireless access point, or the terminal 110 can be directly connected to the server 120 via the Internet; this application does not impose any limitations on this.

[0044] Terminal 110 may be, but is not limited to, mobile phones, tablets, wearable devices (e.g., watches, bracelets, smart helmets, etc.), in-vehicle devices, smart home devices, augmented reality (AR) / virtual reality (VR) devices, laptops, ultra-mobile personal computers (UMPCs), netbooks, personal digital assistants (PDAs), etc.

[0045] Server 120 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Different services can be deployed on the same server or on different servers. The data storage system can store the data that server 120 needs to process. This data storage system can be set up independently, integrated into server 120, or placed in the cloud or on other servers. It should be noted that... Figure 1 The number of terminals and servers is for illustrative purposes only and does not constitute an actual limitation of this application.

[0046] The following will combine Figure 2 The method for predicting periodic information provided in the embodiments of this application will be described in detail.

[0047] Please see Figure 2 , Figure 2 This is a flowchart illustrating a method for predicting periodic information, as shown in an exemplary embodiment of this application. The method can be executed by a computer device, which can be... Figure 1 The implementation environment shown includes terminal 110 or server 120. Of course, this method can also be applied to other implementation environments and executed by terminals or servers in other implementation environments, and this embodiment does not limit this.

[0048] like Figure 2 As shown, in an exemplary embodiment, the method for predicting periodic information includes at least steps S210 to S250, which are described in detail below: Step S210: Obtain multi-source input data.

[0049] In this embodiment, the multi-source input data includes multimodal physiological detection data and periodic correlation data containing historical periodic data. Multimodal physiological detection data refers to physiological time-series data detected by smart wearable devices over a continuous time period, such as skin temperature data, heart rate data, heart rate variability data, respiratory rate data, and sleep data. Historical periodic data refers to data such as the length of historical menstrual cycles and the start date of historical menstrual cycles. Thus, by using continuous-time multimodal physiological time-series data as the main thread, and combining it with historical observational periodic information and historical periodic baselines for joint analysis, it is possible to utilize continuous-time physiological rhythm patterns to assist in judgment even when there is uncertainty in the periodic boundaries, thereby improving the efficiency of utilizing individual periodic information.

[0050] In other embodiments, the period-related data may also include menstrual-related event records, symptom input data, and user background data.

[0051] The menstrual-related event record information may include the start date and end date of the menstrual period recorded by the user, as well as the corresponding business date. It may also include information such as the record creation time, record update time, number of record modifications, record source, historical activity level, and signs of concentrated supplementation after a long period of non-recording.

[0052] Symptom input data can include subjective symptoms such as abdominal pain, fatigue, mood-related symptoms, and bloating, as well as the time of their occurrence.

[0053] User background data may include basic attribute information such as age. In other embodiments, user background data may also include status information such as postpartum status, whether hormonal contraception has been discontinued, and whether there have been long-term changes in daily routine, either actively filled in by the user or legally obtained by the business system.

[0054] Step S220: Perform feature extraction and feature fusion on the multi-source input data to obtain a feature fusion representation.

[0055] In some embodiments, multi-source input data includes input data from different sources, such as multimodal physiological testing data, historical cycle data, menstrual-related event records, symptom input data, and user background data.

[0056] For example, data from different sources can be encoded, features extracted, and jointly fused using a multimodal representation learning module in the target prediction model to form a feature fusion representation for subsequent anomaly source identification, state transition detection, and periodic prediction.

[0057] In some embodiments of this application, feature extraction and feature fusion of multi-source input data can be performed through the following process to obtain a feature fusion representation, including: Step S221: Encode different types of data in the multi-source input data to obtain multiple sub-feature representations.

[0058] Step S222: Perform feature fusion processing on multiple sub-feature representations to obtain a feature fusion representation.

[0059] For example, the multimodal representation learning module includes a physiological encoding unit, a historical cycle encoding unit, a menstrual cycle record encoding unit, a symptom encoding unit, a user background encoding unit, and a fusion unit. In some embodiments, each encoding unit and fusion unit can be implemented using deep learning networks such as recurrent neural networks, temporal convolutional networks, Transformer networks, attention fusion networks, or combinations thereof.

[0060] In this embodiment, input data from different sources are encoded separately using physiological encoding units, historical cycle encoding units, menstrual record encoding units, symptom encoding units, and user background encoding units to obtain multiple sub-feature representations. Then, a fusion unit performs feature fusion processing on these multiple sub-feature representations to obtain a feature fusion representation. This modular encoding and feature fusion approach, rather than simply concatenating different types of inputs after independent processing, allows for the establishment of correlations between information from different sources while preserving local features of each data source. This enables the target prediction model to accurately capture the shift and anomalies in the current physiological state relative to the individual's historical cycle baseline, significantly improving the model's ability to discriminate and predict dynamic shifts in individual cycles.

[0061] For example, the physiological coding unit is used to encode multimodal physiological detection data from multi-source input data within a preset time window to extract rhythmic change features, trend features, fluctuation features, and phase features related to menstruation. For instance, the preset time window may include a current target cycle observation window and a historical cycle observation window.

[0062] For example, the historical cycle encoding unit is used to encode historical cycle data, such as historical cycle length, cycle volatility, historical multi-cycle regularity, and historical abnormal cycle distribution, in order to extract the user's long-term cycle baseline and long-term stability characteristics.

[0063] For example, the menstrual cycle record encoding unit is used to encode menstrual cycle-related event record information to extract information representations related to record compliance. Specifically, prior features of record anomalies can be generated based on the delay between the record creation time and the business date, the difference between the record update time and the creation time, historical record activity, the proportion of delayed records, and signs of concentrated supplementary recording in the menstrual cycle-related event record information. Then, the menstrual cycle record encoding unit encodes the prior features of record anomalies to obtain information representations related to record compliance.

[0064] For example, the symptom coding unit is used to encode symptom input data to extract symptom distribution features within a period and symptom repetition features within a period.

[0065] For example, the user background encoding unit is used to encode user background data to quantify basic attribute information such as age.

[0066] For example, the fusion unit is used to perform feature fusion on the sub-feature representations output by the above-mentioned coding units to obtain a feature fusion representation. In this embodiment of the application, the fusion unit can learn the periodic correlation between the current target period observation window and the historical observation period, the deviation relationship between the multimodal physiological time series data of the current target period and the historical period baseline, the correlation between the recorded abnormal prior information and the current period state, and the correlation between the symptom input and the dynamic changes of the current period by performing feature fusion on the sub-feature representations.

[0067] Step S230: Obtain auxiliary information based on feature fusion representation. The auxiliary information includes at least one of anomaly source information, auxiliary decision information, and state transition information.

[0068] In this embodiment, the auxiliary information can be used not only as an explanatory output but also as internal model auxiliary decision-making information to participate in the generation of periodic prediction results. In some embodiments, the auxiliary information includes at least one of anomaly source information, auxiliary decision-making information, and state transition information. Anomaly source information refers to the probability of the scenario to which the input data belongs; auxiliary decision-making information refers to the auxiliary score used to explain the prediction basis and participate in the adjustment of periodic prediction; and state transition information refers to the risk score of the current period deviating from historical patterns.

[0069] In some embodiments, the scenarios include regular pattern scenarios, recording anomaly scenarios, sensor anomaly scenarios, and physiologically highly variable scenarios. The probability of a regular pattern scenario characterizes the likelihood that the current periodic sample is relatively consistent with the user's historical pattern and that the multimodal physiological rhythm is relatively stable. The probability of a recording anomaly scenario characterizes the likelihood that the anomaly of the current target periodic sample originates from abnormal recording compliance, self-tracking distortion, or unreliable historical observation period boundaries. The probability of a sensory anomaly scenario characterizes the likelihood that the anomaly of the current periodic sample originates from incomplete, unstable, or low-quality multimodal physiological signals collected by the wearable device. The probability of a physiologically highly variable scenario characterizes the likelihood that the anomaly of the current periodic sample originates from the user's actual large periodic fluctuations or unstable multimodal physiological rhythms. By obtaining anomaly source information to distinguish anomaly types, confusion between different anomaly causes can be reduced, avoiding misjudging real physiological fluctuations as abnormal recording compliance or self-tracking distortion, or misjudging recording anomalies as real periodic anomalies, thereby improving the accuracy of subsequent prediction results.

[0070] In some embodiments, the decision support information includes a record reliability score, a multimodal physiological data reliability score, and a period deviation risk score. The record reliability score characterizes the overall reliability of the user's menstrual records, while the multimodal physiological data reliability score characterizes whether the current multimodal physiological data is suitable for period determination. A higher reliability score indicates that the physiological data is more suitable for period determination, while a lower reliability score indicates that the physiological data may be unsuitable as a primary predictive basis due to missing data, noise, or asynchrony. The period deviation risk score indicates the likelihood that the current period deviates from historical patterns.

[0071] In other embodiments, the auxiliary decision-making information may also include state transition risk scores, suspected merging cycle probabilities, and historical observation cycle boundary confidence scores. Thus, by further introducing historical observation cycle boundary confidence, suspected merging cycle probabilities, and cycle quality scores, the system can reduce the impact of low-confidence historical observation cycles on subsequent anomaly source identification, state transition detection, and cycle prediction, thereby improving the model's fault tolerance to historical input contamination scenarios.

[0072] In some embodiments, state transition information includes a state transition risk score, which characterizes the likelihood of a user transitioning from a long-term stable state to a new physiological state.

[0073] Step S240: Determine the prediction adjustment factor based on auxiliary information.

[0074] In some embodiments, combined with Figure 3 As shown, Figure 3 yes Figure 2The flowchart of step S240 in the illustrated embodiment, which describes a method for determining a predicted regulating factor in an exemplary embodiment, may include at least steps S310 to S340, as detailed below: Step S310: Obtain the score thresholds corresponding to the anomaly source information, auxiliary decision-making information, and state transition information, respectively.

[0075] In this embodiment, the score thresholds include an anomaly source probability threshold, record reliability threshold, suspected merging cycle probability threshold, multimodal physiological testing data reliability threshold, and state transition risk threshold. In some embodiments, the score thresholds can be preset based on user historical cycle big data, general physiological benchmarks for the population, and device factory precision parameters, or they can be adaptively updated with long-term user data iteration; no limitation is imposed here.

[0076] Step S320: Compare the anomaly source information, auxiliary decision-making information, and state transition information with the corresponding preset score thresholds to obtain the comparison results.

[0077] In this embodiment of the application, by comparing the anomaly source information, auxiliary decision-making information, and state transition information with the corresponding preset score thresholds, it is possible to determine whether the auxiliary information such as the anomaly source information, auxiliary decision-making information, and state transition information is too low or too high, thereby facilitating subsequent targeted adjustments based on different situations in the comparison results to obtain a more accurate prediction adjustment factor.

[0078] Step S330: Based on the comparison results, determine the representation weights of different types of data in the multi-source input data, determine the attention of the target prediction model to different types of data in the multi-source input data, and determine the confidence level and prediction interval width of the periodic prediction results.

[0079] In some embodiments, the representation weight refers to the weighting ratio of the sub-feature representations generated after the original data of various types are extracted by the corresponding coding unit in the subsequent feature fusion stage. The higher the weight, the stronger the influence of this type of information on the final fusion result, and the lower the weight, the more the contribution of this type of information is suppressed.

[0080] For example, if the comparison results indicate that the record credibility score in the decision support information is high and the probability of suspected merging cycles is low, then the weight of the historical cycle representation will be increased by a preset amount; if the comparison results indicate that the record credibility score is low and the probability of suspected merging cycles is high, then the weight of the historical cycle representation will be decreased by a preset amount; if the comparison results indicate that the reliability score of the multimodal physiological detection data is high, then the weight of the multimodal physiological detection data representation will be increased by a preset amount; if the comparison results indicate that the probability of sensor anomaly scenarios in the anomaly source information is high, or the reliability score of the multimodal physiological detection data in the decision support information is low, then the weight of the multimodal physiological detection data representation will be decreased by a preset amount; if the comparison results indicate that the state transition risk score in the state transition information is high, then the weight of the historical cycle baseline representation will be decreased by a preset amount, and the weight of the recent trend change representation will be increased by a preset amount.

[0081] In some embodiments, attention refers to the degree of attention a target prediction model pays to the representation of various sub-features during the inference process; the higher the attention coefficient, the more the model focuses on learning that type of data feature during computation, and vice versa.

[0082] For example, if the comparison results indicate that the probability of recording abnormal scenarios in the abnormal source information is high, the attention to the historical record boundary will be reduced by a preset amount; if the comparison results indicate that the state transition risk score in the state transition information is high, the attention to the recent trend change will be increased by a preset amount.

[0083] In some embodiments, confidence level refers to the degree of reliability of the prediction result for this cycle, and prediction interval width refers to the range of fluctuation within the prediction period. A higher confidence level indicates higher accuracy of the prediction result for this cycle, while a lower confidence level indicates lower accuracy.

[0084] For example, if the comparison results indicate that the record confidence score is low, and the reliability score of the multimodal physiological test data is low or the state transition risk score is high, then the confidence level will be reduced by a preset amount, and the prediction interval width will be expanded by a preset range; if the comparison results indicate that the record confidence score is high, and the reliability score of the multimodal physiological test data is high or the state transition risk score is low, then the confidence level will be increased by a preset amount, and the prediction interval width will be narrowed by a preset range.

[0085] In this way, by using information on anomaly sources, auxiliary decision-making information, and state transition information as internal auxiliary information to participate in periodic prediction, the auxiliary information step can be used as an explanatory result output and as adjustment information to achieve model optimization. Based on the multi-dimensional comparison results, the feature representation weights, model attention mechanisms, or prediction output constraints are dynamically adjusted, realizing adaptive parameter correction in the model inference process, effectively reducing prediction bias in complex scenarios, and thus improving the reliability of the target prediction model in periodic prediction.

[0086] Step S340: Determine the prediction adjustment factor from the representation weight, attention, confidence level, and prediction interval width.

[0087] In some embodiments, one or more of the following factors can be used as prediction adjustment factors: representation weight, attention, confidence level, and prediction interval width. This allows for flexible selection of adjustment dimensions based on actual business scenarios, adapting to the periodic prediction needs of different review levels, and reducing computational resource overhead while ensuring the reliability of periodic prediction.

[0088] Step S250: Adaptively adjust the target prediction model based on the prediction adjustment factor, and perform periodic prediction based on the adaptively adjusted target prediction model to obtain the periodic prediction result.

[0089] In this embodiment, adaptive adjustment of the target prediction model based on a prediction adjustment factor enables parameter optimization of the model. This allows the model to dynamically and adaptively adjust itself according to the current quality of multi-source data, abnormal disturbances, and changes in physiological cycle states. This weakens the interference of invalid and distorted data, strengthens the representational role of effective features, and accurately adapts to the dynamic shifts and phased state transitions of the individual cycle. Consequently, when using the adaptively adjusted target prediction model for cycle prediction, more reliable cycle prediction results are obtained.

[0090] In some embodiments, the process of predicting periodic information is performed by the periodic prediction head in the target prediction model, and the output of the periodic prediction head is the periodic prediction result.

[0091] In this embodiment, the cycle prediction result may include prediction cycle information and auxiliary information. The auxiliary information serves as an interpretative result of the prediction cycle information, making the cycle prediction result more reliable. The auxiliary information refers to anomaly source information, decision support information, and state transition information.

[0092] In this embodiment of the application, the terminal devices can be categorized and displayed accordingly: For example, if the terminal device is the backend, it can directly display various intermediate scores in the auxiliary information for model monitoring, result interpretation, abnormal sample analysis, and subsequent model iteration.

[0093] For example, if the terminal device is a client, the auxiliary information can be converted into natural language explanations and displayed to the user.

[0094] For example, when the probability of sensor malfunction is high and the reliability score of multimodal physiological detection data is low, the natural language interpretation of the auxiliary information can be "Insufficient recent wear data, low reliability of this prediction".

[0095] When the probability of recording abnormal scenarios is high and the record credibility score is low, the natural language explanation corresponding to the auxiliary information can be "Recent menstrual records may have been supplemented or are missing; it is recommended to supplement and confirm them."

[0096] When the state transition risk score is high, the natural language interpretation of the auxiliary information can be "the recent cycle pattern is significantly different from the historical pattern, and the prediction range has been widened accordingly".

[0097] When the record confidence score is high and the reliability score of the multimodal physiological detection data is high, the natural language interpretation of the auxiliary information can be "the basis for this prediction is relatively sufficient".

[0098] In some embodiments, periodic prediction is performed based on an adaptively adjusted target prediction model to obtain a periodic prediction result, including: inputting a feature fusion representation into the adaptively adjusted target prediction model to perform periodic prediction and obtain a periodic prediction result.

[0099] In this embodiment, the prediction adjustment factor is used to modify the parameters of the model's internal weights, attention, and output layer constraints. Optimization can be achieved by adjusting only the model parameters, without needing to expand the input vector dimension or adding computational steps for feature concatenation and data concatenation. This improves prediction efficiency while ensuring prediction reliability.

[0100] In other embodiments, periodic prediction is performed based on an adaptively adjusted target prediction model to obtain periodic prediction results. This includes: concatenating auxiliary information as additional input features with a feature fusion representation to obtain target fusion features; and inputting the target fusion features into the adaptively adjusted target prediction model for periodic prediction to obtain periodic prediction results. For example, the auxiliary information includes at least one of anomaly source information, auxiliary decision-making information, and state transition information.

[0101] In this embodiment, based on the adaptive adjustment of the model, the input of the prediction cycle head is further optimized. Auxiliary information, such as three types of high-dimensional structured information including anomaly sources, decision support, and state transitions, is directly combined with the basic fusion representation to form the input. This enables the model to learn the correlation between perturbation factors and cycle changes from the underlying input. Furthermore, by combining input feature optimization with adaptive adjustment of model parameters, the model has a stronger fitting ability in complex scenarios, further reducing prediction errors and improving the reliability of prediction results.

[0102] In other embodiments, the adaptive adjustment of periodic prediction can also be achieved in the following ways: after obtaining auxiliary information based on the feature fusion representation, the auxiliary information is used as an additional input feature and concatenated with the feature fusion representation to obtain the target fusion feature; the target fusion feature is input into the target prediction model that has not undergone adaptive adjustment for periodic prediction to obtain the periodic prediction result.

[0103] In this embodiment, optimization can also be achieved simply by optimizing the input of the prediction period head, without having to perform parameter overwriting or weight modification operations during the inference stage. This avoids the risks of model parameter disorder, gradient shift, and historical state contamination that may be caused by dynamic parameter tuning, making the model run more stably and thus improving the reliability of period information prediction.

[0104] In some embodiments of this application, anomaly source information, auxiliary decision information, and state transition information can be obtained through a multi-task auxiliary output layer in the target prediction model. By inputting the feature fusion representation into the multi-task auxiliary output layer, anomaly source information, auxiliary decision information, and state transition information can be obtained.

[0105] That is, the prediction process of anomaly source information, auxiliary decision-making information, and state transition information is executed by the multi-task auxiliary output layer of the target prediction model, such as... Figure 4 As shown, Figure 4 This is a flowchart illustrating an exemplary embodiment of the present application of a method for weakly supervised training of a multi-task auxiliary output layer by constructing prior features, which may include at least steps S410 to S440, as detailed below: Step S410: Construct the target time window.

[0106] In this embodiment, the target time window includes a current target cycle observation window, a historical continuous time window, and a historical observation cycle window. The current target cycle observation window represents the physiological changes observed up to the current moment in the current cycle, such as changes in skin temperature, heart rate, heart rate variability, respiratory rate, sleep patterns, and symptoms observed up to the current moment in the current cycle. The historical continuous time window represents the user's continuous physiological rhythms over a past period; the historical observation cycle window represents periodic segments obtained from the initial division of the user's menstrual records.

[0107] In some embodiments, the historical observation period window includes at least one real period. For example, if there are no user omissions or additions in the historical period data, the historical observation period window is one real period; if there are user omissions or additions in the historical period data, the historical observation period window may be formed by merging multiple real periods.

[0108] For example, in the implementation of this application, the target time window can be constructed through the following process, including: Step S411: Combine the menstrual start date, current date, and multimodal physiological detection data from the menstrual-related event record information to construct the current target cycle observation window.

[0109] Step S412: Use a preset time range as a continuous historical time window. For example, the preset time range could be the past 60 days, 90 days, or 120 days, etc.

[0110] Step S413: Combine the menstrual start date in the menstrual-related event record information to divide the multimodal physiological test data and form at least one historical observation period window.

[0111] In this embodiment of the application, by constructing the current target period observation window, the historical continuous time window and the historical observation period window, it is convenient to extract the corresponding target data from the multi-source input data to construct prior features.

[0112] Step S420: Extract data from the multi-source input data based on the target time window to obtain the target data.

[0113] In some embodiments, data extraction from multi-source input data based on a target time window to obtain target data includes: extracting multimodal physiological detection data from multi-source input data through the current target period observation window, historical continuous time window, and historical observation period window to obtain target data.

[0114] Step S430: Obtain prior features based on the target data, and generate weakly supervised targets for the corresponding tasks based on the prior features.

[0115] In some embodiments, prior features include record behavior rule sub-scores, data quality rule sub-scores, historical periodic baseline rule sub-scores, state transition rule sub-scores, anomaly source rule sub-scores, and target data.

[0116] Step S440: Use prior features as model input and weakly supervised targets as training targets to perform multi-task joint training on the target prediction model, so as to optimize the parameters of the multi-task auxiliary output layer.

[0117] In this embodiment of the application, by constructing prior features and weakly supervised targets to perform multi-task joint training on the target prediction model, it is possible to optimize the parameters of the multi-task auxiliary output layer.

[0118] In some embodiments, the prior features in step S430 include recording behavior rule sub-scores, which represent whether the user's menstrual records are timely, stable, and reliable. These recording behavior rule sub-scores can be used as model input or to construct weakly supervised objectives during the training phase.

[0119] For example, the record behavior rule sub-score can be obtained by: extracting record behavior information of historical periodic data from the target data; calculating the gain influence factor and the decay influence factor to characterize the record credibility of historical periodic data based on the record behavior information; and using the gain influence factor and the decay influence factor as the record behavior rule sub-score.

[0120] The recorded behavioral information includes the number of days of record delay, the number of record modifications, the number of records added in a concentrated period, the data recording interval offset, the start date of menstruation, and the record creation date. Gain factors include the boundary physiological consistency support score and the historical record activity support score, while attenuation factors include the record delay penalty score, the record modification penalty score, the concentrated addition penalty score, and the adjacent record interval deviation penalty score.

[0121] In some embodiments, the recording delay penalty score is used to indicate whether a user retroactively records their menstrual cycle. The function for calculating the recording delay penalty score can be expressed as:

[0122] in, Used to represent the score for recording delay penalty. Used to indicate the number of days of record delay. This is used to represent a preset delay threshold, such as 14 days. In some embodiments, the record delay days are the difference between the record creation date and the menstrual period start date, that is, record delay days = record creation date - menstrual period start date; where the menstrual period start date is the actual date of menstruation entered by the user, and the record creation date is the date the user creates the record in the application.

[0123] For example, if the menstrual period transaction date is April 1st, the record creation date is April 18th, the record delay days are 17 days, and the record delay days are 14, then by calculation... The record delay penalty score is 1. If the user records on the same day, the record delay days are 0, and the record delay penalty score is 0; if the user records 7 days late, the penalty score is calculated... The record delay penalty score is 0.5. Therefore, the closer the record delay penalty score is to 1, the more likely the record is to be a retroactive entry; the closer it is to 0, the more timely the record is.

[0124] In some embodiments, a record modification penalty score is used to indicate whether a user frequently modifies the same menstrual cycle record. The system can count the number of modifications to the same record. The function for calculating the record modification penalty score can be expressed as:

[0125] in, Used to indicate the penalty score for modifying records. Used to indicate the number of times a record has been modified. Used to indicate a preset threshold for the number of modifications, such as 3 times.

[0126] For example, if the record modification count is 0, then by calculation... A record modification penalty score of 0 indicates that the record has not been modified, and the penalty is low. If the record has been modified once, then the penalty is calculated... A record modification penalty score of 0.33 indicates a minor modification. If the record modification count is 3, then a calculation will be performed. A record modification penalty score of 1 indicates that the record has been modified frequently and the record boundary may be unstable.

[0127] In some embodiments, the centralized supplementary entry penalty score is used to indicate whether a user has supplemented multiple historical menstrual events within a short period of time. The system can count the number of historical menstrual records identified as supplementary entries within a preset creation time window. The function for calculating the centralized supplementary entry penalty score can be expressed as:

[0128] in, Used to represent the penalty score for centralized supplementary enrollment. Used to indicate the number of records added in a centralized manner. Used to indicate the preset threshold for the number of entries to be added, such as 3 entries.

[0129] For example, if a user records their menstrual cycle data for February, March, and April all at once on May 1st, the number of records recorded in this concentrated period will be [number missing]. The value is 3, calculated. The penalty score for centralized data entry is 1; if a user enters a new historical record, the number of records entered in centralized data entry will be increased. The value is 1, calculated The penalty score for centralized supplementary enrollment is 0.33; if there is no centralized supplementary enrollment, the penalty score for centralized supplementary enrollment is 0.

[0130] In some embodiments, the adjacent record interval deviation penalty score is used to indicate whether the interval between two adjacent menstrual period start records deviates significantly from the user's historical cycle baseline. The function for calculating the adjacent record interval deviation penalty score can be expressed as:

[0131] in, Used to represent the penalty score for deviation between adjacent records. Used to indicate the proportion of period length. This is used to indicate a preset deviation from a threshold, for example, 1. That is, when the interval between adjacent records is close to twice the historical cycle average, the score is close to 1. The cycle length ratio refers to the ratio between the interval between adjacent records and the historical cycle average, i.e., cycle length ratio = interval between adjacent records / historical cycle average; while the interval between adjacent records is the difference between the start date of the current menstrual period and the start date of the previous menstrual period, i.e., interval between adjacent records = start date of the current menstrual period - start date of the previous menstrual period.

[0132] For example, if the average historical record period for a user is 30 days, and the interval between adjacent records is 62 days, then the period length ratio = 62 / 30 ≈ 2.07. This can be calculated... A deviation penalty score of 1 for the interval between adjacent records indicates that the observation period is significantly longer than expected, potentially indicating missed records or merged periods. If the user's historical average period is 30 days, and the current interval between adjacent records is 31 days, then the period length ratio = 31 / 30 ≈ 1.03. This can be calculated... The deviation penalty score of 0.33 for the adjacent record interval indicates that the record interval is close to the historical periodic baseline and the deviation penalty is low.

[0133] In other embodiments, the function for calculating the deviation penalty score between adjacent records can also be expressed as:

[0134] in, Used to represent the penalty score for deviation between adjacent records. This is used to indicate the deviation of the preset ratio from a threshold, for example, 1. The preset deviation threshold can be determined based on the historical period standard deviation or statistical results of training data.

[0135] In some embodiments, the boundary physiological consistency support score is used to indicate whether multimodal physiological data near the user's recorded menstrual start date supports the view that the location indeed resembles a cycle boundary. The function for calculating the boundary physiological consistency support score can be expressed as:

[0136] in, Used to represent the boundary physiological consistency support score. , , and As weight, Used to represent skin temperature boundary similarity Used to represent the similarity of heart rate variability boundaries Used to express support for sleep changes Used to indicate symptom support.

[0137] For example, by setting a boundary observation window around the start date of menstruation, and comparing changes in skin temperature, heart rate, sleep patterns, and symptom distribution within the current boundary window with corresponding patterns near the user's historical reliable menstrual cycle boundaries, we can obtain skin temperature boundary similarity, heart rate variability boundary similarity, sleep pattern support, and symptom support. Each sub-score can be calculated using correlation coefficients, cosine similarity, trend consistency rate, normalized distance, or symptom count normalization, and uniformly mapped to the 0-2 interval. Then, a weighted sum is performed to obtain the boundary physiological consistency support score.

[0138] If the skin temperature boundary similarity is 0.8, the heart rate variability boundary similarity is 0.7, the sleep variability support is 0.4, and the symptom support is 1.0, and the weights are... , , and All scores are 0.25. Therefore, the boundary physiological consistency support score obtained by calculating the above function is 0.725, indicating that physiological signals and symptoms near this boundary strongly support the menstrual period boundary. If there are no significant changes in skin temperature, heart rate, or sleep before or after this date, and no supporting symptoms, then all the above sub-scores are low, and the boundary physiological consistency support score is close to 0.

[0139] In some embodiments, the Historical Activity Support Score is used to indicate whether a user consistently and stably records their menstrual cycles. The function for calculating the Historical Activity Support Score can be expressed as:

[0140] in, The score used to represent the activity level of historical records can be defined as the number of times a user completes a record within a preset time after the menstrual period, such as within 1 to 3 days.

[0141] For example, if a user recorded their menstrual cycle within 1-2 days after its start in 5 out of the last 6 cycles in China, then calculations would be performed. A historical activity score of 0.83 indicates a high level of user activity. If a user only records data in one of their last six periods, with the rest missed or added later, the score will be calculated... A historical activity score of 0.17 indicates that the user's historical activity is low.

[0142] In other embodiments, after calculating the aforementioned record behavior rule sub-scores, a record credibility score can be obtained by combining these sub-scores. The record credibility score is used to comprehensively characterize whether a user's menstrual cycle records or historical observation period boundaries are credible. In this embodiment, the record credibility score can be used as a weakly supervised objective for model training. For example, this score can be obtained by combining record delay penalty scores, record modification penalty scores, concentrated supplementary recording penalty scores, adjacent record interval deviation penalty scores, boundary physiological consistency support scores, and historical record activity support scores.

[0143] The function for calculating the record credibility score can be expressed as:

[0144] in, Used to represent the credibility score of a record. Used to represent the score for recording delay penalty. Used to indicate the penalty score for modifying records. Used to represent the penalty score for centralized supplementary enrollment. Used to represent the penalty score for deviation between adjacent records. Used to represent the boundary physiological consistency support score. Used to represent the score supporting historical activity. , , , , and The higher the record credibility score, the more reliable the user record; the lower the record credibility score, the more likely the record is to be affected by supplementary entries, modifications, omissions, or unreliable boundaries.

[0145] In some embodiments, the prior features in step S430 include data quality rule sub-scores, which represent whether the multimodal physiological detection data detected by the smart wearable device is complete, continuous, and stable. These scores can be used as model input or to construct weakly supervised objectives during the training phase.

[0146] For example, the data quality rule sub-score can be obtained by: extracting data quality information of multimodal physiological testing data from the target data; calculating a reliability impact factor to characterize the data quality of multimodal physiological testing data based on the data quality information; and using the reliability impact factor as the data quality rule sub-score.

[0147] The data quality information includes the number of days of effective wear, the number of days of individual effective physiological data for each modality, the number of days of collective effective physiological data for each modality, the total number of days in the observation window, the maximum number of consecutive days with missing data, the number of days with abnormal data, and the intensity of the periodic pattern.

[0148] Reliability influencing factors include effective wear coverage, coverage of each modality, continuous loss penalty score, outlier penalty score, multimodal synchronicity score, and periodic correlation mode intensity score.

[0149] In some embodiments, effective wear coverage is used to represent the proportion of valid data collected by the smart wearable device within the observation window. The function for calculating effective wear coverage can be expressed as:

[0150] in, Used to indicate effective wear coverage.

[0151] For example, if the current observation window is 30 days and the effective wearing period is 24 days, then by calculation... The effective wearing coverage rate was 80%. If the current observation window is 30 days, and the effective wearing days are 8 days, then calculations are performed... The effective wearing coverage rate was 26.7%. The higher the effective wearing rate, the more complete the data; the lower the effective wearing rate, the less complete the data.

[0152] In some embodiments, each modal coverage rate is used to represent the data coverage rate of each type of physiological test data in the multimodal physiological test data within the current observation window. These modal coverage rates include skin temperature coverage, heart rate variability coverage, sleep coverage, etc.

[0153] The function for calculating skin temperature coverage can be expressed as:

[0154] The function for calculating heart rate variability coverage can be expressed as:

[0155] The function for calculating sleep coverage can be expressed as:

[0156] In some embodiments, if the coverage rates of each modality, such as skin temperature coverage, heart rate variability coverage, and sleep coverage, are all high, then the reliability of the physiological test data is considered to be high; if the coverage rates of each modality are low, then the risk of sensor abnormalities is considered to be high, and therefore the reliability of the physiological test data is low.

[0157] In some embodiments, the consecutive missing data penalty score is used to indicate whether there are multiple consecutive days without valid data within the observation window. The function for calculating the consecutive missing data penalty score can be expressed as:

[0158] in, Used to represent the penalty score for consecutive missing items. Used to represent the maximum number of consecutive missing days. Used to indicate a preset consecutive missing threshold, such as 7 days.

[0159] For example, if the current observation window is 30 days, the maximum number of consecutive missing days. The preset consecutive missing threshold is 7 days, and the calculation is performed for a period of 6 days. A consecutive missing values ​​penalty score of approximately 0.86 indicates a relatively severe consecutive missing values ​​situation. The maximum number of consecutive missing days... If it is 2 days, then it is calculated A consecutive missing penalty score of approximately 0.29 indicates a mild consecutive missing condition, and the penalty for consecutive missing is low.

[0160] In other embodiments, in multimodal scenarios, the continuous loss penalty scores for modalities such as skin temperature, heart rate variability, and sleep can be calculated separately, and then the maximum value or weighted average value is taken.

[0161] The function for calculating the penalty score for consecutive missing items can also be expressed as:

[0162] in, Used to represent the penalty score for consecutive missing items. Used to represent the penalty score for consecutive missing skin temperature data. The penalty score for consecutive missing data in heart rate variability data. A penalty score used to represent consecutive missing sleep data.

[0163] The function for calculating the penalty score for consecutive missing items can also be expressed as:

[0164] in, Used to represent the penalty score for consecutive missing items. Used to represent the penalty score for consecutive missing skin temperature data. The penalty score for consecutive missing data in heart rate variability data. Used to represent the penalty score for consecutive missing sleep data. , and For weights.

[0165] In some embodiments, the outlier penalty score is used to indicate whether there are obvious anomalous samples, low-quality data, or noisy data within the observation window. This application implements a method to identify the number of outlier samples based on equipment quality indicators, physiologically reasonable ranges, or statistical outlier detection results.

[0166] For example, the function for calculating the outlier penalty score can be expressed as:

[0167] in, Used to represent the penalty score for outliers.

[0168] For example, the function for calculating the outlier penalty score can also be expressed as:

[0169] in, Used to represent the penalty score for outliers.

[0170] For example, if the current observation window is 30 days, and there are outliers in 6 days, then by calculation... The outlier penalty score is 0.20. If the current observation window is 30 days, and outliers exist in 15 days, then the penalty score is calculated... The outlier penalty score is 0.50.

[0171] For example, in a multimodal scenario, the proportion of outliers in each modality can be calculated separately, and then a weighted sum can be performed to obtain the outlier penalty score. Therefore, the function for calculating the outlier penalty score can also be expressed as:

[0172] in, Used to represent outlier penalty scores. , , and For weights.

[0173] In some embodiments, the multimodal synchronicity score is used to indicate whether multiple modalities have valid data simultaneously within the same or similar time windows.

[0174] For example, the function for calculating the multimodal synchronization score can be expressed as:

[0175] in, Used to represent multimodal synchronicity score, multimodal common valid days refers to the number of days in the same day with at least a number of valid data for the target modality, such as skin temperature, sleep and heart rate variability.

[0176] For example, if the current observation window is 30 days, and 20 days of that window have valid data on skin temperature, sleep, and heart rate variability, then by calculating... The multimodal synchronicity score was approximately 0.67, which can be considered moderately high. If only 5 days of valid data containing skin temperature, sleep, and heart rate variability are available, then calculations will be performed... The multimodal synchronization score is approximately 0.17, which indicates that the multimodal synchronization is relatively low.

[0177] For example, the multimodal synchronicity score can also be obtained by averaging the proportion of effective modes per day within the current observation window.

[0178] For example, if the system focuses on four modalities, such as skin temperature, heart rate variability, respiratory rate, and sleep, and all four modalities are effective on a given day, then the effective modality proportion for that day is 1; if only two modalities are effective, then the effective modality proportion for that day is 0.5. The average of the effective modality proportions for each day is then used to obtain the multimodal synchronicity score.

[0179] In some embodiments, the periodic correlation pattern intensity score is used to indicate whether there are phased changes in the current or historical physiological signals that can be used to determine the periodicity.

[0180] For example, the function for calculating the periodic correlation mode intensity score can be expressed as:

[0181] in, Used to represent the intensity score of periodic correlation patterns This is used to represent a similarity comparison between the current cyclical physiological pattern and historical cyclical physiological patterns. In some embodiments, It can be an algorithm such as correlation coefficient, cosine similarity, trend consistency rate, or normalized distance, and mapped to the interval between 0 and 1.

[0182] For example, if a user typically experiences a decrease in skin temperature, changes in heart rate variability, or sleep fluctuations before and after menstruation in their historical cycles, and similar changes occur in the current cycle, then the current cycle's physiological pattern is similar to the historical cycle pattern, and the cycle-related pattern intensity score is high.

[0183] For example, in the embodiments of this application, the intensity of each modality's periodic mode can be calculated separately, and then weighted and combined to obtain the intensity of the periodically correlated mode. Therefore, the function for calculating the intensity score of the periodically correlated mode can also be expressed as:

[0184] in, The intensity score of the periodic correlation mode is used to represent... , , and For weights.

[0185] For example, the skin temperature cycle pattern intensity is 0.8, the heart rate variability cycle pattern intensity is 0.7, the sleep cycle pattern intensity is 0.5, and the respiratory rate cycle pattern intensity is 0.4, with weights... , , and If both are 0.25, then the calculated intensity score of the periodic correlation mode is 0.60.

[0186] In other embodiments, after calculating the aforementioned data quality rule sub-scores, the reliability score of the multimodal physiological testing data can be obtained by combining the data quality rule sub-scores. This reliability score is used to comprehensively characterize whether the current multimodal physiological testing data is sufficiently complete, continuous, and stable, and has a pattern that can be used for periodic determination. In the embodiments of this application, the reliability score of the multimodal physiological testing data can be used as a weakly supervised target for model training.

[0187] For example, the function for calculating the reliability score of multimodal physiological testing data can be expressed as:

[0188] in, Used to represent the reliability score of multimodal physiological test data Used to indicate effective wear coverage. Used to represent multimodal synchronization scores Used to represent the intensity score of periodic correlation patterns Used to represent the penalty score for consecutive missing items. Used to represent outlier penalty scores. , , , , , , and The weights are determined by the reliability score. A higher reliability score indicates that the physiological data is more suitable for periodic determination; a lower reliability score indicates that the current physiological signal may be unsuitable as the main basis for prediction due to missing, noise, or asynchrony.

[0189] In some embodiments, the prior features in step S430 include a historical cycle baseline rule sub-score, which indicates whether the user's past cycles are stable and whether the current cycle deviates from historical patterns. In this embodiment, the historical cycle regularity score, cycle length deviation score, and recent trend change score can be calculated using the lengths of the most recent N historical cycles, the number of days the current cycle has lasted, and the trends of recent cycles. N can be 3, 5, 6, or other preset numbers, or can be adaptively determined based on the integrity of the user's historical data.

[0190] For example, the historical cycle baseline rule sub-score can be obtained by: extracting the cycle feature information of the historical cycle data from the target data; calculating the deviation impact factor based on the cycle feature information, the deviation impact factor being used to characterize the degree of deviation of the historical cycle data from the historical pattern; and using the deviation impact factor as the historical cycle baseline rule sub-score.

[0191] The periodic characteristic information includes the historical period standard deviation, the number of days the current period has lasted, the historical period mean, the length of the most recent period, and the length of the earliest period among the most recent N periods.

[0192] Deviation influencing factors include historical cycle regularity score, cycle length deviation score, and recent trend change score.

[0193] In some embodiments, the historical periodicity score is used to indicate whether a user's performance has been stable over multiple past periods. In this application embodiment, the historical periodicity score can be calculated using the standard deviation or coefficient of variation of the lengths of the most recent N historical periods.

[0194] For example, the function for calculating the historical periodicity score can be expressed as:

[0195] The smaller the standard deviation of the historical period, the more stable the user's historical period and the higher the historical period regularity score; the larger the standard deviation of the historical period, the greater the fluctuation of the user's historical period and the lower the historical period regularity score.

[0196] For example, taking menstrual cycles as an example, if a user's most recent cycles are 29, 30, 30, 31, and 30 days, the score is higher; if a user's most recent cycles are 25, 38, 31, 47, and 29 days, the score is lower.

[0197] In some embodiments, the cycle length deviation score is used to indicate whether the number of days the current cycle has lasted significantly deviates from the user's historical cycle average.

[0198] For example, the function for calculating the deviation of the period length from the fraction can be expressed as:

[0199] The preset deviation threshold can be determined based on the statistical results of training data, product strategy, or the historical fluctuation range of users.

[0200] For example, if a user's historical cycle average is 30 days and the current cycle has lasted for 42 days, then the cycle length deviation score is too high; if a user's historical cycle usually fluctuates between 35 and 45 days and the current cycle has lasted for 42 days, then the score should not be too high.

[0201] In some embodiments, the recent trend change score is used to indicate whether there has been a trend of continuous lengthening, continuous shortening, or increased volatility in recent periods.

[0202] For example, the function for calculating the score of recent trend change can be expressed as:

[0203] For example, if the cycle lengths of the last 5 cycles are 29, 30, 31, 36, and 42 days, the recent trend change score is higher; if the recent cycles have consistently remained between 29 and 31 days, the score is lower.

[0204] In other embodiments, the recent trend change score can also be calculated by the trend slope of the most recent multiple period lengths or the degree of increase in recent period fluctuations.

[0205] In some embodiments, the prior features in step S430 include state transition rule sub-scores, which represent the probability that a user will transition from a historical stable state to a new physiological state stage. These scores are not directly equivalent to the final state transition risk, but rather serve as rule-based prior features input to the model, and can be combined during the training phase to generate weakly supervised state transition risk scores.

[0206] For example, the state transition rule sub-score can be obtained by: extracting periodic feature information, user background information and recorded behavior information from the target data; calculating a probability influencing factor to characterize changes in the user's physiological state based on the periodic feature information, user background information and recorded behavior information; and using the probability influencing factor as the state transition rule sub-score.

[0207] Among them, the potential influencing factors used to characterize changes in a user's physiological state include physiological characterization deviation score, age or state background support score, and state transition inhibition factor.

[0208] In some embodiments, the physiological representation deviation score is used to characterize the degree of deviation of the current target period from the user's historical long-term baseline. In this application embodiment, the physiological representation deviation score can be calculated jointly by period length deviation, historical regularity, recent trend changes, and multimodal physiological pattern deviation.

[0209] For example, the function for calculating the deviation score of physiological representation can be expressed as:

[0210] in, , and The historical regularity score is used as a weight. In this embodiment, the historical regularity score is not used as sole evidence of deviation, but as a moderating factor. The more regular the user's historical period, the more likely the current deviation is to signify a state transition; however, when the user's history itself is highly irregular, even if the current period is long, it should not be easily judged as a state transition.

[0211] Multimodal physiological pattern deviation score is used to indicate whether physiological patterns such as skin temperature, heart rate variability, and sleep deviate from historical cyclical patterns in the current cycle. The function for calculating the multimodal physiological pattern deviation score can be expressed as: Multimodal physiological pattern deviation score = 1 - similarity between current cycle physiological representation and historical cycle baseline representation The similarity can be calculated based on features such as skin temperature trend, heart rate variability trend, sleep rhythm change, and respiratory rate change, or it can be calculated based on the physiological representation vector obtained by model encoding.

[0212] For example, if skin temperature, heart rate variability, and sleep rhythm consistently deviate from historical cycle patterns during the current cycle, the sub-score is higher. If the current cycle's physiological pattern is similar to historical cycles, the sub-score is lower.

[0213] In some embodiments, the age or status background support score is used to characterize whether the user's background information supports the determination of a status transition. In this embodiment, the score can be generated based on the user's age, the postpartum status actively filled in by the user, the status of discontinuing hormonal contraception, long-term changes in daily routines, or other legally obtained background status information.

[0214] For example, the function for calculating age or status background support scores can be expressed as: Age or status background support score = max(age support sub-score, postpartum status support sub-score, discontinuation of hormonal contraception support sub-score, changes in life status support sub-score) For example, the function for calculating age or status context-supported scores can also be expressed as: Age or status background support score = u1 × age support score + u2 × postpartum status support score + u3 × discontinuation of hormonal contraception support score + u4 × life status change support score.

[0215] Where u1, u2, u3, and u4 are weights, the support sub-scores in the above examples can be calculated using rule-based assignment. For example: if the user's age falls within a preset perimenopausal age range, a higher age support sub-score is assigned, such as 0.8; if the user actively fills in their postpartum recovery status, a higher postpartum status support sub-score is assigned, such as 0.7; if the user actively fills in their discontinuation of hormonal contraception, a higher discontinuation of hormonal contraception support sub-score is assigned, such as 0.6; if the user has recently experienced long-term changes in their daily routine, a higher lifestyle change support sub-score is assigned, such as 0.7.

[0216] For example, the score can be increased when a user is in the perimenopausal age group and their cycle has recently started to become longer or more irregular; the score can also be increased when a user actively reports their postpartum recovery status and their cycle is reappearing or gradually returning to normal.

[0217] In this embodiment, age or status background support score serves only as supplementary evidence and does not independently determine the risk of status transition. If only age or background factors exist, but the current cycle length and multimodal physiological testing data do not show significant deviations, the risk of status transition will not significantly increase solely due to the background score.

[0218] In some embodiments, a state transition suppression factor is used to prevent recording anomalies or insufficient sensor data quality from being misjudged as genuine state transitions. This factor may include a recording anomaly suppression factor and a sensor anomaly suppression factor.

[0219] For example, the function for calculating the state transition inhibition factor can be expressed as: State transition inhibition factor = max(recording anomaly inhibition factor, sensing anomaly inhibition factor) For example, the function for calculating the state transition inhibition factor can also be expressed as: State transition inhibition factor = c1 × Recording anomaly inhibition factor + c2 × Sensing anomaly inhibition factor Here, c1 and c2 are weights. A higher state transition inhibition factor indicates that the current cycle deviation is more likely to be explained by recording anomalies or sensing anomalies, and therefore the risk of state transition should be reduced accordingly.

[0220] The sensor anomaly suppression factor is used to indicate whether deviations in current physiological signals may be caused by insufficient device wear, missing data, or poor sensor quality. The function for calculating the sensor anomaly suppression factor can be expressed as:

[0221] in, Used to represent the reliability score of multimodal physiological test data Used to indicate effective wear coverage. Used to represent multimodal synchronization scores Used to represent the penalty score for consecutive missing items. Used to represent outlier penalty scores, b1, b2, b3, b4, b5, b6, and b7 are weights.

[0222] The record anomaly suppression factor is used to indicate whether the current period deviation may be caused by missed entries, supplementary entries, modified records, or suspected period merging. The function for calculating the record anomaly suppression factor can be expressed as:

[0223] in, Used to represent the credibility score of a record. Used to represent the score for recording delay penalty. Used to represent the penalty score for centralized supplementary enrollment. Used to indicate the penalty score for modifying records. The penalty score for deviation between adjacent records is represented by a1, a2, a3, a4, a5, and a6, which are weights. The probability of a suspected merged cycle is used to characterize whether a given historical observation cycle could be formed by merging multiple real cycles. This probability can be calculated from the historical observation cycle quality rule sub-scores. The historical observation cycle quality rule sub-scores are used to determine the reliability of historical observation cycles segmented from user records. In some cases of social division, the historical observation cycle quality rule sub-scores include long-cycle proportion scores, integer multiple proximity scores, intermediate suspected boundary scores, intermediate symptom support scores, and record anomaly correlation scores.

[0224] In some embodiments, the function for calculating the probability of a suspected merger cycle can be expressed as:

[0225] Among them, h1, h2, h3, h4, and h5 are weights. The long-cycle ratio score is used to determine whether a certain historical observation cycle is significantly longer than the user's historical cycle average. The integer multiple close score is used to determine whether the length of a certain observation cycle is close to a preset integer multiple of the historical cycle average, such as 2 times or 3 times. The intermediate suspected boundary score is used to determine whether there are unrecorded menstrual boundaries within a longer observation cycle. The intermediate symptom support score is used to determine whether there is a cluster of menstrual-related symptoms within a longer historical observation cycle. The record anomaly support score is used to represent this.

[0226] In some embodiments, the long-period ratio score can be calculated by segmented scoring or by linear normalization. For example, if the historical average period is 30 days and the length of a certain historical observation period is 60 days, then the length of that historical observation period is approximately twice the historical average, and the long-period ratio score is higher.

[0227] In some embodiments of this application, integer multiples of the approximate score can be calculated by rule-based scoring or by using linear normalization.

[0228] For example, a function that calculates integer multiples of fractions can be expressed as: The nearest integer multiple of the fraction = 1 - min(|period length ratio - nearest integer multiple| / tolerance range, 1) For example, if the historical average period is 30 days, the observation period is 60 days, the period length ratio is 2.0, the most recent integer multiple is 2, and the difference between the period length ratio and the most recent integer is 0, then by calculation, the integer multiple is close to a fraction of 1. It can be understood that since the observation period length is 60 days, which is twice the historical average of 30 days, this suggests that the observation period may be formed by the merging of two actual periods.

[0229] In some embodiments, the function for calculating intermediate suspected boundary scores can be expressed as: The intermediate suspected boundary score = e1 × skin temperature boundary similarity 1 + e2 × heart rate variability boundary similarity + e3 × sleep change similarity + e4 × symptom support score. Here, e1, e2, e3, and e4 are weights.

[0230] It can be seen that the intermediate suspected border score is affected by skin temperature, heart rate variability, sleep, and symptoms. If there are no physiological changes and symptoms during the observation period, the intermediate suspected border score is low; if there are physiological changes and / or symptoms during the observation period, the intermediate suspected border score is high.

[0231] In some embodiments, the function for calculating intermediate symptom support scores can be expressed as: R_middle_symptom = min(number of menstrual-related symptoms in the middle candidate window / T_symptom, 1) R_middle_symptom represents the middle symptom support score, and T_symptom represents the preset threshold for the number of symptoms, for example, 3.

[0232] For example, if the intermediate candidate window is days 29-31, and the user records symptoms such as abdominal pain, fatigue, and mood swings, then the number of menstrual-related symptoms in the intermediate candidate window is 3, and the intermediate symptom support score calculated by R_middle_symptom = min(3 / 3, 1) is 1. If the number of menstrual-related symptoms in the intermediate candidate window is 1, then the intermediate symptom support score calculated by R_middle_symptom = min(1 / 3, 1) is approximately 0.33. If the number of menstrual-related symptoms in the intermediate candidate window is 0, that is, no menstrual-related symptoms are recorded, then the intermediate symptom support score calculated by R_middle_symptom = min(0 / 3, 1) is 0.

[0233] In other embodiments, if the user records the severity of menstrual-related symptoms within the intermediate candidate window, the function for calculating the intermediate symptom support score can also be expressed as: R_middle_symptom = min(weighted sum of symptom intensity / T_severity, 1) Where T_severity is used to represent the preset threshold, for example, 6.

[0234] For example, if a user records menstrual-related symptoms such as abdominal pain, fatigue, and mood swings in the middle candidate window, and the severity of abdominal pain is 3, the severity of fatigue is 2, and the severity of mood swings is 1, then the weighted sum of symptom intensity is 6. The middle symptom support score is 1 by calculating R_middle_symptom = min(6 / 6, 1). Thus, by considering the severity of recorded symptoms in this embodiment, situations with few symptom types but high intensity can be reflected, thereby improving the accuracy of subsequent predictions.

[0235] In other embodiments, after calculating the aforementioned state transition rule sub-scores, a state transition risk score can be obtained by combining the sub-scores. This state transition risk score represents the probability that a user will transition from a stable state to a new physiological state. In this embodiment, the state transition risk score can be used as a weakly supervised objective for model training.

[0236] For example, the function for calculating the state transition risk score can be expressed as:

[0237] in, The state transition risk score is represented by p1, p2, and p3, which are weights. The clip is used to restrict the result to the interval [0,1].

[0238] In some embodiments, the prior features in step S430 further include anomaly source rule sub-scores, which are used to summarize the rule sub-scores obtained in the preprocessing stage into evidence strengths for different anomaly sources. These rule sub-scores can be used as prior features for model input, or they can be used in the training stage to construct soft probability labels for the anomaly source discrimination task. In this embodiment, the anomaly source evidence score includes regularity evidence score, recorded anomaly evidence score, sensor anomaly evidence score, and physiological high variability evidence score.

[0239] For example, the anomaly source rule sub-score can be obtained by: calculating the record behavior rule sub-score, data quality rule sub-score, and historical period baseline rule sub-score based on the target data; calculating the classification evidence score to characterize the anomaly source type based on the record behavior rule sub-score, data quality rule sub-score, and historical period baseline rule sub-score respectively; and using the classification evidence score as the anomaly source rule sub-score.

[0240] The classification evidence score is the sub-score of the abnormal source rule, which includes the normal regularity evidence score, the record abnormality evidence score, the sensor abnormality evidence score, and the physiological high variability evidence score.

[0241] In some embodiments, the record anomaly evidence score is used to indicate whether the anomaly of the current sample is more likely to originate from missing, supplemented, modified, or unreliable anomalies at the boundaries of historical observation periods. In this embodiment, the record anomaly evidence score can be calculated based on the record behavior rule sub-score. For example, the function for calculating the record anomaly evidence score can be expressed as: E_record = r1 × D_delay + r2 × D_backfill + r3 × D_modify + r4 × D_interval + r5 × Probability of suspected merger cycle + r6 × (1 - C_boundary) + r7 × (1 - C_activity) Where E_record represents the record abnormal evidence score, D_delay is the record delay penalty score, D_backfill is the centralized supplementary recording penalty score, D_modify is the record modification penalty score, D_interval is the adjacent record interval deviation penalty score, C_boundary is the boundary physiological consistency support score, C_activity is the historical record activity support score, and r1, r2, r3, r4, r5, r6 and r7 are weights.

[0242] In some embodiments, the sensor anomaly evidence score is used to indicate whether the anomaly of the current sample is more likely to originate from insufficient device usage, missing data, poor sensor quality, or indistinguishable multimodal signals. In this embodiment, the sensor anomaly evidence score can be calculated using data quality rule sub-scores. For example, the function for calculating the sensor anomaly evidence score can be expressed as: E_sensor = s1 × (1 - Ft) + s2 × (1 - skin temperature coverage) + s3 × (1 - Heart Rate Variability Coverage) + s4 × (1 - sleep coverage) + s5 × D_missing + s6 × D_outlier + s7 × (1 - C_sync) + s8 × (1 - C_pattern) Wherein, E_sensor is used to represent the sensor abnormality evidence score, Ft is used to represent the effective wear coverage, D_missing is used to represent the continuous missing penalty score, D_outlier is used to represent the outlier penalty score, C_sync is used to represent the multimodal synchronicity score, C_pattern is used to represent the periodic correlation pattern intensity score, and s1, s2, s3, s4, s5, s6, s7 and s8 are weights.

[0243] In some embodiments, the physiological high variability evidence score is used to indicate whether the anomaly in the current sample is more likely to originate from the user's actual cyclical fluctuations or multimodal physiological rhythm changes, rather than recorded or sensory anomalies. In this embodiment, the physiological high variability evidence score can be calculated using the historical cyclic baseline rule sub-score. For example, the function for calculating the physiological high variability evidence score can be expressed as: E_var = v1 × Period length deviation fraction + v2 × deviation score of physiological representation + v3 × Recent Trend Change Score + v4 × C_pattern + v5 × clip_1 + v6 × clip_2 - v7 × E_record - v8 × E_sensor Wherein, E_var is used to represent the physiological high variability evidence score, C_pattern is used to represent the periodic correlation pattern strength score, clip_1 is used to represent the recording reliability score, clip_2 is used to represent the reliability score of multimodal physiological detection data, E_record is used to represent the recording abnormal evidence score, E_sensor is used to represent the sensing abnormal evidence score, and v1, v2, v3, v4, v5, v6, v7 and v8 are weights.

[0244] In some embodiments, the regularity evidence score is used to indicate whether the current sample conforms to the user's historical pattern and that there are no obvious recording anomalies, sensory anomalies, or physiological high variability. For example, the function for calculating the regularity evidence score can be expressed as: E_normal = n1 × clip_1 + n2 × clip_2 + n3 × C_pattern + n4 × C_boundary - n5 × period length deviation fraction - n6 × E_record - n7 × E_sensor - n8 × E_var Wherein, E_normal represents the evidence score for regularity, C_pattern represents the strength score for periodic correlation patterns, C_boundary represents the support score for boundary physiological consistency, clip_1 represents the record credibility score, clip_2 represents the reliability score for multimodal physiological detection data, E_record represents the evidence score for record anomalies, E_sensor represents the evidence score for sensor anomalies, E_var represents the evidence score for high physiological variability, and n1, n2, n3, n4, n5, n6, n7, and n8 are weights.

[0245] In some embodiments, after obtaining prior features, the multi-dimensional prior features can be weighted and fused, and then a weakly supervised target for model training can be generated based on the fusion result.

[0246] The weakly supervised objectives for the corresponding tasks include the anomaly source classification task objective, the decision support task objective, the state transition risk task objective, and the periodic prediction task objective. The prior features include the record behavior rule sub-score, the data quality rule sub-score, the historical periodic baseline rule sub-score, the state transition rule sub-score, the anomaly source rule sub-score, and the target data.

[0247] Please continue reading. Figure 5 , Figure 5 This is a flowchart illustrating an exemplary embodiment of the present application of a method for generating a weakly supervised target based on prior features, which may include at least steps S510 to S540, detailed below: Step S510: Calculate the classification probability label based on the sub-score of the anomaly source rule, and use the classification probability label as the target of the anomaly source classification task.

[0248] In some embodiments, the anomaly source rule sub-scores include regularity evidence scores, record anomaly evidence scores, sensor anomaly evidence scores, and physiological high variability evidence scores. A classification probability label is obtained by combining these anomaly source rule sub-scores. This classification probability label is then used as a weakly supervised objective for model training.

[0249] For example, the function for calculating the classification probability label can be expressed as: P_normal_target = E_normal / (E_normal + E_record + E_sensor + E_var) P_record_target = E_record / (E_normal + E_record + E_sensor + E_var) P_sensor_target = E_sensor / (E_normal + E_record + E_sensor + E_var) P_var_target = E_var / (E_normal + E_record + E_sensor + E_var) Wherein, P_normal_target is used to represent the classification probability label of the normal regularity type, P_record_target is used to represent the classification probability label of the record abnormality type, P_sensor_target is used to represent the classification probability label of the sensor abnormality type, P_var_target is used to represent the classification probability label of the physiological high variability type, E_normal is used to represent the evidence score of the normal regularity type, E_record is used to represent the evidence score of the record abnormality type, E_sensor is used to represent the evidence score of the sensor abnormality type, and E_var is used to represent the evidence score of the physiological high variability type.

[0250] Step S520: Calculate the record credibility score based on the record behavior rule sub-score; calculate the reliability score of the multimodal physiological detection data based on the data quality rule sub-score; and calculate the risk score of cycle deviation based on the historical cycle baseline rule sub-score; and use the record credibility score, the reliability score of the multimodal physiological detection data, and the risk score of cycle deviation as auxiliary decision-making task objectives.

[0251] In some embodiments, the record behavior rule sub-score includes a gain influence factor and a decay influence factor. The gain influence factor includes a boundary physiological consistency support score and a historical record activity support score. The decay influence factor includes a record delay penalty score, a record modification penalty score, a concentrated supplementary recording penalty score, and an adjacent record interval deviation penalty score. In this embodiment, the record credibility score is used as a weakly supervised objective for model training.

[0252] For example, the function for calculating the record credibility score can be expressed as:

[0253] in, Used to represent the credibility score of a record. Used to represent the score for recording delay penalty. Used to indicate the penalty score for modifying records. Used to represent the penalty score for centralized supplementary enrollment. Used to represent the penalty score for deviation between adjacent records. Used to represent the boundary physiological consistency support score. Used to represent the score supporting historical activity. , , , , and For weights.

[0254] In some embodiments, the data quality rule sub-scores include effective wear coverage, coverage of each modality, consecutive missing value penalty score, outlier penalty score, multimodal synchronicity score, and periodic correlation pattern strength score. In this embodiment, the reliability score of the multimodal physiological testing data can be obtained by combining the various data quality rule sub-scores. This reliability score of the multimodal physiological testing data is then used as a weakly supervised objective for model training.

[0255] For example, the function for calculating the reliability score of multimodal physiological testing data can be expressed as:

[0256] in, Used to represent the reliability score of multimodal physiological test data Used to indicate effective wear coverage. Used to represent multimodal synchronization scores Used to represent the intensity score of periodic correlation patterns Used to represent the penalty score for consecutive missing items. Used to represent outlier penalty scores. , , , , , , and The weights are determined by the reliability score. A higher reliability score indicates that the physiological data is more suitable for periodic determination; a lower reliability score indicates that the current physiological signal may be unsuitable as the main basis for prediction due to missing, noise, or asynchrony.

[0257] In some embodiments, the historical cycle baseline rule sub-score includes a historical cycle regularity score, a cycle length deviation score, and a recent trend change score. For example, in this application embodiment, the weighted sum of each historical cycle baseline rule sub-score can be performed to obtain a risk score for cycle deviation.

[0258] In some embodiments, the state transition rule sub-score includes a physiological representation deviation score, an age or state background support score, and a state transition inhibition factor. In this embodiment, a state transition risk score can be obtained by combining the various state transition rule sub-scores. This state transition risk score is then used as a weakly supervised target for model training.

[0259] For example, the function for calculating the state transition risk score can be expressed as:

[0260] in, The state transition risk score is represented by p1, p2, and p3, which are weights. The clip is used to restrict the result to the interval [0,1].

[0261] In other embodiments, the auxiliary decision-making task objectives may also include historical observation period boundary reliability, suspected merged period probability, and historical observation period quality score. The historical observation period boundary reliability score characterizes the reliability of the historical observation period boundaries obtained from the initial segmentation based on user menstrual records. The suspected merged period probability characterizes whether a given historical observation period could be formed by the merging of multiple real periods. The historical observation period quality score comprehensively characterizes whether a given historical observation period is suitable as a valid historical input for subsequent anomaly source identification, state transition detection, and period prediction.

[0262] Therefore, in this embodiment of the application, the historical observation period boundary credibility, the probability of suspected merging period, and the historical observation period quality score calculated based on prior features can also be used as weak supervision targets for model training.

[0263] Step S530: Calculate the state transition risk score based on the sub-scores of the state transition rule, and use the state transition risk score as the state transition risk task objective.

[0264] In some embodiments, the state transition rule sub-score includes a physiological representation deviation score, an age or state background support score, and a state transition inhibition factor. In this embodiment, a state transition risk score can be obtained by combining the various state transition rule sub-scores. This state transition risk score is then used as a weakly supervised target for model training.

[0265] For example, the function for calculating the state transition risk score can be expressed as:

[0266] in, The state transition risk score is represented by p1, p2, and p3, which are weights. The clip is used to restrict the result to the interval [0,1].

[0267] Step S540: Use the periodic indication information in the target data as the target of the periodic prediction task.

[0268] In this embodiment, the period indication information refers to the menstrual start date, prediction interval, or period stage in the target data. By using the period indication information as a weakly supervised objective for model training, the reliability of the model's period prediction in complex scenarios can be improved.

[0269] In other embodiments, the cycle prediction task is used to output cycle-related prediction results. Taking the menstrual cycle scenario as an example, the cycle prediction results include, but are not limited to: the prediction interval of the next menstrual period, the most likely date, and the prediction confidence level.

[0270] The periodic prediction task is based on feature fusion representation and combines anomaly source information, auxiliary decision information and state transition information. The adaptive prediction adjustment unit adjusts the feature fusion method, prediction weight allocation, attention weight or confidence output to generate the corresponding periodic prediction results.

[0271] For example, when the record confidence score is low or the probability of suspected merging cycles is high, the cycle prediction task reduces the weight of historical observation cycle representation to avoid mistaking excessively long observation cycles caused by omissions as the user's true cycle length; when the reliability score is low or the probability of sensor anomalies is high, the cycle prediction task reduces the weight of wearable physiological time series representation and lowers the prediction confidence; when the state transition risk score is high, the cycle prediction task reduces the constraint weight of long-term historical baseline, increases the weight of recent cycle change representation, and expands the prediction interval.

[0272] In other embodiments, the periodic prediction result output by the periodic prediction task may include both point prediction results and interval prediction results. For example, when the data quality is high, the record reliability is high, and the user's state is stable, the system can output a narrower prediction interval and a higher prediction confidence level; when the record is unreliable, the sensor data is insufficient, or the risk of state transition is high, the system can output a wider prediction interval and a lower prediction confidence level, or prompt the user to supplement menstrual records and continue wearing the device to improve prediction reliability.

[0273] It should be noted that the steps in this embodiment are consistent with the corresponding steps in the foregoing embodiments. Therefore, for a detailed description of these steps, please refer to the description in the foregoing embodiments. This embodiment will not repeat them here.

[0274] Please continue reading. Figure 6 , Figure 6 This is a network architecture diagram of a target prediction model illustrated in an exemplary embodiment of this application. Figure 6 As shown, the target prediction model includes a shared representation learning layer (which can also be a multimodal representation learning module), a multi-task auxiliary output layer, an adaptive prediction adjustment unit, and a periodic prediction head.

[0275] The shared representation learning layer includes physiological coding units, historical cycle coding units, menstrual cycle record coding units, symptom coding units, user background coding units, and fusion units.

[0276] The physiological coding unit is used to encode multimodal physiological detection data within the current target cycle observation window, as well as historical multimodal physiological detection data within the most recent continuous time range, to extract rhythmic variation features, trend features, fluctuation features, and stage features related to the menstrual cycle. In some embodiments, historical observation cycle information can be further used to extract cycle-level representations from continuous-time physiological detection data, and participate in subsequent joint fusion together with continuous-time representations.

[0277] The historical cycle coding unit is used to encode the historical cycle length, cycle volatility, historical multi-cycle regularity, and historical abnormal cycle distribution, and to extract the user's long-term cycle baseline and long-term stability characteristics.

[0278] The menstrual cycle record encoding unit is used to encode prior features of record anomalies or weakly supervised signals, and to extract information representations related to record compliance. In some embodiments, prior features of record anomalies may be generated based on the delay between the record creation time and the business date, the difference between the record update time and the creation time, historical record activity, the proportion of delayed records, and signs of concentrated supplementary recording.

[0279] The symptom coding unit is used to encode the subjective symptoms (including emotions) recorded by the user and their occurrence time, and to extract the symptom distribution characteristics and the periodic symptom repetition characteristics within the period.

[0280] The user background encoding unit is used to quantify basic attribute information such as age.

[0281] The fusion unit is used to jointly model the results output by the above-mentioned coding units to generate a unified multimodal latent representation. In some embodiments, the fusion unit is used to construct the periodic correlation between the current target period observation window and the historical observation period, the deviation relationship between the multimodal physiological time series data of the current target period and the historical period baseline, the correlation between recorded abnormal prior information and the current period state, and the correlation between symptom input and the dynamic changes of the current period.

[0282] In this embodiment, instead of simply splicing together various inputs after processing them independently, a modular encoding and joint fusion approach is used to preserve the local features of each data source and establish the correlation between information from different sources. This approach is more suitable for handling the problem of judging the current changes in the menstrual cycle relative to the historical baseline.

[0283] The multi-task auxiliary output layer includes an anomaly source classification header, an auxiliary decision-making information output header, and a state transition detection header.

[0284] The anomaly source classification head is used to determine the scene to which the current target periodic sample belongs based on fused feature representation. The output of the anomaly source classification head includes the probability of a regular scene, the probability of a recorded abnormal scene, the probability of a sensor-sensored abnormal scene, and the probability of a physiologically highly variable scene.

[0285] For example, during the model inference phase, the anomaly source classification head generates raw discriminant scores corresponding to regular patterns, recorded anomalies, sensor anomalies, and physiologically highly variable scenarios based on fused feature representations. These raw discriminant scores are then converted into anomaly source probabilities using a normalization function. During the model training phase, classification evidence scores can be calculated based on the rule sub-scores obtained in the preprocessing phase. These scores are then normalized to obtain classification probability labels, which are used as weakly supervised targets to achieve supervised training of the anomaly source classification task.

[0286] In this embodiment, a unified model is used to directly output the probability distribution of normal scenarios and scenarios with multiple abnormal sources. Instead of first performing a simple normal / abnormal binary classification and then further splitting the abnormal sources by manual rules, it is based on fused feature representation and directly outputs the probabilities of multiple scenarios. This can avoid introducing multi-level errors and rule adaptation deviations during the transmission of binary classification results, achieve accurate judgment of abnormal types, reduce the cost of manual intervention, and improve the accuracy of scenario classification.

[0287] The auxiliary decision information output header is used to output an auxiliary score that can explain the current prediction basis and participate in the periodic prediction adjustment. This auxiliary score is generated by the unified model based on fused feature representation through the corresponding auxiliary output header, rather than being calculated separately by the independent rule system.

[0288] In some embodiments, the decision support information includes one or more of the following: record confidence score, reliability score of multimodal physiological testing data, probability of suspected merging cycles, state transition risk score, risk score of current cycle deviation from long-term baseline, confidence score of historical observation cycle boundaries, and quality score of historical observation cycles.

[0289] In this embodiment, the record reliability score is used to characterize whether the user's menstrual records or historical observation cycle boundaries are reliable; the reliability score of the multimodal physiological detection data is used to characterize whether the current multimodal physiological detection data has a sufficiently clear and stable cycle correlation pattern to support subsequent predictions; the suspected merged cycle probability is used to characterize whether a certain historical observation cycle may be formed by the merging of multiple real cycles; the state transition risk score is used to characterize whether the user has transitioned from a long-term stable state to a new physiological cycle stage; the current cycle deviation risk score is used to characterize the degree of deviation of the current cycle from the user's long-term historical pattern, providing an initial basis for subsequent state transition detection and cycle prediction adjustment; the historical observation cycle boundary reliability score is used to characterize whether the historical observation cycle boundaries obtained based on the initial division of the user's menstrual records are reliable; and the historical observation cycle quality score is used to characterize whether a certain historical observation cycle is suitable as an effective historical input for subsequent anomaly source discrimination, state transition detection, and cycle prediction.

[0290] The state transition detection head is used to handle boundary scenarios during state transitions, such as the onset of perimenopause, the postpartum recovery period, the recovery phase after discontinuing hormonal contraception, or the long-term stage of circadian rhythm changes.

[0291] In some embodiments, the state transition detection head is used to generate a phased representation of the current target cycle based on the multimodal physiological time-series data observed up to the current moment of the current target cycle; then, it generates a historical observation cycle representation based on the multimodal physiological time-series windows corresponding to up to three historical observation cycles, and performs weighted modeling by combining the boundary confidence of the historical observation cycles, the probability of suspected merged cycles, or the cycle quality score; then, it compares the differences between the current target cycle and the historical observation cycles and the long-term baseline based on the joint fusion representation; then, it further analyzes the deviation of the current target cycle from the long-term baseline by combining age, cross-cycle change trends, and the initial risk score of the current cycle deviating from the long-term baseline, and finally outputs the state transition risk score.

[0292] In this embodiment, the state transition risk score is not simply a judgment of whether the current cycle is getting longer, but a comprehensive judgment of whether the current cycle is transitioning from a stable historical state to a new physiological state. If the user's past cycles have been highly regular, but the current cycle begins to show a sustained long cycle and a continuous deviation of the multimodal physiological rhythm from the previous pattern, and this deviation cannot be fully explained by recording anomalies or sensing anomalies, then the model can output a high state transition risk score.

[0293] For example, if a user's past cycles have consistently been 29 to 31 days, with timely recording and complete wear data, but their current cycle has extended to day 42, and skin temperature, heart rate variability, and sleep rhythm are consistently deviating from historical patterns, the model can output a high risk score for a state transition. Conversely, if the current cycle appears longer, but the user has significant omissions, missed entries, or incomplete wear data, the model should not simply classify it as a state transition. Instead, it should consider both recording anomalies and sensor abnormalities to reduce the risk of a state transition. This avoids misidentification caused by relying solely on cycle length, distinguishes between abnormal data interference and genuine physiological state changes, and allows the target prediction model to better handle boundary scenarios prone to misjudgment, such as the onset of perimenopause, thereby improving cycle prediction accuracy.

[0294] The adaptive prediction and adjustment unit is used to receive anomaly source information, auxiliary decision-making information, and state transition information, and convert them into adjustment information for the periodic prediction head.

[0295] In some embodiments, the adaptive prediction adjustment unit can influence the periodic prediction task in the following ways: 1. Anomaly source information, auxiliary decision-making information, and state transition information are directly used as additional input features, concatenated with the fused feature representation, and then input into the cycle prediction head. This allows the cycle prediction head to directly learn the relationship between different auxiliary decision-making information and the final menstrual cycle prediction result, improving the accuracy of cycle prediction.

[0296] 2. Based on auxiliary information such as anomaly source information, decision support information, and state transition information, generate weights for different data sources or different representation branches.

[0297] 3. Based on auxiliary information such as anomaly source information, decision support information, and state transition information, attention is generated for different data sources or different representation branches.

[0298] 4. Based on auxiliary information such as anomaly source information, decision support information, and state transition information, generate confidence levels and prediction interval widths for calibrating periodic prediction results.

[0299] In this way, anomaly source information, auxiliary decision-making information, and state transition information can not only be output as interpretive results, but also participate in the adaptive generation of the final periodic prediction results as auxiliary decision-making information within the model. This enables multi-path adaptive optimization of prediction logic, balancing prediction accuracy and result interpretability, and adapting to various periodic disturbances and physiological state change scenarios.

[0300] The cycle prediction header is used to output cycle-related prediction results. For example, in the menstrual cycle scenario, the cycle prediction results include, but are not limited to: the prediction interval for the next menstrual period, the most likely date, and the prediction confidence level.

[0301] In some embodiments, the periodic prediction results obtained by the target prediction model in a single inference process may include anomaly source information, decision support information, state transition risk scores, and periodic prediction information. Anomaly source information, decision support information, and state transition risk scores can serve as internal decision support information within the model, or as interpretable results provided to the system or user, helping to explain the basis for prediction and improving the transparency and credibility of the model results.

[0302] The anomaly source information includes the probability of regular, predictable scenarios, the probability of recorded abnormal scenarios, the probability of sensor-related abnormal scenarios, and the probability of physiologically highly variable scenarios.

[0303] The decision support information includes the record confidence score, the reliability of multimodal physiological test data, the risk score of the current period deviating from the long-term baseline, the confidence score of the historical observation period boundary, the probability of suspected merged periods, and the quality score of historical observation periods.

[0304] The cycle prediction information includes the prediction range, most likely date, and prediction confidence level for the next menstrual period.

[0305] In this embodiment of the application, the terminal devices can be categorized and displayed accordingly: For example, if the terminal device is the backend, it can directly display various intermediate scores in the auxiliary information for model monitoring, result interpretation, abnormal sample analysis, and subsequent model iteration.

[0306] For example, if the terminal device is a client, the auxiliary information can be converted into natural language explanations and displayed to the user.

[0307] For example, when the probability of sensor malfunction is high and the reliability score of multimodal physiological detection data is low, the natural language interpretation of the auxiliary information can be "Insufficient recent wear data, low reliability of this prediction".

[0308] When the probability of recording abnormal scenarios is high and the record credibility score is low, the natural language explanation corresponding to the auxiliary information can be "Recent menstrual records may have been supplemented or are missing; it is recommended to supplement and confirm them."

[0309] When the state transition risk score is high, the natural language interpretation of the auxiliary information can be "the recent cycle pattern is significantly different from the historical pattern, and the prediction range has been widened accordingly".

[0310] When the record confidence score is high and the reliability score of the multimodal physiological detection data is high, the natural language interpretation of the auxiliary information can be "the basis for this prediction is relatively sufficient".

[0311] Please continue reading. Figure 7 , Figure 7 This is a schematic diagram illustrating the predictive application of periodic information, as shown in an exemplary embodiment of this application. Figure 7 The process begins by acquiring multi-source input data, including multimodal physiological testing data, along with historical cycle data, menstrual-related event records, symptom input, and user background data. Then, features are extracted from different data types separately, followed by feature fusion to obtain a unified feature fusion representation. Next, auxiliary information is generated based on the feature fusion representation, including at least one of the following: anomaly source information, decision support information, and state transition information.

[0312] Ancillary information can be used not only as an explanatory output, but also as an internal decision-making aid within the model, participating in the generation of periodic prediction results.

[0313] Specifically, in some embodiments, auxiliary information can be directly concatenated with the feature fusion representation as additional input features to obtain target fusion features. Then, the target fusion features are directly input into the target prediction model without adaptive adjustment for periodic prediction to obtain periodic prediction results.

[0314] In this embodiment, optimization can be achieved simply by optimizing the input of the prediction period head, without having to perform parameter overwriting or weight modification operations during the inference stage. This avoids the risks of model parameter disorder, gradient shift, and historical state contamination that may be caused by dynamic parameter tuning, making the model run more stably and thus improving the reliability of period information prediction.

[0315] In other embodiments, a prediction adjustment factor can be determined using auxiliary information. Then, the target prediction model is adaptively adjusted based on this factor. After obtaining the adaptively adjusted target prediction model, the initial feature fusion representation can be directly input into the adaptively adjusted target prediction model for periodic prediction, thereby obtaining a periodic prediction result. Alternatively, the concatenated target fusion features can be directly input into the target prediction model without adaptive adjustment for periodic prediction, thereby obtaining a periodic prediction result.

[0316] In this embodiment, the model's internal weights, attention, and output layer constraints are modified using a prediction adjustment factor, achieving optimization simply by adjusting the model parameters. Furthermore, based on adaptive model adjustment, the input to the prediction cycle head is further optimized. Three types of high-dimensional structured information—anomaly sources, decision support, and state transitions—are directly integrated with the basic fusion representation to form the input. This allows the model to learn the correlation between perturbation factors and cycle changes from the underlying input. By combining input feature optimization with adaptive model parameter adjustment, the model exhibits stronger fitting capabilities in complex scenarios, further reducing prediction errors and improving the reliability of prediction results.

[0317] In other embodiments, the target prediction model can be supervised through multi-task joint training, including: acquiring training data, which includes multimodal physiological time-series data, periodic history data, menstrual-related event records, symptom input data, user background data, prior features, and weakly supervised targets constructed based on the prior features. The prior features are used as model input for representation learning, and the weakly supervised targets serve as supervisory signals during the training phase, constraining the outputs of the anomaly source classification head, the auxiliary decision information output head, and the state transition detection head. After training, during the inference phase, the model automatically outputs the anomaly source probability and various auxiliary scores based on the input data, rather than the final judgment result being directly given by the rule system.

[0318] For example, the sub-scores themselves are neither the loss function nor the final output of the model. The rule sub-scores are first combined into a weakly supervised objective, and then the loss is calculated between the weakly supervised objective and the model output.

[0319] Taking the training process of recording credibility scores as an example: The raw data includes record creation date, business date, number of modifications, interval between adjacent records, physiological signals, etc. The raw data is preprocessed to obtain rule sub-scores, such as record delay penalty score, modification penalty score, centralized supplementary recording penalty score, boundary physiological consistency support score, etc. Then, these rule sub-scores are combined to obtain weakly supervised objectives, such as weakly supervised scores for record credibility. The model output is the predicted score for record credibility.

[0320] The loss function can be expressed as: L_record = loss(S_record_pred, S_record_target) Where L_record represents the record confidence score loss, S_record_pred represents the predicted record confidence score, and S_record_target represents the weakly supervised record confidence score.

[0321] In some embodiments, the multi-task loss function of a multi-objective prediction model can be expressed as: L_total=L_period+λ1L_abnormal+λ2L_record+λ3L_signal+λ4L_deviation+λ5L_boundary+λ6L_merge+λ7L_quality+λ8L_transition Where L_total represents the total loss, L_period represents the periodic prediction loss, L_abnormal represents the anomaly source classification loss, L_record represents the record credibility score loss, L_signal represents the physiological detection data reliability loss, L_deviation represents the current period's deviation from the long-term baseline risk score loss, L_boundary represents the historical observation period boundary credibility score loss, L_merge represents the suspected merged period probability loss, L_quality represents the historical observation period quality score loss, L_transition represents the state transition risk score loss, and λ1 to λ8 are the loss weights for each task.

[0322] For anomaly source identification tasks, cross-entropy loss or multi-label binary cross-entropy loss can be used; for continuous score prediction tasks, mean squared error loss, mean absolute error loss, binary cross-entropy loss, or ranking loss can be used; for periodic prediction tasks, date regression loss, interval prediction loss, distribution prediction loss, or a combination thereof can be used.

[0323] In this embodiment of the application, through multi-task joint training, the anomaly source classification task, the auxiliary decision information prediction task, and the state transition detection task can jointly constrain the feature fusion representation, so that the period prediction task can not only learn the historical period length and the current physiological signal, but also learn whether the record is reliable, whether the sensor data is available, whether the historical period boundary is reliable, and whether the user may enter a new physiological state stage, thereby improving the prediction robustness in complex scenarios.

[0324] Please continue reading. Figure 8 , Figure 8 This is a schematic diagram illustrating the predictive application of periodic information, which is another exemplary embodiment of this application. Figure 8 In the process, the terminal device acquires multimodal physiological detection data collected by smart wearable devices, including skin temperature, heart rate, heart rate variability, respiratory rate, and sleep data. It also acquires cycle history data, menstrual-related event records, symptom input, and user background data.

[0325] Next, the multi-source input data is preprocessed. By constructing the current target period observation window, historical continuous time window, and historical observation period window, target data is extracted from the multi-source input data, and prior features are constructed based on the target data. Prior features include various types of rule sub-scores and target data. Rule sub-scores include, for example, record behavior rule sub-scores, data quality rule sub-scores, historical period baseline rule sub-scores, state transition rule sub-scores, and anomaly source rule sub-scores.

[0326] Then, the multi-modal representation learning layer of the target prediction model encodes and fuses the multi-source input data in modules to obtain fused feature representations.

[0327] The fused feature representations are input into the multi-task auxiliary output layer of the target prediction model for prediction, thereby obtaining anomaly source information, auxiliary decision-making information, and state transition information.

[0328] Then, through the adaptive prediction adjustment unit of the target prediction model, the model parameters are optimized based on anomaly source information, auxiliary decision-making information, and state transition information by means of feature concatenation, adjustment of gating weights, attention, confidence, etc.

[0329] Finally, the period prediction result is output and displayed through the period prediction head.

[0330] During the training phase, prior features can be obtained from the target data, and weakly supervised targets for the corresponding tasks can be generated based on these prior features. Then, the prior features are used as model inputs, and the weakly supervised targets are used as training objectives to perform multi-task joint training on the target prediction model. This optimizes the parameters of the multi-task auxiliary output layer and the periodic prediction head.

[0331] Combination Figure 9 As shown, Figure 9 This is a structural diagram of a periodic information prediction device illustrated in an exemplary embodiment of this application. Figure 9 As shown, the exemplary device for predicting periodic information specifically includes: an acquisition module 910, a feature fusion module 920, a first prediction module 930, an adjustment module 940, and a second prediction module 950; wherein, the acquisition module 910 is configured to acquire multi-source input data; wherein, the multi-source input data includes multimodal physiological detection data and periodic correlation data containing historical periodic data; the feature fusion module 920 is configured to perform feature extraction and feature fusion on the multi-source input data to obtain a feature fusion representation; the first prediction module 930 is configured to acquire auxiliary information based on the feature fusion representation, the auxiliary information including at least one of anomaly source information, auxiliary decision information, and state transition information; the adjustment module 940 is configured to determine a prediction adjustment factor based on the auxiliary information; the second prediction module 950 is configured to adaptively adjust the target prediction model based on the prediction adjustment factor, and perform periodic prediction based on the adaptively adjusted target prediction model to obtain a periodic prediction result.

[0332] By employing the periodic information prediction device in this application embodiment, on the one hand, by acquiring multi-source input data and performing feature extraction and feature fusion on the multi-source input data, it is possible to establish the correlation between different types of input data, forming a feature fusion representation with global representation capabilities, which facilitates obtaining more accurate auxiliary information for subsequent predictions and improves the prediction reliability of periodic information; on the other hand, by acquiring auxiliary information such as anomaly source information, auxiliary decision-making information, and state transition information through feature fusion representation, and using this auxiliary information to determine the prediction adjustment factor, and then adaptively adjusting the target prediction model based on the prediction adjustment factor, the model can achieve corresponding adjustments in complex scenarios such as inaccurate or abnormal input data, accurately matching the current data state, anomaly type, and periodic change trend, thereby reducing the periodic prediction error in complex scenarios and further improving the prediction reliability of periodic information.

[0333] It should be noted that the periodic information prediction device and the periodic information prediction method provided in the above embodiments belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the periodic information prediction device provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0334] Embodiments of this application also provide an electronic device, including: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, cause the electronic device to implement the periodic information prediction method provided in the above embodiments.

[0335] Figure 10 A schematic diagram of a computer system suitable for implementing the embodiments of this application is shown. It should be noted that... Figure 10 The computer system 1000 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0336] like Figure 10As shown, the computer system 1000 includes a Central Processing Unit (CPU) 1001, which can perform various appropriate actions and processes based on programs stored in Read-Only Memory (ROM) 1002 or programs loaded from storage portion 1008 into Random Access Memory (RAM) 1003, such as performing the methods described in the above embodiments. Various programs and data required for system operation are also stored in RAM 1003. The CPU 1001, ROM 1002, and RAM 1003 are interconnected via bus 1004. An Input / Output (I / O) interface 1005 is also connected to bus 1004.

[0337] The following components are connected to I / O interface 1005: an input section 1006 including a keyboard, mouse, etc.; an output section 1007 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a model interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 1009 performs communication processing via a model such as the Internet. A drive 1010 is also connected to I / O interface 1005 as needed. Removable media 1011, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 1010 as needed so that computer programs read from them can be installed into storage section 1008 as needed.

[0338] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from the model via communication section 1009, and / or installed from removable medium 1011. When the computer program is executed by central processing unit (CPU) 1001, it performs various functions defined in the system of this application.

[0339] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable computer program. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0340] Another aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for predicting periodic information as described above. This computer-readable storage medium may be included in the electronic device described in the above embodiments, or it may exist independently and not assembled into the electronic device.

[0341] Another aspect of this application provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the periodic information prediction method provided in the various embodiments described above.

[0342] The above description is merely a preferred exemplary embodiment of this application and is not intended to limit the implementation of this application. Those skilled in the art can easily make corresponding modifications or alterations based on the main concept and spirit of this application. Therefore, the scope of protection of this application should be determined by the scope of protection claimed in the claims.

[0343] It is understood that in the specific embodiments of this application, data related to the menstrual cycle (such as cycle information, date information, etc.) are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

Claims

1. A method for predicting periodic information, characterized in that, include: Acquire multi-source input data; wherein, the multi-source input data includes multimodal physiological detection data and periodic correlation data containing historical periodic data; Feature extraction and feature fusion are performed on the multi-source input data to obtain a feature fusion representation; Auxiliary information is obtained based on the feature fusion representation, and the auxiliary information includes at least one of anomaly source information, auxiliary decision information, and state transition information. The prediction adjustment factor is determined based on the auxiliary information; The target prediction model is adaptively adjusted based on the aforementioned prediction adjustment factor, and periodic prediction is performed based on the adaptively adjusted target prediction model to obtain the periodic prediction result.

2. The method according to claim 1, characterized in that, The step of extracting and fusing features from the multi-source input data to obtain a feature fusion representation includes: Encode different types of data in the multi-source input data to obtain multiple sub-feature representations; The multiple sub-feature representations are subjected to feature fusion processing to obtain the feature fusion representation.

3. The method according to claim 1, characterized in that, Determining the prediction adjustment factor based on the auxiliary information includes: Obtain the score thresholds corresponding to the anomaly source information, the auxiliary decision-making information, and the state transition information, respectively; The anomaly source information, the auxiliary decision-making information, and the state transition information are compared with the corresponding preset score thresholds to obtain the comparison results; Based on the comparison results, the representation weights of different types of data in the multi-source input data are determined, the attention of the target prediction model to different types of data in the multi-source input data is determined, and the confidence and prediction interval width of the periodic prediction results are determined. The prediction adjustment factor is determined from the representation weights, the attention, the confidence level, and the prediction interval width.

4. The method according to claim 1, characterized in that, The step of performing periodic prediction based on the adaptively adjusted target prediction model to obtain periodic prediction results includes: The feature fusion representation is input into the adaptively adjusted target prediction model for periodic prediction to obtain the periodic prediction result; or, The auxiliary information is used as an additional input feature and concatenated with the feature fusion representation to obtain the target fusion feature; The target fusion features are input into the adaptively adjusted target prediction model for periodic prediction to obtain the periodic prediction result.

5. The method according to claim 1, characterized in that, The process of acquiring the auxiliary information is executed by the multi-task auxiliary output layer of the target prediction model, and the multi-task output layer optimizes the parameters in the following way: Construct the target time window; Based on the target time window, data extraction is performed on the multi-source input data to obtain the target data; Prior features are obtained based on the target data, and weakly supervised targets for the corresponding tasks are generated based on the prior features. The prior features are used as model inputs, and the weakly supervised target is used as the training target to perform multi-task joint training on the target prediction model, so as to optimize the parameters of the multi-task auxiliary output layer.

6. The method according to claim 5, characterized in that, The prior features include recording behavioral rule sub-scores, and obtaining the prior features based on the target data includes: Extract historical periodic data recording behavior information from the target data; Based on the recorded behavior information, a gain influence factor and a decay influence factor are calculated to characterize the recording reliability of the historical periodic data. The gain influence factor and the attenuation influence factor are used as sub-scores of the recording behavior rule.

7. The method according to claim 5, characterized in that, The prior features include data quality rule sub-scores, and the step of obtaining prior features based on the target data includes: Data quality information of multimodal physiological detection data is extracted from the target data; Based on the data quality information, a reliability impact factor for characterizing the data quality of the multimodal physiological detection data is obtained. The reliability impact factor is used as the sub-score of the data quality rule.

8. The method according to claim 5, characterized in that, The prior features include historical periodic baseline rule sub-scores, and the step of obtaining prior features based on the target data includes: Extract periodic feature information from historical periodic data from the target data; Based on the aforementioned periodic characteristic information, a deviation influence factor is calculated, which is used to characterize the degree of deviation of the historical periodic data from historical patterns. The deviation impact factor is used as the historical cycle baseline rule sub-score.

9. The method according to claim 5, characterized in that, The prior features include state transition rule sub-scores, and the step of obtaining prior features based on the target data includes: Extract the periodic feature information, user background information, and recorded behavior information of historical periodic data from the target data; Based on the periodic feature information, the user background information, and the recorded behavior information, a potential influencing factor for characterizing changes in the user's physiological state is obtained. The probability influence factor is used as the sub-score of the state transition rule.

10. The method according to claim 5, characterized in that, The prior features include anomaly source rule sub-scores, and the step of obtaining prior features based on the target data includes: Based on the target data, the following sub-scores are calculated: record behavior rule sub-score, data quality rule sub-score, and historical period baseline rule sub-score. Based on the record behavior rule sub-score, the data quality rule sub-score, and the historical period baseline rule sub-score, a classification evidence score for characterizing the anomaly source type is obtained; The classification evidence score is used as the sub-score of the anomaly source rule.

11. The method according to claim 5, characterized in that, The weakly supervised objectives for the corresponding tasks include anomaly source classification task objectives, decision support task objectives, state transition risk task objectives, and periodic prediction task objectives. The prior features include record behavior rule sub-scores, data quality rule sub-scores, historical periodic baseline rule sub-scores, state transition rule sub-scores, anomaly source rule sub-scores, and the target data. Generating the weakly supervised objectives for the corresponding tasks based on the prior features includes: The classification probability label is calculated based on the sub-score of the anomaly source rule, and the classification probability label is used as the target of the anomaly source classification task. The record credibility score is calculated based on the record behavior rule sub-score; the reliability score of the multimodal physiological detection data is calculated based on the data quality rule sub-score; and the risk score of cycle deviation is calculated based on the historical cycle baseline rule sub-score; and the record credibility score, the reliability score of the multimodal physiological detection data, and the risk score of cycle deviation are used as the target of the auxiliary decision-making task. The state transition risk score is calculated based on the state transition rule sub-score, and the state transition risk score is used as the state transition risk task objective. The periodic indication information in the target data is used as the target of the periodic prediction task.

12. A device for predicting periodic information, characterized in that, include: The acquisition module is configured to acquire multi-source input data; wherein, the multi-source input data includes multimodal physiological detection data and periodic correlation data containing historical periodic data; The feature fusion module is configured to extract and fuse features from the multi-source input data to obtain a feature fusion representation. The first prediction module is configured to obtain auxiliary information based on the feature fusion representation, wherein the auxiliary information includes at least one of anomaly source information, auxiliary decision information, and state transition information. The adjustment module is configured to determine a predictive adjustment factor based on the auxiliary information; The second prediction module is configured to adaptively adjust the target prediction model based on the prediction adjustment factor, and perform periodic prediction based on the adaptively adjusted target prediction model to obtain the periodic prediction result.

13. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the electronic device to implement the method for predicting periodic information as described in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, It stores computer-readable instructions that, when executed by the processor of a computer, cause the computer to perform a method for predicting periodic information as described in any one of claims 1 to 11.

Images