A female physiological state inference method and system based on multi-modal physiological signals

Abstract

This invention discloses a method and system for inferring female physiological states based on multimodal physiological signals. The method acquires multimodal physiological signals from a user through a wearable device, including but not limited to skin temperature, heart rate variability, resting heart rate, sleep data, and electrodermal signals; constructs an individual baseline model for the user; based on the deviation characteristics of the current signal from the individual baseline, constructs a state space containing multiple candidate physiological states; performs constraint contraction processing on the state space, the constraints including at least two of the following: state transition order constraint, state duration constraint, signal change direction constraint, and multimodal signal collaborative constraint; and determines the final physiological state from the contracted state space. The method also includes a cycle segmentation mechanism, data sufficiency judgment and degradation control, closed-loop intervention effect evaluation, risk probability calculation, and intelligent suggestion generation based on a language model. This invention upgrades female cycle management from calendar-time-based calculation to constrained state inference based on real-time multimodal physiological signals, applicable to applications such as premenstrual syndrome prediction, ovulation period identification, assisted conception, contraception reference, early pregnancy monitoring, postpartum recovery tracking, and perimenopausal management.

Description

Technical Field

[0001] This invention relates to the fields of physiological signal processing and digital health technology, and in particular to a method and system for modeling, inferring and evaluating the state of a woman's menstrual cycle based on multimodal physiological signals collected by wearable devices. Background Technology

[0002] The female menstrual cycle is a complex physiological process driven by hormonal changes, typically including the menstrual period, follicular phase, ovulation phase, and luteal phase. Accurately identifying a user's current physiological state is of significant value in various applications, including premenstrual syndrome (PMS) prediction, ovulation identification, determining the window of opportunity for conception, assisted contraception, monitoring physiological changes in early pregnancy, postpartum recovery tracking, perimenopausal management, and the daily management of cycle-related diseases such as polycystic ovary syndrome (PCOS) and endometriosis.

[0003] Current female cycle management products primarily rely on the following technological approaches: First, cycle prediction methods based on calendar time. This method takes the user's recorded menstrual cycle start and end dates as input, calculates the average or trend of historical cycle lengths, and then estimates future menstrual and ovulation dates. Essentially, this method is a time-based estimation and does not involve identifying the user's current physiological state. For users with irregular cycles (such as patients with polycystic ovary syndrome or perimenopausal women), the prediction bias is significant.

[0004] Second, there are auxiliary judgment methods based on single physiological signals. For example, the basal body temperature (BBT) method infers whether ovulation has occurred by measuring the biphasic change in morning resting body temperature (see Pierson EE, Yau C. Modelling temperature data for menstrual cycle length prediction. arXiv:1606.05612,2016). However, a single body temperature signal is easily affected by factors such as sleep quality, ambient temperature, illness, and travel, leading to misjudgment. Similarly, methods relying solely on heart rate variability (HRV) or solely on resting heart rate (RHR) also face the same noise interference problem.

[0005] Third, end-to-end prediction methods based on machine learning (see CN113892907A, Biorhythm Detection Method, Device, Equipment, and Medium Based on Wearable Devices). These methods directly map input data to prediction results, without constraining the rationality of physiological states within the model. For example, the model might output a result of "jumping directly from the luteal phase to the ovulation phase," which is physiologically impossible, but the end-to-end model lacks a mechanism to exclude such unreasonable states.

[0006] In addition, some literature has proposed using Hidden Markov Models (HMMs) to model the menstrual cycle (see SymulL, Hsieh P, Engel M, et al. Inferring time-varying menstrual cycle related physiological parameters using wearable data. npj Digital Medicine, 2019). However, this method is mainly used for probability estimation of cycle length and ovulation day, and has not established a multimodal state space construction and constraint contraction mechanism based on individual baselines.

[0007] The above-mentioned existing technologies have the following common problems: (1) They are based on time estimation and have not established a state recognition mechanism based on real-time physiological signals; (2) Single signals are easily affected by non-periodic factors; (3) They lack a mechanism to verify candidate states by physiological constraints; (4) They do not consider the pollution problem of abnormal cycles on state inference; (5) They still output deterministic results when the data is insufficient, and lack reliability protection.

[0008] Therefore, there is a need for a method that can utilize multimodal physiological signals, construct state space based on individual baseline deviations, and shrink state through physiological constraints to achieve reliable inference of female physiological states. Summary of the Invention

[0009] The technical problem that this invention aims to solve is that existing methods for managing female menstrual cycles rely on calendar time calculations or single signal judgments, which cannot accurately identify the user's current physiological state and lack a mechanism to verify the physiological rationality of the inference results.

[0010] To address the aforementioned technical problems, this invention provides a method for inferring a woman's physiological state based on multimodal physiological signals, comprising the following steps: S1, acquiring multimodal physiological signal data of the user through a wearable device. The multimodal physiological signal data includes at least two or more physiological signals, including but not limited to: skin temperature, heart rate variability, resting heart rate, sleep data, electrodermal signal, blood oxygen saturation, respiratory rate, and sweat composition data. The multimodal physiological signal data may also include user subjective report data, including but not limited to symptom self-assessment scale data, menstrual cycle records, or mood record data.

[0011] S2. Preprocess the multimodal physiological signal data, including outlier removal, time alignment, and daily aggregation.

[0012] S3. Based on the preprocessed multimodal physiological signal data, construct an individual baseline model for the user. The individual baseline model is used to describe the stable range and deviation pattern of the user in each physiological signal dimension.

[0013] S4. Calculate the deviation features of the physiological signal within the current time window relative to the individual baseline model, and construct a state space containing multiple candidate physiological states based on the deviation features. Each candidate state in the state space corresponds to a confidence value.

[0014] S5. Perform constraint shrinkage processing on the state space. The constraint shrinkage processing includes at least two of the following constraints: state transition order constraint, state duration constraint, signal change direction constraint, and multimodal signal cooperative constraint.

[0015] S6. From the state space after constraint contraction, select the candidate state with the highest confidence as the final physiological state output.

[0016] Furthermore, the determination of the final physiological state depends on the synergistic relationship between at least two physiological signals; a single physiological signal is not the sole basis for determining the final physiological state.

[0017] Furthermore, the method also includes a periodic segmentation step: when abnormal intervals are detected in the physiological cycle data, the data is divided into different periodic segments, and the state inferences between different periodic segments are isolated from each other.

[0018] Furthermore, the method also includes a data sufficiency judgment step: before performing state space construction, the sufficiency of the currently available data is judged; when the data does not meet the sufficiency conditions, a downgrade process is performed.

[0019] Furthermore, the method also includes an indirect inference step for hormone-dominant state: based on the combined change structure of at least two physiological signals, the user's current hormone-dominant state is indirectly inferred.

[0020] Furthermore, the method also includes an intervention effect evaluation step: acquiring the user's intervention behavior data, and evaluating the impact of the intervention behavior on the physiological state by comparing changes in physiological signals and state distribution before and after the intervention.

[0021] Furthermore, the method also includes a risk probability calculation step: based on the confidence distribution during the final physiological state and / or the constraint contraction process, calculate the risk probability associated with specific physiological events such as the onset of premenstrual syndrome, ovulation window, and cycle abnormalities.

[0022] Furthermore, the method also includes an intelligent suggestion generation step: taking the final physiological state, risk probability and / or physiological signal data as input, and providing them to a language model or generative model, which then generates personalized health suggestions or interpretation information.

[0023] The present invention also provides a female physiological state inference system based on multimodal physiological signals, including a data acquisition module, a preprocessing module, a baseline modeling module, a state space construction module, a constraint contraction module, and a state output module.

[0024] The present invention also provides a computer-readable storage medium and a computer program product for implementing the above-described method.

[0025] The beneficial effects of the present invention include: (1) upgrading from time estimation to state inference based on real-time multimodal signals, which can identify the user's current real physiological state.

[0026] (2) By constructing and confining the state space, we can ensure that the inference results are physiologically reasonable and avoid unreasonable state transitions.

[0027] (3) Multimodal signal collaborative modeling reduces the impact of noise interference on a single signal.

[0028] (4) The period segmentation mechanism makes the system robust to abnormal periods (such as long periods related to PCOS and perimenopausal fluctuations).

[0029] (5) Data sufficiency protection mechanism to avoid outputting unreliable results when data is insufficient.

[0030] (6) The intervention effect evaluation function upgrades the system from a simple recording tool to a verifiable health management tool. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0032] Figure 2 A schematic diagram of the state space construction and constraint shrinkage process.

[0033] Figure 3 This is a schematic diagram of a periodic segmentation mechanism.

[0034] Figure 4 This is a schematic diagram of the data sufficiency assessment and downgrade mechanism.

[0035] Figure 5 This is a closed-loop diagram for evaluating the intervention effect.

[0036] Figure 6 This is a schematic diagram of the module architecture of the system of the present invention. Detailed Implementation

[0037] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0038] It should be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of the invention. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. In the event of any discrepancy between the terminology used herein and its ordinary meaning, the definition herein shall prevail.

[0039] It should be understood that when a component is referred to as "connected" or "coupled" to another component, it can be directly connected or coupled to the other component, or there can be an intermediate component between them. The term "multiple" as used in this article refers to two or more.

[0040] Example 1: Acquisition and Preprocessing of Multimodal Physiological Signals. This example describes the acquisition and preprocessing process of multimodal physiological signal data.

[0041] When a user wears a smartwatch (such as an Apple Watch), the system collects the following physiological signals through the wearable device's sensors: (1) Skin temperature: continuously collected by the temperature sensor at the bottom of the watch during the user's sleep. The wrist temperature signal quality is highest during sleep because there is minimal interference from the external environment during sleep. The system extracts the temperature data during each night's sleep and calculates the representative temperature value for that night (such as the median or mean).

[0042] (2) Heart Rate Variability (HRV): The degree of variability between adjacent heartbeats is collected by the photoplethysmography (PPG) sensor of the watch. The system aggregates the data on a daily basis to obtain a representative HRV value for each day. In the specific implementation, the HRV index provided by Apple Watch is SDNN (standard deviation of adjacent NN intervals). This method is also applicable to other HRV indices such as rMSSD.

[0043] (3) Resting heart rate (RHR): The system extracts the user's heart rate data in a resting state and aggregates it on a daily basis.

[0044] (4) Sleep data: including total sleep duration, proportion of sleep stages (deep sleep, light sleep, REM), etc.

[0045] The preprocessing steps include: (a) Outlier removal: For each dimension of the signal, calculate the mean and standard deviation within a sliding window (e.g., 14 days), and mark data points that deviate from the mean by more than 3 standard deviations as outliers and remove them.

[0046] (b) Time alignment: Align the sampling data from different sensors to a unified date axis. Since different signals have different sampling frequencies and time periods (e.g., temperature is only collected during sleep, while HRV may be collected throughout the day), aggregation and alignment are required on a daily basis.

[0047] (c) Daily aggregation: Aggregate multiple samples from each day into a single representative value. For temperature, use the median during sleep; for HRV and RHR, use the daily mean or daily median.

[0048] Example 2: Construction of Individual Baseline Model. This example describes the method for constructing an individual baseline model.

[0049] The goal of the individual baseline model is to describe the "normal range" of the user across each signal dimension, serving as a reference for subsequent deviation calculations. The individual baseline model can be constructed directly based on the user's historical data, or it can be obtained by continuously calibrating the user's data based on the population statistical baseline. The construction method is as follows: (1) Initial baseline establishment: Collect data for at least 14 consecutive days after the user's first use (preferably 28 days, i.e., a complete cycle), and calculate the overall statistical characteristics of each signal, including the mean, standard deviation, and quartile range. In the cold start phase when the user data is less than 14 days, the system can use the population statistical baseline as the initial value, and gradually calibrate it to the individual baseline as the user data accumulates.

[0050] (2) Establishment of phase baselines: After accumulating data from at least two complete cycles, phase baselines can be further established. This involves statistically analyzing the signal characteristics of the user during each stage of the menstrual period, follicular phase, before and after ovulation, and luteal phase. For example, a user's wrist temperature baseline during the luteal phase is 0.2-0.4°C higher than that during the follicular phase, and their HRV baseline is 10-20% lower than that during the follicular phase.

[0051] (3) Baseline dynamic update: As new data is continuously input, the baseline model is updated in a sliding window manner. The update window is preferably the data of the most recent 3-6 periods to adapt to the natural drift of the user's physiological state.

[0052] The baseline model is stored in a local database (such as SQLite) to ensure that user privacy data does not leave the local device.

[0053] Example 3: Construction of State Space. This example describes the process of constructing a candidate state space based on signal deviation features.

[0054] The system defines the following set of female physiological states: S = {menstrual period, early follicular phase, late follicular phase, ovulation period, early luteal phase, mid-luteal phase, late luteal phase}.

[0055] For the current time window (e.g., signal data from the last 3-5 days), the system calculates the deviation feature vector of each signal relative to the individual baseline: D = (d_temp, d_hrv, d_rhr, d_sleep). Where d_temp is the temperature deviation (current value minus the baseline mean, then divided by the baseline standard deviation), d_hrv, d_rhr, and d_sleep are similar.

[0056] Based on the deviation feature vector D, the system calculates the initial confidence level for each state in the state set S. The calculation method can be: (a) rule-based method: establish mapping rules based on known physiological laws. For example, if d_temp > 1.0 (temperature is significantly higher than baseline), the initial confidence level of the luteal phase-related state increases; if d_hrv < -1.0 (HRV is significantly lower than baseline), it also points to the luteal phase.

[0057] (b) Probabilistic model-based approach: Using Gaussian mixture model or hidden Markov model, the posterior probability of the current deviation feature belonging to each state is calculated based on the signal distribution corresponding to each state in historical data.

[0058] (c) Machine learning model-based approach: Train a classification model to output the probability distribution of each candidate state.

[0059] Regardless of the method used, the output is a set of candidate states and their corresponding confidence values, i.e., the state space SS = {(s_i, c_i) | s_i ∈ S, c_i ∈ [0,1]}.

[0060] Example 4: Constraint Shrinkage Process. This example describes the specific process of performing constraint shrinkage on the state space.

[0061] The goal of constraint contraction is to use prior physiological knowledge to screen and contract the initial state space, eliminating candidate states that do not conform to physiological laws.

[0062] (1) State transition sequence constraint: In the female menstrual cycle, state transitions follow a fixed physiological sequence: menstruation → follicular phase → ovulation → luteal phase → menstruation. The system checks the transition relationship between the current candidate state and the previous determined state. If the transition of the candidate state does not conform to the above sequence (such as jumping directly from the early luteal phase to the early follicular phase without going through the menstrual phase), the confidence of the candidate state is set to 0 or significantly reduced.

[0063] (2) Duration constraint of each state: Each physiological state has a physiologically reasonable duration range. For example, the follicular phase usually lasts 7-21 days, the luteal phase usually lasts 10-16 days, and the menstrual period usually lasts 3-7 days. If the implied duration of a candidate state is not within a reasonable range (e.g., the luteal phase lasts only 2 days), its confidence level is reduced.

[0064] (3) Constraints on the direction of signal change: Each state transition is accompanied by a signal change in a specific direction. For example, when transitioning from the follicular phase to the luteal phase, wrist temperature should rise (due to the increase in progesterone after ovulation, which leads to an increase in body temperature). If the actual observed direction of temperature change does not match the expectation, the confidence level of the corresponding candidate state is reduced.

[0065] (4) Multimodal signal co-constraint: The confirmation of certain states requires the coordinated changes of multiple signals. For example, the confirmation of ovulation requires not only a rise in temperature, but also a coordinated change in HRV decrease and RHR increase. If there is only a single signal change and other signals do not show a consistent trend, the confidence of the candidate state is reduced.

[0066] After constraint contraction, the system selects the state with the highest confidence from the remaining candidate states as the final physiological state output. If the highest confidence is lower than a preset threshold (e.g., 0.5), the output result is marked as "low confidence".

[0067] Example 5: A specific example of multimodal signal collaborative judgment. This example illustrates the process of multimodal signal collaborative judgment using a specific case.

[0068] The signal deviation characteristics of a user on day 15 are as follows: wrist temperature increased by 0.3°C relative to baseline (d_temp = 1.5), HRV decreased by 15% relative to baseline (d_hrv = -1.2), and RHR increased by 3 bpm relative to baseline (d_rhr = 0.8).

[0069] Individual analysis of each signal: Increased temperature may be caused by increased progesterone after ovulation, or by a cold or fever; decreased HRV may be caused by increased sympathetic nerve activity during the luteal phase, or by stress or lack of sleep; increased RHR may be caused by physiological changes during the luteal phase, or by insufficient recovery after exercise.

[0070] Multimodal synergistic analysis: When three signals—temperature increase, HRV decrease, and RHR increase—occur simultaneously, and the time windows of these changes are consistent (all occurring within 2-3 days), their synergistic pattern strongly points to "ovulation has occurred, and the luteal phase has begun." The reliability of this synergistic judgment is far higher than that of a single signal.

[0071] If only a temperature increase is observed at the same time, but there are no significant changes in HRV and RHR, the system will not identify the state as the luteal phase, but will retain multiple candidate states (such as "possible ovulation" and "non-cyclical temperature fluctuations") and mark them as low confidence, waiting for subsequent data confirmation.

[0072] Example 6: Periodic Segmentation Mechanism. This example describes a segmentation isolation mechanism under abnormal periodic conditions.

[0073] In real-world usage scenarios, women may experience the following abnormal cycles: (a) prolonged cycles due to stress (e.g., cycles exceeding 45 days); (b) irregular cycles caused by PCOS; (c) perimenopausal cycle fluctuations; and (d) signal interruptions caused by travel, illness, etc.

[0074] When the system detects any of the following abnormal interval conditions, it will trigger period segmentation: (a) The current period length exceeds the preset threshold (e.g., 45 days), which is determined to be an abnormally long period.

[0075] (b) If continuous data is missing for more than a preset number of days (e.g., 7 days), it is determined as a discontinuous signal.

[0076] (c) Multiple signals appear simultaneously with abnormal fluctuations that do not conform to any known state pattern.

[0077] After segmentation is triggered, the system assigns data before the abnormal interval to the previous period segment and data after the abnormal interval to the new period segment. The state inference of the new period segment is based solely on the data within that segment and is unaffected by the previous period segment. This avoids data from abnormal periods contaminating the current state inference.

[0078] Segmentation can be done using hard segmentation (complete isolation) or soft segmentation (weight decay, i.e., the influence weight of the previous period segment decays exponentially over time).

[0079] Example 7: Data Sufficiency Assessment and Degradation Mechanism. This example describes the protection mechanism when data is insufficient.

[0080] In actual use, users may encounter the following situations where data is insufficient: (a) new users have just started using the watch and there is insufficient historical data to establish a reliable baseline; (b) users are not wearing the watch, resulting in missing data; (c) data is missing for certain signal dimensions (e.g., the watch only collects heart rate but not temperature).

[0081] Before performing state space construction, the system performs the following sufficiency checks: (a) Data continuity check: at least M days (e.g., 5 days) within the most recent N days (e.g., 7 days) have valid data.

[0082] (b) Data coverage check: Does the number of available signal dimensions meet the minimum requirement (at least 2 signals)?

[0083] (c) Baseline reliability check: Whether the individual baseline model has been established based on sufficient historical data (at least 14 days).

[0084] When any sufficiency condition is not met, the system performs a degradation process: First-level degradation: When the signal dimension is insufficient but the time data is sufficient, it switches to simplified inference based on the available signal and marks the output as "limited confidence".

[0085] Level 2 downgrade: When time data is insufficient but historical cycle records are available, switch to calendar prediction based on historical statistics and clearly mark it as "based on historical calculations, not real-time signal inference".

[0086] Level 3 Degradation: When historical data is also insufficient, the system does not output state inference results, but only displays the original signal trend, prompting the user to continue wearing the device to accumulate data.

[0087] Example 8: Indirect Inference of Hormone-Dominant State. This example describes a method for indirectly inferring hormone-dominant state using multimodal signals.

[0088] During a woman's menstrual cycle, different hormones dominate at different stages: estrogen dominates the follicular phase, followed by a luteinizing hormone (LH) surge before ovulation, and progesterone dominates the luteal phase. Directly measuring hormone levels requires blood or urine tests, which cannot be achieved using wearable devices.

[0089] This method makes indirect inferences based on the following signal-hormone mapping relationship: (a) Indirect evidence of elevated progesterone: sustained increase in wrist temperature (progesterone acts on the hypothalamic thermoregulatory center after ovulation) + decrease in HRV (progesterone affects the balance of the autonomic nervous system) + slight increase in RHR. When the synergistic changes of the above three signals are simultaneously satisfied, the system infers that the current state is progesterone-dominated (luteal phase).

[0090] (b) Indirect evidence of estrogen dominance: wrist temperature in the low baseline range + HRV at a high level + RHR in the low baseline range. When the above synergistic pattern persists, the system infers that the current state is estrogen dominance (follicular phase).

[0091] (c) Indirect evidence of ovulation: the turning point from low to high temperature + a brief drop in HRV + possible changes in sleep quality.

[0092] Key limitation: The above inference relies on the coordinated changes of at least two signals. If only a single signal changes (e.g., only an increase in temperature), the system will not determine the hormone-dominated state, but will instead mark the result as "pending confirmation".

[0093] Example 9: Intervention Effectiveness Evaluation. This example describes the specific implementation method of the intervention feedback loop.

[0094] Users can record intervention behaviors in the system, such as starting regular exercise, adjusting sleep time, supplementing with specific nutrients (such as magnesium and vitamin B6), and reducing caffeine intake.

[0095] The system records the start time of the intervention and continuously tracks changes in the following indicators in subsequent cycles: (a) Signal level: compare changes in the statistical characteristics of signals before and after the intervention. For example, the average temperature increase during the luteal phase in the three cycles before the intervention was 0.2°C, and the average temperature increase in the three cycles after the intervention was 0.35°C.

[0096] (b) State level: Compare the changes in state distribution before and after intervention. For example, before intervention, the "high risk of PMS" state often occurred in the late luteal phase. After intervention, did the frequency of this state decrease?

[0097] (c) Symptom level: Combine the symptom data actively recorded by users (such as mood swings, abdominal pain, headache, etc.) to compare the changes in symptom severity and frequency before and after intervention.

[0098] The evaluation of intervention effectiveness requires comparison of data from at least one complete period before and after the intervention, preferably three periods each, to reduce the interference of natural fluctuations. The evaluation results are presented to the user in a visual format.

[0099] Example 10: State-space implementation based on Hidden Markov Model. This example describes a specific algorithmic implementation of the state-space construction and constraint contraction process.

[0100] A Hidden Markov Model (HMM) is used, where the hidden states correspond to the set of female physiological states S, and the observed values ​​correspond to the deviation feature vector D of multimodal physiological signals.

[0101] The model parameters include: (a) a state transition probability matrix A, where A[i][j] represents the probability of transitioning from state i to state j. State transition order constraints are achieved by setting physiologically impossible transition probabilities to 0 (e.g., directly from the early luteal phase to the early follicular phase). (b) an observation probability distribution B, where B[i] describes the conditional probability distribution of physiological signal deviation characteristics in state i. (c) an initial state distribution π.

[0102] In this implementation, the state space corresponds to the forward probability vector of the HMM (i.e., the posterior probability of each hidden state at each time step), and the constraint shrinking process corresponds to the constraint effect of the state transition probability in the forward algorithm and the update effect of the observation probability.

[0103] This implementation is a specific example of the method of the present invention and does not limit the scope of protection of the present invention. The state space construction and constraint shrinkage process of the present invention is also applicable to other implementations such as Conditional Random Fields (CRF), Bayesian networks, rule engines, or hybrid methods.

[0104] Furthermore, the constraint contraction process is not limited to implementation in the form of explicit rules. When using deep learning models (such as recurrent neural networks, Transformers, etc.) for state inference, if the model output is equivalent to performing physiological constraint-based screening or contraction on candidate states through the distribution characteristics of training data, the design of the loss function, or the structure of the model architecture (such as the masking mechanism of the state transition matrix) during the model training process, then the implicit constraints in this training process also belong to the implementation form of the constraint contraction process described in this invention.

[0105] Example 11: Application Scenarios for PMS Prediction. This example describes a specific scenario where the above method is applied to the prediction of premenstrual syndrome (PMS).

[0106] PMS typically occurs in the late luteal phase (3-7 days before menstruation), and symptoms include mood swings, bloating, breast tenderness, and fatigue. Accurate prediction of PMS requires accurately identifying whether the user has entered the luteal phase and the specific stage of that phase.

[0107] The system's workflow is as follows: (a) continuously collect the user's multimodal physiological signals; (b) when the status inference result changes from "follicular phase" to "early luteal phase", the system confirms that ovulation has occurred; (c) starting from this change point, and combining the statistical characteristics of the user's historical luteal phase length, predict the start and end dates of the PMS high-risk window; (d) issue an early warning notification 2-3 days before the arrival of the PMS high-risk window.

[0108] The accuracy of this prediction directly depends on the precision with which the method of this invention identifies the ovulation time and the start of the luteal phase. Compared to calendar-based calculation methods, this method can adapt to the natural variations in user cycle length.

[0109] Example 12: System Architecture and Data Flow. This example describes the overall architecture and data flow of the system.

[0110] The system is deployed on the user's mobile terminal (such as iPhone), and data storage uses a local database (such as the GRDB framework based on SQLite). All physiological data and inference results are stored locally on the user's device and are not uploaded to a remote server.

[0111] The data flow of the system is as follows: (1) Data acquisition layer: Physiological signal data is periodically synchronized from the smartwatch to the local database through Apple HealthKit API or other wearable device SDK.

[0112] (2) Data processing layer: Perform preprocessing (outlier removal, time alignment, daily aggregation) and write the processed data to the local database.

[0113] (3) Baseline modeling layer: Individual baseline models are built and updated based on historical data, and baseline parameters are stored in a local database.

[0114] (4) State inference layer: performs state space construction, constraint shrinking and final state determination, and writes the inference results to the local database.

[0115] (5) Application layer: Based on the inference results, provide users with a variety of application functions, including but not limited to: PMS prediction and early warning, ovulation period identification and conception window prompt, pregnancy preparation assistance and luteal function assessment, contraceptive safe period reference, early pregnancy physiological change monitoring, postpartum cycle recovery tracking, perimenopausal status management, PCOS and endometriosis-related cycle abnormality monitoring, cycle health report generation, intervention effect evaluation, etc.

[0116] Each layer communicates with other layers through well-defined data interfaces, and each module can be updated and replaced independently without affecting the overall system architecture.

[0117] Example 13: Risk Probability Calculation. This example describes a method for calculating the risk probability of a specific physiological event based on state inference results.

[0118] After completing state inference (S6), the system can further calculate the risk probability associated with specific physiological events. Taking PMS risk as an example, the calculation process is as follows: (a) Obtain the confidence value c_luteal_late of the "late luteal phase" state from the constrained state space.

[0119] (b) Calculate the current PMS risk probability by combining the frequency of PMS symptoms in the late luteal phase of the user's historical cycle with p_hist: P_pms = c_luteal_late × p_hist × w_signal, where w_signal is the weighting coefficient of the current signal deviation.

[0120] (c) Similarly, the ovulation window probability is calculated based on the confidence level of the "ovulation period" candidate state and the intensity of the temperature transition signal; the cycle anomaly probability is calculated based on the degree to which the current cycle length deviates from the historical mean and the degree of anomaly in the signal pattern.

[0121] The risk probability can replace or supplement the final physiological state as the output form of the system. On the user interface, the system can simultaneously display the current state (e.g., "mid-luteal phase") and the related risk (e.g., "PMS risk: moderate, 62%)".

[0122] Example 14: Intelligent Recommendation Generation Based on Language Model. This example describes a method for combining state inference results with a language model to generate personalized health recommendations.

[0123] The system provides the following information as structured input to the language model or generative model: (a) the current inferred physiological state and its confidence level; (b) the relevant risk probability; (c) a summary of recent signal change trends; (d) the user's historical periodic characteristics; and (e) the user's recorded intervention behaviors and their effect evaluation results.

[0124] Based on the above input, the language model generates personalized health advice or interpretations. For example, when the system infers that the user is in the mid-luteal phase and has a moderate risk of PMS, the language model can generate the following advice: "You may enter a high-risk period for PMS in 3 days. It is recommended to adjust your sleep schedule in advance. Based on your data from the past 3 cycles, regular exercise during this phase has helped reduce the severity of your PMS symptoms."

[0125] The language model can be deployed and run on a local device, or it can be accessed via a cloud interface. When using the cloud model, the system only transmits anonymized structured summary information and does not transmit the original physiological signal data to protect user privacy.

Claims

1. A method for inferring female physiological state based on multimodal physiological signals, wherein the method is not used for disease diagnosis and treatment, characterized in that, Includes the following steps: S1. Acquire multimodal physiological signal data of the user through a wearable device. The multimodal physiological signal data includes at least two physiological signals, including but not limited to: skin temperature, heart rate variability, resting heart rate, sleep data, skin conductance signal, blood oxygen saturation, respiratory rate, and sweat composition data. The multimodal physiological signal data may also include subjective report data from the user. S2. Preprocess the multimodal physiological signal data, including outlier removal, time alignment, and daily aggregation; S3. Based on the preprocessed multimodal physiological signal data, construct an individual baseline model for the user. The individual baseline model is used to describe the stable range and deviation pattern of the user in each physiological signal dimension. The individual baseline model can be directly constructed based on the user's historical data, or it can be obtained based on the population statistical baseline after continuous calibration of the user's data. S4. Calculate the deviation features of the physiological signal within the current time window relative to the individual baseline model, and construct a state space containing multiple candidate physiological states based on the deviation features. Each candidate state in the state space corresponds to a confidence value. S5. Perform constraint shrinkage processing on the state space, wherein the constraint shrinkage processing includes at least two of the following constraints: - State transition order constraint: Based on the physiological sequence of each state in the female menstrual cycle, exclude candidate states that do not conform to the transition order; - State duration constraint: Based on the reasonable duration range of each physiological state, exclude candidate states with abnormal duration; - Signal change direction constraint: Based on the signal change direction characteristics corresponding to each physiological state, exclude candidate states with mismatched signal change directions; - Multimodal signal coordination constraint: Based on the coordination change relationship between at least two physiological signals, exclude candidate states with uncoordinated signal changes. S6. From the state space after constraint contraction, select the candidate state with the highest confidence as the final physiological state output. The determination of the final physiological state depends on the synergistic relationship between at least two physiological signals; a single physiological signal is not the sole basis for determining the final physiological state.

2. The method as described in claim 1, characterized in that, The constraint shrinkage process includes performing any one or more operations on the state space, such as filtering, pruning, reordering, probability reweighting, or probability convergence. The constraint shrinkage process can be implemented in the form of explicit rules, model post-processing, or implicit constraints during model training, as long as its effect is to reduce the number of candidate states or adjust the confidence distribution of candidate states.

3. The method as described in claim 1, characterized in that, It also includes the periodic segmentation step: When abnormal intervals are detected in the physiological cycle data, the data is divided into different cycle segments. The state inference between different cycle segments is isolated from each other, and the state inference of the current cycle segment is based only on the data within that cycle segment. The abnormal interval includes at least one of the following: a period length greater than a preset threshold, abnormal signal fluctuation, or signal discontinuity.

4. The method as described in claim 1, characterized in that, It also includes data sufficiency assessment and downgrade steps: Before performing step S4, the sufficiency of the currently available data is determined, and the sufficiency determination includes at least one of data continuity, data coverage, and data validity. When the data does not meet the preset sufficiency conditions, a downgrade process is performed, which includes at least one of the following operations: prohibiting the execution of complete state space construction and constraint shrinkage, reducing the confidence index of the output result, and switching to approximate prediction based on historical statistical data.

5. The method as described in claim 1, characterized in that, It also includes the indirect inference steps for hormone-dominant states: Based on the combined change structure of at least two physiological signals, a signal-state mapping relationship is established, and the user's current hormone-dominant state is indirectly inferred using the mapping relationship. The hormone-dominant state includes at least one of the following: follicular estrogen-dominant state, ovulation state, and luteal progesterone-dominant state.

6. The method as described in claim 1, characterized in that, It also includes steps for evaluating the effectiveness of the intervention: The intervention behavior data of the user is obtained, and the impact of the intervention behavior on the user's physiological state is assessed by comparing changes in physiological signals, changes in state distribution and / or changes in symptom risk in at least one period before and after the intervention behavior.

7. A system for inferring female physiological state based on multimodal physiological signals, characterized in that, include: The data acquisition module is used to acquire multimodal physiological signal data of users through wearable devices; The preprocessing module is used to perform outlier removal, time alignment, and daily aggregation on the multimodal physiological signal data. The baseline modeling module is used to build an individual baseline model for a user based on the preprocessed data. The individual baseline model can be built directly based on the user's historical data, or it can be obtained based on the population statistical baseline after continuous calibration of the user's data. The state space construction module is used to construct a state space containing multiple candidate physiological states based on the deviation characteristics of the current signal relative to the individual baseline. The constraint shrinking module is used to shrink the state space based on at least two of the following constraints: state transition order constraint, state duration constraint, signal change direction constraint, and multimodal signal cooperative constraint. The state output module is used to determine and output the final physiological state from the contracted state space.

8. The system as described in claim 7, characterized in that, It also includes at least one of the following modules: The periodic segmentation module is used to divide the data into mutually isolated periodic segments when abnormal intervals are detected; The data sufficiency judgment module is used to determine whether the data meets the sufficiency conditions before state inference; The degradation control module is used to perform degradation processing when data is insufficient; The intervention feedback module is used to assess the impact of intervention behaviors on physiological states.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6 and 11 to 12.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 6 and 11 to 12.

11. The method as described in claim 1, characterized in that, It also includes the risk probability calculation step: Based on the confidence distribution during the final physiological state and / or the constraint contraction process, calculate the risk probability associated with a specific physiological event, which includes at least one of premenstrual syndrome onset, ovulation window, and cycle abnormality. The risk probability, as an output of the method, can replace or supplement the output of the final physiological state.

12. The method as described in claim 1, characterized in that, It also includes the intelligent suggestion generation step: The final physiological state, risk probability, and / or physiological signal data are used as input to a language model or generative model, which then generates personalized health advice or interpretation information based on the input.

Images